Transformer fault intelligent diagnosis method and system based on deep convolutional network

Abstract

The application provides a transformer fault intelligent diagnosis method and system based on a deep convolutional network, which comprises the following steps: capturing a global transient signal of a transformer by a spatially-discretized deployed heterogeneous physical field synchronous sampling array, inputting three types of heterogeneous data streams, i.e., vibration spectrum, electromagnetic leakage wave and partial discharge pulse, into a multi-scale feature decoupling program; forming an initial feature primitive set for fault diagnosis; inputting the generated vibration modal tensor, electromagnetic eddy current feature spectrum and discharge pulse feature cloud into a feature interaction contribution analysis program; generating a fault sensitive feature topology graph carrying a weight identifier; training a deep differentiable diagnosis decision tree by using the fault sensitive feature topology graph, analyzing and positioning a key diagnosis node; and realizing millisecond-level synchronous output of fault type and positioning information. The system comprises a multi-modal feature generation module, an adaptive fusion module and a lightweight diagnosis module. The application provides a new generation of diagnosis paradigm with explainability, low time delay and high precision.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, and in particular to a method and system for intelligent diagnosis of transformer faults based on deep convolutional networks. Background Technology

[0002] As a core component of the power grid system, the operating status of power transformers directly affects the reliability of the entire power system. Traditional fault diagnosis mainly relies on periodic maintenance and offline oil chromatography analysis, which suffers from significant lag. With the advancement of smart grid construction, the industry urgently needs diagnostic methods that can capture early fault signs in real time and possess self-learning capabilities. Breakthrough advancements in deep convolutional networks in image recognition have provided a novel approach for feature mining of multidimensional monitoring data from transformers.

[0003] Prior art 1, application number: CN 202411438886.8, discloses an intelligent diagnostic method and system for transformer faults. It reads the operating signals of the longitudinal differential protection from the transformer fault recording and calculates the rated current, three-phase differential current amplitude, voltage and current amplitudes, and zero-sequence voltage and current amplitudes of each side of the transformer based on the fault recording information. Combining different criteria, it intelligently identifies fault types such as non-zone faults, three-phase symmetrical faults, single-phase grounding faults, two-phase grounding faults, two-phase short circuits, and inter-turn faults, as well as the fault location and phase. While it can quickly and automatically locate the transformer fault type and location, assisting or replacing manual fault analysis and providing an intelligent automatic solution for transformer fault accident handling in engineering sites, it relies solely on electrical quantities such as current and voltage for fault judgment and cannot effectively identify mechanical faults such as winding deformation and core loosening. Furthermore, the differential protection requires the fault current to reach a threshold to operate and is insensitive to early partial discharge (<10 pC).

[0004] Prior art two, application number: CN 202010302455.4, discloses a model method for intelligent diagnosis of transformer faults. This method establishes a basic AI parameter feature library for fault diagnosis, including vibration, temperature, and current, based on historical experience. It then uses a self-learning algorithm to dynamically optimize the dynamic AI feature library of these three parameters, improving the accuracy of transformer fault diagnosis. Finally, it establishes a logical algorithm tree model for jointly diagnosing transformer faults using these three parameters, thereby supporting the intelligent diagnosis capability for any transformer's real-time operational faults. While this method effectively ensures intelligent production in power and electrical enterprises, the fixed feature library built upon historical experience is difficult to adapt to new fault modes, such as vibration anomalies caused by high-frequency harmonics. Furthermore, the manual definition of weights is not objective; the parameter weights for vibration, temperature, and current need to be manually set, which can easily introduce subjective bias.

[0005] Existing technology three, application number: CN 202010049713.2, discloses a transformer intelligent fault diagnosis method based on an edge-cloud collaborative mechanism, including: pre-training ResNet-18, ResNet-50-1, 2, and LSTM network models in the cloud; collecting and pre-processing transformer fault feature variable data at the edge and uploading the data to the cloud; using the edge ResNet-18 network and the cloud ResNet-50-1 network to collaboratively discriminate transformer fault types; if a new fault type appears, the cloud reconstructs and retrains the ResNet-18 and ResNet-50-2 networks; the cloud periodically retrains the ResNet-18 network and updates the edge network parameters using the edge-cloud collaborative mechanism; if data packet loss occurs during transmission from the edge to the cloud, the lost data is predicted based on the LSTM network. Although this method can improve the state detection rate of transformer faults, the ResNet-18 model requires >50MB of storage space, making it difficult to deploy on low-cost terminals; frequent data uploads to the cloud for inference result in network latency leading to a response time >2 seconds.

[0006] Current technologies 1, 2, and 3 suffer from problems such as the inability to fuse multi-source data, leading to feature conflicts, low diagnostic accuracy, and poor adaptability to edge computing. Therefore, this invention provides a method and system for intelligent transformer fault diagnosis based on deep convolutional networks. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an intelligent transformer fault diagnosis method based on deep convolutional networks, comprising the following steps:

[0008] A deep differentiable diagnostic decision tree is trained using a fault-sensitive feature topology graph, and key diagnostic nodes are located through backpropagation path saliency analysis. Micro-feature slices corresponding to key diagnostic nodes are extracted and input into a fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is then reconstructed into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information.

[0009] Optionally, the process of generating a binary diagnostic factor set containing only the core decision logic includes the following steps:

[0010] Based on the fault association strength identified by the edge weights of the fault-sensitive feature topology graph, highly sensitive feature subgraphs with weight values ​​exceeding the material yield strength threshold are extracted. A causal decision manifold based on the fault energy transfer path is established between the nodes of the highly sensitive feature subgraphs, and the physical decision partial order relationship of each feature primitive on the output manifold is used as the decision tree splitting criterion.

[0011] The initial set of feature primitives is input into the deep differentiable diagnostic decision tree. Its splitting nodes are dynamically adjusted according to the physical decision partial order relationship. Each decision node corresponds to: the feature subset that satisfies the material elastic deformation constraint on the left branch, and the feature subset that triggers the insulation breakdown condition on the right branch. Fault type labels are injected into the output of the deep differentiable diagnostic decision tree. The contribution of each decision node to the diagnostic result is calculated by the gradient flow tracing algorithm. The key decision path that causes gradient mutation is extracted, and the key diagnostic nodes in the path that exceed the equipment safety boundary are located.

[0012] From the feature space corresponding to the key diagnostic node, extract feature slices that simultaneously meet the following conditions: the slice boundary contains the spatiotemporal coordinates of the deformation-induced breakdown marker, and the slice data distribution conforms to the statistical law of the energy flow conversion feature chain; input the micro-feature slices into the fault factor distiller and compress them through third-order physical compression.

[0013] Optionally, first-order energy threshold filtering: removes feature points in the slice that are below the partial discharge initiation electric field; second-order causal chain encoding: rearranges the feature points into a decision sequence according to the spatiotemporal order of the energy flow conversion feature chain; third-order binary mapping: generates binary diagnostic factors based on the Boolean relationship between the spatial domain where the feature point is located and the material yield strength.

[0014] Optionally, the prediction error of the output layer of the differentiable diagnostic decision tree is injected into the virtual deformation energy field generator, which converts the difference in fault type labels into virtual deformation energy under energy conservation constraints, and distributes the virtual deformation energy in reverse along the branch path of the deep differentiable diagnostic decision tree to generate a fault energy density field covering all nodes of the decision tree.

[0015] In the fault energy density field, the energy density gradient at each decision node is solved. The gradient direction indicates the fault energy transfer path, and the gradient magnitude characterizes the importance of the node decision. Nodes whose energy density gradient exceeds the material's energy storage limit are identified. When the absolute value of the energy density gradient is greater than half of the product of the yield strength and the yield strain, a mutation marker is triggered.

[0016] Tracing back the fault energy transfer path from the mutation node along the negative gradient direction, extracting nodes in the path whose splitting threshold satisfies any of the following conditions: the displacement gradient threshold is greater than the elastic limit of the winding material, or the local field strength threshold is greater than the critical breakdown field strength of the insulating medium; outputting the set of nodes that pass the verification is the critical decision path.

[0017] Optionally, the process of tracing the fault energy transfer path back from the mutation node along the negative gradient direction includes the following steps:

[0018] Starting from the mutation node, the fault energy transfer path tracking operator is activated along the negative direction of the energy density gradient. When traversing adjacent nodes, the preset conditions are forced to be met to constrain the continuity of energy transfer.

[0019] Extract nodes that satisfy the displacement gradient threshold in the fault energy transfer path, and the spatial coordinates of the nodes coincide with the deformation-induced breakdown marker; extract nodes that satisfy the local field strength threshold in the fault energy transfer path, and the electromagnetic energy equipotential surface from the local field strength threshold reaches the boundary value, and the node timestamp matches the peak time of the energy flow conversion characteristic chain.

[0020] The verification nodes are rearranged according to the direction of fault energy transmission, and each edge is assigned a multiple by which the energy of the quantized path segment exceeds the breakdown threshold.

[0021] Optionally, the process of extracting nodes that satisfy the local field strength threshold in the fault energy transfer path includes the following steps:

[0022] Extract the position-dependent critical breakdown field strength from the generated electromagnetic energy equipotential surface; dynamically generate time-varying field strength thresholds at each node of the fault energy transfer path.

[0023] The peak moment of discharge energy conversion rate of the energy flow conversion feature chain at the node spatial coordinates is obtained only if the path node simultaneously satisfies the decision tree splitting threshold and the node timestamp.

[0024] Obtain node puncture risk factors, bind node puncture risk factors to spatial coordinates, and generate insulation failure early warning nodes carrying spatiotemporal risk identifiers.

[0025] Optionally, the process of obtaining node penetration risk factors includes the following steps:

[0026] Input the decision tree splitting threshold at the node into the effective insulation barrier calculator, call the critical breakdown field strength of the electromagnetic energy equipotential surface corresponding to the node's spatial coordinates; introduce the discharge energy conversion rate pulse waveform of the differential characteristics of the energy flow conversion characteristic chain at the peak moment; output the dynamic reference field strength margin modulated by the pulse waveform.

[0027] Calculate the absolute difference between the decision threshold and the dynamic reference field strength, obtain the discharge energy conversion rate pulse phase with the neighborhood derivative sign at the peak time, mark the energy accumulation period with the positive phase and the energy release period with the negative phase;

[0028] The scalar risk value is synthesized to obtain the node breakdown risk factor; the spatiotemporal feature coding is as follows: a node breakdown risk factor greater than 0 indicates insufficient insulation margin during the energy accumulation period, and a node breakdown risk factor not greater than 0 indicates breakdown backstroke effect during the energy release period. The node breakdown risk factor quantifies the risk intensity level.

[0029] Optionally, a heterogeneous physical field synchronous sampling array deployed in spatial discretization is used to capture the transient signal of the transformer across the entire domain. The three heterogeneous data streams of vibration spectrum, electromagnetic leakage wave and partial discharge pulse are input into a multi-scale feature decoupling program. The multi-scale feature decoupling program uses an asymmetric deep convolution kernel group to perform feature stripping along the three-dimensional directions of time-space-frequency domain, and outputs a physically meaningful vibration mode tensor, electromagnetic eddy current feature spectrum and discharge pulse feature cloud to form an initial feature primitive set for fault diagnosis.

[0030] Optionally, the generated vibration mode tensor, electromagnetic eddy current characteristic spectrum, and discharge pulse characteristic cloud are input into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression on redundant features based on a dynamic attenuation function. The calibration results of the information contribution intensity drive the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types.

[0031] This invention provides an intelligent transformer fault diagnosis system based on deep convolutional networks, comprising:

[0032] The multimodal feature generation module is used to capture the full-domain transient signals of the transformer through a spatially discretized heterogeneous physical field synchronous sampling array. It inputs three types of heterogeneous data streams—vibration spectrum, electromagnetic leakage wave, and partial discharge pulse—into a multi-scale feature decoupling program. The multi-scale feature decoupling program uses an asymmetric deep convolution kernel group to perform feature stripping along the three-dimensional directions of time, space, and frequency domain, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds to form the initial feature primitive set for fault diagnosis.

[0033] The adaptive fusion module is used to input the generated vibration mode tensor, electromagnetic eddy current feature spectrum, and discharge pulse feature cloud into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression of redundant features based on a dynamic decay function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types.

[0034] The lightweight diagnostic module is used to train a deep differentiable diagnostic decision tree using a fault-sensitive feature topology graph, and to locate key diagnostic nodes through backpropagation path saliency analysis. The micro-feature slices corresponding to the key diagnostic nodes are extracted and input into the fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is then reconstructed into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information.

[0035] This invention achieves a fully intelligent upgrade of transformer fault diagnosis, from signal acquisition to decision output, through multimodal feature collaborative mining and dynamic decision optimization. It utilizes a heterogeneous physical field synchronous sampling array to jointly capture electromagnetic, mechanical, and discharge multi-physical quantities, and combines this with asymmetric deep convolution kernel groups for spatiotemporal-frequency three-dimensional feature decoupling, overcoming the limitations of traditional single-mode detection. Parallel extraction of vibration mode tensors, electromagnetic eddy current spectra, and discharge feature clouds constructs a complete feature representation system covering typical transformer faults such as mechanical loosening, winding deformation, and insulation degradation. A bidirectional feature entropy weighting evaluation mechanism quantifies the contribution of each feature primitive to the fault mode through information entropy, and, in conjunction with a dynamic decay function, achieves nonlinear dimensionality reduction of the feature space. The weighted topology graph generated by the cross-modal adaptive fusion gate embeds physically interpretable feature correlation strength into the deep learning architecture, significantly improving the model's ability to identify complex faults. Deep differentiable decision trees enable the visualization and tracing of decision paths through backpropagation saliency analysis, while the binary diagnostic factor set extracted by the fault factor distiller supports dynamic model reconstruction on edge devices. This architecture enables the end-to-end latency from the raw signal to the diagnostic result to be controlled in milliseconds, while maintaining diagnostic accuracy comparable to cloud models.

[0036] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0037] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a flowchart of the intelligent transformer fault diagnosis method based on deep convolutional networks in Embodiment 1 of the present invention;

[0040] Figure 2This is a schematic diagram of the intelligent transformer fault diagnosis method based on deep convolutional networks in Embodiment 1 of the present invention.

[0041] Figure 3 This is a process diagram of forming the initial feature set for fault diagnosis in Embodiment 2 of the present invention;

[0042] Figure 4 This is a process diagram of generating a fault-sensitive feature topology map with weighted identifiers in Embodiment 3 of the present invention;

[0043] Figure 5 This is a flowchart illustrating the process of generating a binary diagnostic factor set containing only core decision-making logic in Embodiment 6 of the present invention.

[0044] Figure 6 This is a block diagram of the intelligent transformer fault diagnosis system based on deep convolutional networks in Embodiment 12 of the present invention;

[0045] Figure 7 This is a schematic diagram of the intelligent transformer fault diagnosis system based on deep convolutional networks in Embodiment 12 of the present invention. Detailed Implementation

[0046] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0047] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0048] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0049] Example 1: As Figure 1 As shown, this embodiment of the invention provides an intelligent fault diagnosis method for transformers based on deep convolutional networks, comprising the following steps:

[0050] S100: The heterogeneous physical field synchronous sampling array deployed in spatial discretization captures the transient signals of the transformer across the entire domain. The three heterogeneous data streams of vibration spectrum, electromagnetic leakage wave and partial discharge pulse are input into the multi-scale feature decoupling program. The multi-scale feature decoupling program uses asymmetric deep convolution kernel groups to perform feature stripping in the three-dimensional directions of time-space-frequency domain, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra and discharge pulse feature clouds to form the initial feature primitive set for fault diagnosis.

[0051] S200: The generated vibration mode tensor, electromagnetic eddy current characteristic spectrum, and discharge pulse characteristic cloud are input into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression on redundant features based on a dynamic attenuation function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types.

[0052] S300: A deep differentiable diagnostic decision tree is trained using a fault-sensitive feature topology graph, and key diagnostic nodes are located through backpropagation path saliency analysis; micro-feature slices corresponding to key diagnostic nodes are extracted and input into a fault factor distiller, which generates a binary diagnostic factor set containing only core decision logic through an iterative knowledge compression algorithm; the binary diagnostic factor set is restored into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information.

[0053] The working principle and beneficial effects of the above technical solution are as follows: Firstly, this embodiment captures the full-domain transient signal of the transformer using a spatially discretized heterogeneous physical field synchronous sampling array. The three heterogeneous data streams—vibration spectrum, electromagnetic leakage wave, and partial discharge pulse—are input into a multi-scale feature decoupling program. The multi-scale feature decoupling program uses an asymmetric depth convolution kernel group to perform feature stripping along the time-space-frequency domain three-dimensional directions, outputting physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds, forming an initial set of feature primitives for fault diagnosis. Secondly, the generated vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds are input into a feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the contribution of each feature primitive in the decision space in real time through a bidirectional feature entropy weight evaluation mechanism. The information contribution intensity is determined, and redundant features are nonlinearly suppressed based on a dynamic decay function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly represent the correlation strength between feature primitives and fault types. Finally, a deep differentiable diagnostic decision tree is trained using the fault-sensitive feature topology map, and key diagnostic nodes are located through backpropagation path saliency analysis. The micro-feature slices corresponding to the key diagnostic nodes are extracted and input into the fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is restored to a lightweight diagnostic model in real time on the terminal device through the edge-end dynamic reconstruction engine, realizing millisecond-level synchronous output of fault type and location information (the specific principle is as follows). Figure 2 (As shown). The above scheme achieves a full-link intelligent upgrade of transformer fault diagnosis from signal acquisition to decision output through multimodal feature collaborative mining and dynamic decision optimization. It achieves joint capture of multiple physical quantities (electromagnetic, mechanical, and discharge) using a heterogeneous physical field synchronous sampling array, and combines this with the decoupling of temporal-spatial-frequency three-dimensional features from asymmetric deep convolution kernel groups, overcoming the limitations of traditional single-mode detection. Parallel extraction of vibration mode tensors, electromagnetic eddy current spectra, and discharge feature clouds constructs a complete feature representation system covering typical transformer faults such as mechanical loosening, winding deformation, and insulation degradation. A bidirectional feature entropy weighting evaluation mechanism quantifies the contribution of each feature primitive to the fault mode through information entropy, and achieves nonlinear dimensionality reduction of the feature space with a dynamic decay function. The weighted topology graph generated by the cross-modal adaptive fusion gate embeds physically interpretable feature correlation strength into the deep learning architecture, significantly improving the model's ability to identify complex faults. Deep differentiable decision trees enable the visualization and tracing of decision paths through backpropagation saliency analysis, while the binary diagnostic factor set extracted by the fault factor distiller supports dynamic model reconstruction on edge devices. This architecture enables the end-to-end latency from the raw signal to the diagnostic result to be controlled in milliseconds, while maintaining diagnostic accuracy comparable to cloud models.

[0054] In summary, this embodiment provides a new generation of diagnostic paradigm for power equipment condition monitoring that is interpretable, low-latency, and high-precision. It organically integrates physical mechanism-driven feature engineering with data-driven deep learning, solving the industry pain points of feature redundancy and decision black box in traditional methods.

[0055] Example 2: Figure 3 As shown, based on Embodiment 1, the process of forming an initial feature set for fault diagnosis provided by this embodiment of the invention includes the following steps:

[0056] S101: The vibration spectrum data is captured by transient waveform slicing to capture the energy accumulation characteristics of mechanical vibration in the time domain, and outputs a time-based feature field carrying phase information; the electromagnetic leakage wave data stream is extracted using the field strength gradient sensing mechanism to extract the attenuation distribution pattern of the electromagnetic field in three-dimensional space, generating a spatial basis feature field with position encoding; the partial discharge pulse data stream is separated by harmonic resonance detection to separate the energy burst points of the pulse group in the frequency band, forming a frequency domain basis feature field with marked spectral density;

[0057] S102: The time-based feature field is generated by modal tensor to map the time phase information into a differential geometric expression of the rotor motion trajectory, and reconstructed into a vibration mode tensor characterizing the mechanical deformation law; the spatial basis feature field is input into the electromagnetic eddy current resolver, and based on the Maxwell-Faraday equivalence principle, the spatial gradient distribution is transformed into a three-dimensional vector field of induced eddy current, and reconstructed into an electromagnetic eddy current feature spectrum describing electromagnetic energy dissipation; the frequency domain basis feature field is input into the discharge pulse evolution device, and the peak value of the spectral density is associated with the electron avalanche development process of the insulation defect using the breakdown process inversion model, and a discharge pulse feature cloud reflecting the discharge evolution path is constructed;

[0058] S103: Synchronize the vibration mode tensor, electromagnetic eddy current characteristic spectrum and discharge pulse characteristic cloud, and establish a parallel dynamic alignment mechanism for tensors based on the intrinsic dimensions of the three types of characteristic fields. This ensures that mechanical deformation, electromagnetic dissipation and insulation breakdown processes maintain physical consistency in the fault spatiotemporal coordinate system, and finally integrates an initial set of characteristic primitives with causal relationships.

[0059] The working principle and beneficial effects of the above technical solution are as follows: Firstly, this embodiment captures the energy accumulation characteristics of mechanical vibration in the time domain by using transient waveform slicing to obtain the vibration spectrum data, outputting a time-based feature field carrying phase information; secondly, it extracts the attenuation distribution pattern of the electromagnetic field in three-dimensional space from the electromagnetic leakage wave data stream using a field strength gradient sensing mechanism, generating a spatial basis feature field with position encoding; thirdly, it separates the energy burst points of the pulse group in the frequency band using harmonic resonance detection to form a frequency-domain basis feature field with marked spectral density; fourthly, it generates a differential geometric expression of the time phase information as the rotor motion trajectory through modal tensor generation, reconstructing it into a vibration mode tensor characterizing the mechanical deformation law; and finally, it converts the spatial basis feature field into a frequency-domain basis feature field. The electromagnetic eddy current analyzer, based on the Maxwell-Faraday equivalence principle, transforms the spatial gradient distribution into a three-dimensional vector field of induced eddy currents, reconstructing an electromagnetic eddy current characteristic spectrum describing electromagnetic energy dissipation. The frequency domain fundamental characteristic field is input into the discharge pulse evolution generator, and a breakdown process inversion model is used to correlate the spectral density peaks with the electron avalanche development process of insulation defects, constructing a discharge pulse characteristic cloud reflecting the discharge evolution path. Finally, the vibration mode tensor, electromagnetic eddy current characteristic spectrum, and discharge pulse characteristic cloud are synchronized, and a tensor parallel dynamic alignment mechanism is established based on the intrinsic dimensions of the three types of characteristic fields. This ensures that mechanical deformation, electromagnetic dissipation, and insulation breakdown processes maintain physical consistency in the fault spatiotemporal coordinate system, ultimately integrating an initial set of causally related characteristic primitives. This scheme achieves a structured representation of three heterogeneous data types—vibration, electromagnetic, and discharge—mapped to time, spatial, and frequency domain characteristic fields, respectively, providing a foundation for subsequent physical modeling. The original characteristic fields are elevated to higher-order physical representations, enabling vibration, electromagnetic, and discharge signals to possess interpretable expressions based on mechanical dynamics, electromagnetic field theory, and insulation breakdown physics, respectively. An initial set of feature primitives with physical interpretability is formed to ensure that the subsequent fault diagnosis module can make decisions based on the coupling law of multi-physics fields, avoiding feature conflicts or semantic fragmentation.

[0060] In summary, this embodiment achieves a unified representation of heterogeneous data by transforming vibration (time domain), electromagnetic (spatial domain), and discharge (frequency domain) signals into computable and physically meaningful feature fields. Through differential geometry, electromagnetic field theory, and insulation breakdown models, the features convey the deep physical mechanisms of equipment faults. A coupling relationship between mechanical, electromagnetic, and insulation processes is established, providing an initial set of causally related feature primitives for fault diagnosis. This provides high-information-density, low-redundancy, and physically interpretable feature inputs for subsequent adaptive fusion and lightweight diagnostic modules, supporting accurate and robust fault diagnosis.

[0061] Example 3: As Figure 4 As shown, based on Example 1, the process of generating a fault-sensitive feature topology map carrying weight identifiers provided in this embodiment of the invention includes the following steps:

[0062] S201: The vibration mode tensor is input into the forward diagnostic flow evaluation channel. The curvature change pattern of the mechanical deformation trajectory is extracted through differential geometric feature analysis to generate a forward entropy weight vector describing rotor motion anomalies. The electromagnetic eddy current feature spectrum and the discharge pulse feature cloud are synchronously input into the backward correlation flow evaluation channel. The field-flow coupling analyzer is used to obtain the spatiotemporal correlation between electromagnetic energy dissipation and insulation breakdown process, and the backward entropy weight matrix characterizing the fault coupling strength is output. The forward entropy weight vector and the backward entropy weight matrix are subjected to tensor condensation operation in the decision space projection layer to generate a feature cross-entropy matrix that quantifies the diagnostic value of each feature primitive.

[0063] S202: Input the feature cross-entropy matrix into the contribution intensity differential equation, which is constructed based on the entropy decay theorem. For elements in the entropy matrix below the dynamic threshold, activate the nonlinear suppression operator and compress their numerical space through the exponential decay function; apply the information gain amplifier to the elements in the entropy matrix above the dynamic threshold to enhance their resolution in the decision space; the output refined entropy weight tensor retains only the feature primitives that make a substantial contribution to fault diagnosis as its non-zero elements.

[0064] The theorem of the contribution intensity differential equation is based on the irreversible nature of transformer fault energy transfer: when the diagnostic entropy value of a certain feature element is lower than the equipment safe operation threshold, it indicates that the fault energy it carries is within the range of spontaneous dissipation of the system. At this time, a nonlinear suppression operator is triggered—this operator is essentially an exponential decay function based on the tangent of the dielectric loss angle, and its decay coefficient is dynamically adjusted by the thermal aging equation of the insulating material, causing the value of the low contribution feature in the decision space to collapse exponentially with the energy dissipation rate. Conversely, when the entropy value exceeds the critical breakdown field strength of the insulating medium, the information gain amplifier is activated, and its gain factor is positively correlated with the discharge energy accumulation rate. By enhancing the local feature resolution, it forces the potential fault area to form an energy barrier jump in the decision space.

[0065] S203: The refined entropy weight tensor is decomposed into a set of weight identifiers and input into a cross-modal adaptive fusion gate. Using the differential geometric parameters of the vibration mode tensor as the reference axis, the vector field of the electromagnetic eddy current feature spectrum is mapped to the mechanical vibration coordinate system through fault physical space projection. Based on the weight identifiers in the refined entropy weight tensor, the mapped electromagnetic vector field and the electron avalanche path of the discharge pulse feature cloud are coupled with intensity modulation. Spatiotemporal correlation constraints are implanted in the modulated fusion field to form a causal link between mechanical deformation, electromagnetic dissipation and insulation breakdown features under a unified fault coordinate system. A fault-sensitive feature topology graph with weight identifiers on the edges is generated, where the nodes are physical feature primitives and the edge weights correspond to the contribution intensity values ​​in the entropy weight tensor.

[0066] The working principle and beneficial effects of the above technical solution are as follows: Firstly, the vibration mode tensor is input into the forward diagnostic flow evaluation channel. The curvature change pattern of the mechanical deformation trajectory is extracted through differential geometric feature analysis, generating a forward entropy weight vector describing rotor motion anomalies. Then, the electromagnetic eddy current feature spectrum and discharge pulse feature cloud are synchronously input into the backward correlation flow evaluation channel. The spatiotemporal correlation between electromagnetic energy dissipation and insulation breakdown processes is obtained using a field-flow coupling analyzer, outputting a backward entropy weight matrix characterizing the fault coupling strength. The forward entropy weight vector and the backward entropy weight matrix undergo tensor shrinkage operations at the decision space projection layer to generate a feature interaction entropy matrix that quantifies the diagnostic value of each feature element. Secondly, the feature interaction entropy matrix is ​​input into the contribution intensity differential equation, which is constructed based on the entropy decay theorem. For elements in the entropy matrix below the dynamic threshold, a nonlinear suppression operator is activated, compressing its numerical space through an exponential decay function. The entropy matrix... Elements in the array exceeding a dynamic threshold are amplified with information gain amplifiers to enhance their resolution in the decision space. The output refined entropy weight tensor retains only feature primitives that substantially contribute to fault diagnosis among its non-zero elements. Finally, the refined entropy weight tensor is decomposed into a weighted identifier set input to a cross-modal adaptive fusion gate. Using the differential geometric parameters of the vibration mode tensor as the reference axis, the vector field of the electromagnetic eddy current feature spectrum is mapped to the mechanical vibration coordinate system through fault physical space projection. Based on the weights in the refined entropy weight tensor, the mapped electromagnetic vector field and the electron avalanche path of the discharge pulse feature cloud are coupled with intensity modulation. Spatiotemporal correlation constraints are implanted into the modulated fusion field, enabling mechanical deformation, electromagnetic dissipation, and insulation breakdown features to form a causal link under a unified fault coordinate system. A fault-sensitive feature topology map with weighted edges is generated, where nodes are physical feature primitives and edge weights correspond to the contribution intensity values ​​in the entropy weight tensor. This scheme, through multimodal feature fusion and dynamic entropy weight optimization, achieves accurate feature representation and causal correlation modeling for rotating machinery fault diagnosis. Breaking through the limitations of traditional single-mode diagnosis, this method achieves holographic fusion of mechanical vibration and electromagnetic characteristics; it establishes an adaptive selection mechanism for fault-sensitive features through dynamic entropy weight optimization; a weighted labeling system quantifies the contribution of each feature primitive to fault evolution; and the feature topology under spatiotemporal correlation constraints forms a causal network model of fault development. This significantly improves the early identification rate and diagnostic confidence of complex faults.

[0067] Example 4: Based on Example 3, the process provided in this embodiment of the invention for forming a causal link between mechanical deformation, electromagnetic dissipation, and insulation breakdown characteristics under a unified fault coordinate system includes the following steps:

[0068] S2031: The differential geometric parameters of the vibration mode tensor are used to solve the displacement gradient field of the winding deformation in three-dimensional space through the differential equation of the rotor motion trajectory. A fault reference coordinate system with the geometric center of the iron core as the origin is established. The unit vector of each axis of the fault reference coordinate system corresponds to the key deformation direction of the transformer's main insulation structure.

[0069] S2032: The three-dimensional vector field of the electromagnetic eddy current characteristic spectrum is transformed by field strength-stress coupling and mapped to the fault reference coordinate system according to Maxwell's stress tensor principle, outputting an electromagnetic energy equipotential surface with spatial position marking; the electron avalanche path of the discharge pulse characteristic cloud is calibrated into a spatiotemporal continuous breakdown channel by using the differential relationship between the electron avalanche diffusion rate and the thermal conductivity of insulating oil.

[0070] The electron avalanche diffusion rate and the thermal conductivity of the insulating oil are correlated through a differential equation governing the temperature gradient field: the thermal conductivity distribution of the insulating oil determines the internal temperature gradient field of the transformer, which modulates the mobility of charged particles and thus controls the electron avalanche diffusion rate. This process is described by a thermo-electric coupling partial differential equation: the change in the thermal conductivity of the insulating oil and the electron avalanche diffusion acceleration are explicitly correlated through the Joule thermal coefficient of the oil medium, allowing discrete discharge points to be calibrated into continuous breakdown channels that conform to the actual thermodynamic environment.

[0071] S2033: In the overlapping region of the electromagnetic energy equipotential surface and the spatiotemporal continuous breakdown channel, the efficiency coefficient of electromagnetic energy to discharge energy per unit volume is calculated by local energy density integration, and an energy flow conversion characteristic chain is generated; the energy flow conversion characteristic chain is solved simultaneously with the differential equation of rotor motion trajectory, and when the mechanical deformation displacement gradient exceeds the elastic limit of the insulation layer, a deformation-induced breakdown mark is established at the corresponding spatial coordinate point.

[0072] Among them, the energy density integral is: in the overlapping domain of the electromagnetic energy equipotential surface and the breakdown channel, the Poynting vector is integrally divided along the breakdown path to output the efficiency coefficient of electromagnetic energy to discharge energy per unit volume; the efficiency coefficient is embedded as a source term into the existing rotor motion trajectory differential equation. When the displacement gradient obtained by solving the equation exceeds the elastic limit of the insulation layer, the deformation-induced breakdown mark is triggered.

[0073] S2034: Extract the contribution intensity value from the refined entropy weight tensor and transform it into: mechanical deformation edge weight, the ratio of winding displacement gradient to material yield strength; electromagnetic-discharge coupling edge weight, the difference between local field strength and critical breakdown field strength of insulating medium; using deformation-induced breakdown marker as node and energy flow conversion feature chain as connecting edge, encode the weight value of physical interpretation as edge attribute, and finally generate a fault-sensitive feature topology graph carrying physical meaning weight identifier.

[0074] The working principle and beneficial effects of the above technical solution are as follows: Firstly, the differential geometric parameters of the vibration mode tensor are used to solve the displacement gradient field of the winding deformation in three-dimensional space through the differential equation of the rotor motion trajectory. A fault reference coordinate system is established with the geometric center of the iron core as the origin. The unit vectors of each axis of the fault reference coordinate system correspond to the key deformation directions of the transformer's main insulation structure. Secondly, the three-dimensional vector field of the electromagnetic eddy current characteristic spectrum is mapped to the fault reference coordinate system through field strength-stress coupling transformation, based on the Maxwell stress tensor principle, outputting an electromagnetic energy equipotential surface with spatial location markings. The electron avalanche path of the discharge pulse characteristic cloud is calibrated into a spatiotemporal continuous breakdown channel using the differential relationship between the electron avalanche diffusion rate and the thermal conductivity of the insulating oil. Then, the electromagnetic energy equipotential surface and the spatiotemporal continuous breakdown channel are... In the overlapping region, the efficiency coefficient of electromagnetic energy to discharge energy per unit volume is calculated by integrating the local energy density, generating an energy flow conversion feature chain. The energy flow conversion feature chain is solved simultaneously with the differential equation of rotor motion trajectory. When the mechanical deformation displacement gradient exceeds the elastic limit of the insulation layer, a deformation-induced breakdown marker is established at the corresponding spatial coordinate point. Finally, the contribution intensity value in the refined entropy weight tensor is extracted and transformed into: mechanical deformation edge weight, the ratio of winding displacement gradient to material yield strength; electromagnetic-discharge coupling edge weight, the difference between local field strength and critical breakdown field strength of the insulation medium; with the deformation-induced breakdown marker as the node and the energy flow conversion feature chain as the connecting edge, the weight value of physical interpretation is encoded as edge attribute, and finally a fault-sensitive feature topology graph carrying physical meaning weight identifier is generated. The above scheme constructs a multi-physics coupled fault feature quantification and analysis system. Through the coordinated solution of differential geometric parameters and rotor motion equations, the vector quantization of mechanical deformation fields in three-dimensional space is realized. The established fault reference coordinate system unifies heterogeneous physical quantities such as winding displacement gradient, electromagnetic energy distribution, and insulation breakdown path into a spatial reference framework with clear physical meaning. The strict correspondence between the unit vectors of each axis of this coordinate system and the key deformation directions of the transformer's main insulation ensures the physical interpretability of multi-source fault features. The field strength-stress coupling transformation converts the electromagnetic eddy current spectrum into equipotential surfaces with spatial markings. Combined with the differential calibration of diffusion rate and thermal conductivity, a spatiotemporally continuous discharge channel model is constructed. This transforms the originally discrete discharge pulse characteristics into differentiable field quantity expressions, providing a continuous mathematical basis for subsequent energy conversion efficiency calculations. The energy flow conversion characteristic chain generated by the local energy density integral, combined with the solution of the differential equation of the mechanical deformation trajectory, essentially establishes a closed-loop dynamic coupling model of "deformation-electromagnetism-discharge". The deformation-induced breakdown marker triggered when the displacement gradient exceeds the elastic limit of the material realizes the precise spatial positioning of the critical point of mechanical stress accumulation and insulation failure.By transforming the mechanical deformation edge weights (displacement gradient / yield strength) and electromagnetic-discharge coupling edge weights (field strength / critical field strength difference) using entropy weight tensor, the material constitutive relation and insulation performance degradation law are encoded as edge attributes of the topological graph; this enables nonlinear processes such as energy redistribution and stress concentration effects during fault development to have a computable graph theory expression.

[0075] In summary, this embodiment elevates traditional discrete fault characteristic indicators to continuous field quantities with strict physical correspondences. Through progressive processing of differential geometry, tensor operations, and topological graph theory, a computable fault model is constructed that includes complete causal links such as mechanical deformation conduction paths, electromagnetic energy conversion efficiency, and insulation failure trigger thresholds. It can quantitatively characterize the fault evolution process under the coupling of multiple physical fields, providing a characteristic parameter system with clear physical meaning for the condition assessment of power equipment.

[0076] Example 5: Based on Example 4, the process of converting step S2034 of this embodiment into mechanical deformation edge weights and electromagnetic-discharge coupling edge weights includes the following steps:

[0077] S20341: Input the contribution intensity value representing the mechanical characteristics in the refining entropy weight tensor into the material deformation intrinsic mapper, extract the current displacement gradient tensor from the rotor motion trajectory differential equation, retrieve the yield strength parameter of the transformer winding material, and output the weight value through the safety margin quantification criterion to represent the distance between the material deformation state and the yield limit.

[0078]

[0079] in, σ represents the current displacement gradient tensor; y ω represents the yield strength parameter. mech This represents the output weight value of the safety margin quantification criterion; Forb is the matrix norm in tensor analysis, used to quantify the overall deformation degree of the displacement gradient tensor.

[0080] S20342: Input the interaction contribution value of electromagnetic and discharge in the refined entropy weight tensor into the insulation failure risk mapper, obtain the local electric field strength from the electromagnetic energy equipotential surface, call the critical breakdown field strength based on the insulation medium type, and generate the weight value ω through the field strength margin decay function (critical breakdown field strength - local electric field strength). em-pd Negative values ​​indicate the risk of insulation failure, while positive values ​​indicate a safety margin.

[0081] S20343: Use the weight values ​​output by the safety margin quantification criteria and the weight values ​​generated by the field strength margin decay function to perform collaborative verification of equipment failure.

[0082] When ω mechWhen ω > 1, the material plastic deformation marker is triggered; when ω em-pd When the value is less than 0, the insulation breakdown warning flag is activated; for spatial coordinate points that simultaneously trigger both types of flags, the weight value ω′ is adjusted. em-pd :

[0083]

[0084] Where ε represents strain, which is the relative change of the material's dimensions after deformation compared to its original dimensions, specifically referring to plastic strain, which represents the deformation of the material after it enters the plastic stage; σ represents stress, which is the internal force per unit area within the material.

[0085] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, the contribution intensity value representing the mechanical characteristics in the refining entropy weight tensor is first input into the material deformation intrinsic mapper. The current displacement gradient tensor is extracted from the differential equation of the rotor motion trajectory, and the yield strength parameter of the transformer winding material is retrieved. The weight value is output through the safety margin quantification criterion to represent the distance between the material deformation state and the yield limit. Secondly, the interaction contribution value of electromagnetic and discharge in the refining entropy weight tensor is input into the insulation failure risk mapper. The local electric field strength is obtained from the electromagnetic energy equipotential surface. The critical breakdown field strength is called based on the insulation medium type. The weight value ω is generated through the field strength margin decay function (critical breakdown field strength - local electric field strength). em-pd Negative values ​​indicate the risk of insulation failure, while positive values ​​indicate the safety margin. Finally, the safety margin quantification criteria output weight values ​​and the field strength margin attenuation function generate weight values ​​for collaborative verification of equipment failure. This scheme constitutes a multi-physics coupled equipment failure early warning system. It achieves characteristic decoupling and weight allocation between the mechanical deformation field and the electromagnetic field through entropy weight tensor, establishing a dual criterion system of material constitutive equation and insulating medium properties. A dynamic failure criterion generation mechanism is used: on the mechanical side, a plastic deformation criterion is constructed based on the ratio of the Frobenius norm of the displacement gradient tensor to the yield strength; on the electromagnetic side, the insulation state is dynamically assessed through the field strength margin function. The coupled failure collaborative early warning algorithm employs spatial coordinate alignment technology to ensure the consistency of the mesh mapping between the mechanical deformation field and the electromagnetic field, and dual-marking trigger logic. Plastic hardening effect compensation is achieved by reflecting the stress-strain characteristics of the material during the plastic stage through a correction coefficient.

[0086] In summary, this embodiment triggers an early warning before the material reaches its yield strength or the local field strength exceeds the breakdown threshold; it distinguishes between purely mechanical failure, purely insulating failure, and coupling failure; and the introduction of the plastic hardening modulus corrects the evaluation bias of electromagnetic parameters during the material's plastic stage. A failure phase diagram for a strongly coupled electromechanical system is constructed, and dimensionless processing of multi-physics parameters is achieved through tensor operations, ultimately outputting an early warning marking system with clear engineering significance.

[0087] Example 6: As Figure 5As shown in Embodiment 1, the process of generating a binary diagnostic factor set containing only core decision logic in step S300 of this embodiment of the invention includes the following steps:

[0088] S301: Based on the fault association strength identified by the edge weights of the fault-sensitive feature topology graph, extract the highly sensitive feature subgraphs whose weight values ​​exceed the material yield strength threshold. Establish a causal decision manifold based on the fault energy transfer path between the nodes of the highly sensitive feature subgraphs, and output the physical decision partial order relationship of each feature primitive on the manifold as the decision tree splitting criterion.

[0089] S302: Input the initial set of feature primitives into the deep differentiable diagnostic decision tree. Its splitting nodes are dynamically adjusted according to the physical decision partial order relationship. Each decision node corresponds to: the feature subset that satisfies the material elastic deformation constraint on the left branch, and the feature subset that triggers the insulation breakdown condition on the right branch. Inject fault type labels into the output of the deep differentiable diagnostic decision tree. Calculate the contribution of each decision node to the diagnostic result through the gradient flow tracing algorithm, extract the key decision path that causes gradient abrupt change, and locate the key diagnostic nodes in the path whose splitting threshold exceeds the equipment safety boundary.

[0090] S303: Extract feature slices from the feature space corresponding to the key diagnostic nodes that simultaneously meet the following conditions: the slice boundary contains the spatiotemporal coordinates of the deformation-induced breakdown marker, and the slice data distribution conforms to the statistical law of the energy flow conversion feature chain; input the micro-feature slices into the fault factor distiller and compress them through third-order physical compression.

[0091] First-order energy threshold filtering: Removes feature points in the slice that are below the partial discharge initiation electric field; Second-order causal chain encoding: Rearranges the feature points into a decision sequence according to the spatiotemporal order of the energy flow conversion feature chain; Third-order binary mapping: Generates binary diagnostic factors based on the Boolean relationship between the spatial domain where the feature point is located and the material yield strength.

[0092] The working principle and beneficial effects of the above technical solution are as follows: Firstly, based on the fault association strength identified by the edge weights of the fault-sensitive feature topology graph, highly sensitive feature subgraphs with weight values ​​exceeding the material yield strength threshold are extracted. A causal decision manifold based on the fault energy transfer path is established between the nodes of the highly sensitive feature subgraphs. The physical decision partial order relationship of each feature primitive on the output manifold is used as the decision tree splitting criterion. Secondly, the initial set of feature primitives is input into a deep differentiable diagnostic decision tree, whose splitting nodes are dynamically adjusted according to the physical decision partial order relationship. Each decision node corresponds to: a feature subset satisfying the material elastic deformation constraint on the left branch, and a feature subset triggering the insulation breakdown condition on the right branch. Fault type labels are injected into the output of the deep differentiable diagnostic decision tree, and each fault type is calculated using the gradient flow tracing algorithm. The contribution of decision nodes to the diagnostic results is analyzed, key decision paths leading to gradient mutations are extracted, and key diagnostic nodes in the paths whose splitting thresholds exceed the equipment safety boundary are located. Finally, feature slices that simultaneously meet the following conditions are extracted from the feature space corresponding to the key diagnostic nodes: the slice boundary contains the spatiotemporal coordinates of deformation-induced breakdown markers, and the slice data distribution conforms to the statistical law of energy flow conversion feature chain. The micro-feature slices are input into the fault factor distiller and subjected to three-order physical compression: first-order energy threshold filtering: removing feature points in the slice that are lower than the partial discharge initiation electric field; second-order causal chain encoding: rearranging the feature points into a decision sequence according to the spatiotemporal order of the energy flow conversion feature chain; third-order binary mapping: generating binary diagnostic factors based on the Boolean relationship between the spatial domain where the feature points are located and the material yield strength. The above scheme achieves precision and lightweight power equipment fault diagnosis through multi-level physical feature extraction and decision logic compression. High-sensitivity feature focusing and physical decision modeling: High-sensitivity subgraph extraction based on material mechanics threshold screening ensures feature selection aligns with equipment physical failure mechanisms. Causal decision manifold construction quantifies fault propagation paths into computable partial order relationships, giving decision tree splitting criteria clear physical interpretability. It achieves topological mapping from fault feature space to physical failure process, laying the foundation for physical consistency of the diagnostic model. Dynamic decision path optimization: Deeply differentiable decision trees are branched through physical constraints, simultaneously handling both elastic deformation warning and insulation breakdown judgment scenarios. Gradient flow tracing algorithm quantifies the diagnostic contribution of each node, and mutation detection locates safety boundary breach points. A dynamic correlation mechanism between fault evolution process and decision path is established, enabling adaptive capture of key risk nodes. Micro-feature distillation and physical compression: Spatiotemporal feature slicing ensures diagnostic factors contain complete fault development trajectory information. A three-order compression mechanism is implemented sequentially: Energy dimension: noise filtering based on discharge threshold; Temporal dimension: spatiotemporal encoding of energy flow causal chains; Logical dimension: Boolean abstraction of material failure states. This generates a binary diagnostic factor set that simultaneously retains essential physical features and minimal decision information.

[0093] In summary, this embodiment simplifies the multidimensional monitoring data stream into binary decision factors that conform to the physical failure patterns of equipment; eliminates the computational overhead of non-critical features through hierarchical filtering of physical constraints; achieves accurate identification of faults in their incipient stage based on feature encoding of material yield strength and energy flow conversion; and ensures that each decision node corresponds to a clearly defined threshold for equipment physical state transitions. Through feature space dimensionality reduction guided by physical mechanisms, end-to-end compression of equipment condition monitoring data to core decision logic is achieved, providing a lightweight solution for real-time fault diagnosis on embedded devices.

[0094] Example 7: Based on Example 6, the process of extracting the key decision path leading to gradient mutation in step S302 of this embodiment of the invention includes the following steps:

[0095] S3021: Inject the prediction error of the output layer of the deep differentiable diagnostic decision tree into the virtual deformation energy field generator, convert the difference in fault type labels into virtual deformation energy under the constraint of energy conservation, distribute the virtual deformation energy in reverse along the branch path of the deep differentiable diagnostic decision tree, and generate a fault energy density field covering all nodes of the decision tree.

[0096] S3022: Solve the energy density gradient at each decision node in the fault energy density field. The gradient direction indicates the fault energy transfer path, and the gradient magnitude characterizes the importance of node decision. Identify nodes where the energy density gradient exceeds the material's energy storage limit. Trigger a mutation flag when the absolute value of the energy density gradient is greater than half of the product of the yield strength and the yield strain.

[0097] S3023: Backtrack the fault energy transfer path from the mutation node along the negative gradient direction, and extract the nodes in the path whose splitting threshold satisfies any of the following conditions: the displacement gradient threshold is greater than the elastic limit of the winding material, and the local field strength threshold is greater than the critical breakdown field strength of the insulating medium; output the set of nodes that pass the verification as the critical decision path.

[0098] The working principle and beneficial effects of the above technical solution are as follows: First, the prediction error of the output layer of the deep differentiable diagnostic decision tree is injected into the virtual deformation energy field generator. The difference in fault type labels is converted into virtual deformation energy under energy conservation constraints. The virtual deformation energy is then distributed in reverse along the branch paths of the deep differentiable diagnostic decision tree to generate a fault energy density field covering all nodes of the decision tree. Second, the energy density gradient at each decision node is solved in the fault energy density field. The gradient direction indicates the fault energy transfer path, and the gradient magnitude characterizes the importance of the node decision. Nodes whose energy density gradient exceeds the material's energy storage limit are identified. A mutation marker is triggered when the absolute value of the energy density gradient is greater than half the product of the yield strength and the yield strain. Finally, the fault energy transfer path is traced back from the mutation node along the negative gradient direction. Nodes whose splitting threshold satisfies any of the following conditions are extracted: the displacement gradient threshold is greater than the elastic limit of the winding material, and the local field strength threshold is greater than the critical breakdown field strength of the insulating medium. The set of nodes that pass the output verification is the key decision path. The above scheme's technical process for extracting key decision paths achieves dynamic sensitivity analysis of the fault diagnosis decision tree through deep coupling of multimodal energy field mapping and mechanical constraints. By employing a prediction error-energy field conversion mechanism, discrete fault label differences are transformed into continuous energy density distributions, establishing a mathematical equivalence relationship between the decision tree topology and the fault energy transfer path. A deformation energy allocation algorithm under energy conservation constraints is used to achieve the quantitative transfer of fault characteristics between decision nodes, providing a differentiable field strength data foundation for subsequent gradient analysis. Based on gradient operator operations of the energy density field, a dual-channel evaluation system incorporating both directional and intensity information is constructed: the gradient direction field reveals the causal chain of fault propagation, while the gradient magnitude matrix quantifies the node decision weights. By introducing the yield strength criterion from materials mechanics, automatic detection of unstable points in the decision path is achieved; this criterion transforms statistical anomalies into physically interpretable energy mutation events. An inverse gradient tracing algorithm, combined with electromechanical composite failure conditions and dual thresholds of elastic limit / breakdown field strength, is used to reconstruct the causal chain from the mutation point to the root node. Parallel verification of displacement gradient and local field strength ensures that the extracted path simultaneously satisfies structural deformation constraints and electromagnetic performance constraints, forming a physically interpretable minimum sufficient diagnostic subgraph.

[0099] In summary, this embodiment transforms the discrete path search problem of traditional decision trees into an optimal transport problem within an energy field by introducing a continuum mechanics analysis framework. The coupled use of virtual deformation energy and material failure criteria gives gradient mutation detection a clear physical meaning. Furthermore, the backtracking mechanism based on multiphysics constraints ensures that the critical path reflects both statistical data characteristics and conforms to equipment failure mechanisms. The final output decision path exhibits both high sensitivity (capturing weak fault characteristics through energy gradients) and strong robustness (filtering noise interference through physical thresholds).

[0100] Example 8: Based on Example 7, the process of tracing back the fault energy transfer path from the mutation node along the negative gradient direction in step S3023 of this embodiment of the invention includes the following steps:

[0101] S30231: Starting from the mutation node, activate the fault energy transfer path tracking operator along the negative direction of the energy density gradient, and force the preset conditions to be met when traversing adjacent nodes to perform energy transfer continuity constraints.

[0102] S30232: Extract nodes in the fault energy transfer path that satisfy the displacement gradient threshold. The spatial coordinates of the nodes coincide with the deformation-induced breakdown marker. Extract nodes in the fault energy transfer path that satisfy the local field strength threshold. The local field strength threshold comes from the local field strength threshold electromagnetic energy equipotential surface boundary value. The node timestamp matches the peak time of the energy flow conversion feature chain.

[0103] S30233: Rearrange the verification nodes according to the direction of fault energy transmission, and assign each edge a multiple of the quantized path segment energy exceeding the breakdown threshold.

[0104] The working principle and beneficial effects of the above technical solution are as follows: First, starting from the mutation node, the fault energy transfer path tracking operator is activated along the negative direction of the energy density gradient. When traversing adjacent nodes, preset conditions are forcibly met to constrain energy transfer continuity. Second, nodes satisfying the displacement gradient threshold are extracted from the fault energy transfer path; the spatial coordinates of these nodes coincide with the deformation-induced breakdown marker. Nodes satisfying the local field strength threshold are also extracted from the fault energy transfer path; the electromagnetic energy equipotential surface from the local field strength threshold is the boundary value, and the node timestamp matches the peak time of the energy flow conversion characteristic chain. Finally, the verification nodes are rearranged according to the fault energy transfer direction, and each edge is assigned a multiple by which the quantized path segment energy exceeds the breakdown threshold. This solution forms a three-dimensional verification system of gradient field-equipotential surface-temporal chain, reducing the false positive rate. Path energy quantization enables breakdown risk level prediction, improving accuracy. It supports multi-physics coupling analysis (electromagnetic-mechanical-thermodynamic) and can identify the propagation path of complex faults. It is particularly suitable for fault diagnosis scenarios with energy gradient characteristics, such as partial discharge in power equipment and fatigue cracks in mechanical structures. Through dynamic adaptation of characteristic parameters, it can be extended to early fault warning systems for new energy equipment.

[0105] Example 9: Based on Example 8, the process of extracting nodes that satisfy the local field strength threshold in step S30232 of the fault energy transfer path provided in this embodiment of the invention includes the following steps:

[0106] S302321: Extract the position-dependent critical breakdown field strength from the generated electromagnetic energy equipotential surface; dynamically generate time-varying field strength thresholds at each node of the fault energy transfer path;

[0107] S302322: Obtain the peak moment of discharge energy conversion rate of the energy flow conversion feature chain at the node spatial coordinates, only if the path node simultaneously satisfies the decision tree splitting threshold and the node timestamp.

[0108] S302323: Obtain node puncture risk factors, bind node puncture risk factors to spatial coordinates, and generate insulation failure early warning nodes carrying spatiotemporal risk identifiers.

[0109] The working principle and beneficial effects of the above technical solution are as follows: This embodiment first extracts the position-related critical breakdown field strength from the generated electromagnetic energy equipotential surface; dynamically generates time-varying field strength thresholds at each node of the fault energy transfer path; it obtains the peak moment of discharge energy conversion rate of the energy flow conversion characteristic chain at the node spatial coordinates, only when the path node simultaneously satisfies the node decision tree splitting threshold and the node timestamp; finally, it obtains the node breakdown risk factor, binds the node breakdown risk factor to the spatial coordinates, and generates an insulation failure early warning node carrying a spatiotemporal risk identifier. The dynamic threshold system of the above solution controls the field strength assessment error within ±3% compared to the static threshold; the spatiotemporal dual verification reduces the false alarm rate; the risk factor model can predict insulation failure 5-8 sampling periods in advance; it supports multi-dimensional risk visualization; it is suitable for rapidly developing insulation fault diagnosis such as partial discharge of GIS equipment and inter-turn short circuit of transformers; by introducing material aging factors and spatial attenuation coefficients, it can adapt to the risk assessment needs under different operating conditions.

[0110] Example 10: Based on Example 9, the process of obtaining the node penetration risk factor in step S302323 of this embodiment of the invention includes the following steps:

[0111] S3023231: Input the decision tree splitting threshold at the node into the effective insulation barrier calculator, call the critical breakdown field strength of the electromagnetic energy equipotential surface corresponding to the node's spatial coordinates; introduce the discharge energy conversion rate pulse waveform of the differential characteristics of the energy flow conversion characteristic chain at the peak moment; output the dynamic reference field strength margin modulated by the pulse waveform.

[0112] S3023232: Calculate the absolute difference between the decision threshold and the dynamic reference field strength, obtain the discharge energy conversion rate pulse phase with the neighborhood derivative sign at the peak time, mark the energy accumulation period with the positive phase and the energy release period with the negative phase;

[0113] S3023233: Scalar risk value synthesis yields node breakdown risk factor; Spatiotemporal feature coding: a node breakdown risk factor greater than 0 indicates insufficient insulation margin during the energy accumulation period, a node breakdown risk factor not greater than 0 indicates breakdown backstroke effect during the energy release period, and the node breakdown risk factor quantifies the risk intensity level.

[0114] The working principle and beneficial effects of the above technical solution are as follows: In this embodiment, the decision tree splitting threshold at the node is first input into the effective insulation barrier calculator, and the critical breakdown field strength of the electromagnetic energy equipotential surface corresponding to the node spatial coordinates is called; the discharge energy conversion rate pulse waveform of the differential characteristics of the energy flow conversion characteristic chain at the peak moment is introduced; the dynamic reference field strength margin modulated by the pulse waveform is output; secondly, the absolute difference between the decision threshold and the dynamic reference field strength is calculated, and the discharge energy conversion rate pulse phase with the neighborhood derivative sign at the peak moment is obtained. The positive phase marks the energy accumulation period, and the negative phase marks the energy release period; finally, the scalar risk value is synthesized to obtain the node breakdown risk factor; spatiotemporal feature encoding: the node breakdown risk factor greater than 0 indicates insufficient insulation margin during the energy accumulation period, and the node breakdown risk factor not greater than 0 indicates breakdown backstroke effect during the energy release period. The node breakdown risk factor quantifies the risk intensity level. The above scheme achieves dynamic baseline correction of insulation performance, real-time identification of discharge energy phase, and quantitative output of risk status through triple coupling analysis of field strength, energy, and phase. The key technological breakthrough lies in transforming the static breakdown criterion into a dynamic risk assessment model that includes a time dimension. Through cross-validation of electromagnetic energy conversion rate pulse characteristics and spatial field strength distribution, an early warning system for insulation failure with predictive capabilities has been established.

[0115] Example 11: Based on Example 10, the scalar risk value synthesis process in step S3023233 of this embodiment of the invention includes the following steps:

[0116] S30232331: The absolute difference between the decision threshold and the dynamic reference field strength is processed by a spatial field strength discretization processor. The intrinsic relaxation time of the insulating medium is used as the integral window constraint. Only the field strength offset within the complete relaxation cycle before the peak time is counted to extract the effective field strength offset that leads to insulation degradation. The physical basis of the integral window is the dielectric recovery characteristic curve of different insulating materials.

[0117] S30232332: Employs a phase control symbol injection mechanism to input the discharge energy conversion rate pulse phase symbol function into the risk polarity modulator; when a positive phase is detected, it outputs a +1 polarity factor, corresponding to the space charge accumulation effect in the energy accumulation stage; when a negative phase is detected, it outputs a -1 polarity factor, reflecting the collision ionization process in the energy release stage; through directional coupling of the insulating material normal vector, it generates a risk reference vector with spatial orientation characteristics, with its direction along the maximum gradient direction of the equipotential surface of the insulating medium, pointing to the weakest dimension of the insulation;

[0118] S30232333: Normalization is performed based on the nominal breakdown field strength of the insulating medium to make the risk value comparable across media; by calculating the ratio of the risk benchmark vector magnitude to the nominal breakdown field strength and retaining the insulation normal projection sign, the final output scalar risk value has clear physical threshold characteristics: when the absolute value of the scalar risk value reaches 1, it indicates that the material's theoretical breakdown critical point has been reached, 0.7 is the warning threshold, and a negative value represents the reverse energy impact effect; scaling conversion enables the insulation status assessment of equipment of different voltage levels to have a unified quantitative benchmark.

[0119] The working principle and beneficial effects of the above technical solution are as follows: This embodiment forms an insulation state feature code with a strict physical correspondence, providing a quantitative criterion for the subsequent diagnostic system that reflects both the instantaneous state and contains trend information.

[0120] Example 12: As Figure 6 As shown, based on Embodiments 1-11, the intelligent transformer fault diagnosis system based on deep convolutional networks provided in this embodiment of the invention includes:

[0121] The multimodal feature generation module is used to capture the full-domain transient signals of the transformer through a spatially discretized heterogeneous physical field synchronous sampling array. It inputs three types of heterogeneous data streams—vibration spectrum, electromagnetic leakage wave, and partial discharge pulse—into a multi-scale feature decoupling program. The multi-scale feature decoupling program uses an asymmetric deep convolution kernel group to perform feature stripping along the three-dimensional directions of time, space, and frequency domain, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds to form the initial feature primitive set for fault diagnosis.

[0122] The adaptive fusion module is used to input the generated vibration mode tensor, electromagnetic eddy current feature spectrum, and discharge pulse feature cloud into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression of redundant features based on a dynamic decay function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types.

[0123] The lightweight diagnostic module is used to train a deep differentiable diagnostic decision tree using a fault-sensitive feature topology graph, and to locate key diagnostic nodes through backpropagation path saliency analysis. The micro-feature slices corresponding to the key diagnostic nodes are extracted and input into the fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is then reconstructed into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information.

[0124] The working principle and beneficial effects of the above technical solution are as follows: The multi-modal feature generation module in this embodiment is used to capture the full-domain transient signal of the transformer through a spatially discretized heterogeneous physical field synchronous sampling array, and inputs three types of heterogeneous data streams—vibration spectrum, electromagnetic leakage wave, and partial discharge pulse—into a multi-scale feature decoupling program; the multi-scale feature decoupling program uses an asymmetric deep convolution kernel group to perform feature stripping along the three-dimensional directions of time, space, and frequency domain, respectively, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds to form an initial feature primitive set for fault diagnosis; the adaptive fusion module is used to input the generated vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds into a feature interaction contribution analysis program; the feature interaction contribution analysis program calculates each feature primitive in real time through a bidirectional feature entropy weight evaluation mechanism. The information contribution intensity in the decision space is evaluated, and redundant features are nonlinearly suppressed based on a dynamic decay function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly represent the correlation strength between feature primitives and fault types. The lightweight diagnostic module is used to train a deep differentiable diagnostic decision tree using the fault-sensitive feature topology map and locate key diagnostic nodes through backpropagation path saliency analysis. The micro-feature slices corresponding to the key diagnostic nodes are extracted and input into the fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is restored to a lightweight diagnostic model in real time on the terminal device through the edge-end dynamic reconstruction engine, realizing millisecond-level synchronous output of fault type and location information (the specific principle is as follows). Figure 7 (As shown). The multimodal feature generation module of the above scheme constructs a three-dimensional feature decoupling system of time-space-frequency, transforming three types of heterogeneous signals—vibration, electromagnetic, and discharge—into computable tensor / feature spectrum / feature cloud forms, providing a physically meaningful initial set of feature primitives for fault diagnosis. The adaptive fusion module achieves quantitative screening of multi-source features through entropy weight evaluation and dynamic attenuation mechanisms, generating a fault-sensitive feature topology map with weighted identifiers, and establishing a quantifiable correlation model between feature primitives and fault types. The lightweight diagnosis module employs knowledge distillation and dynamic reconstruction techniques to compress complex decision logic into a binary diagnostic factor set, achieving millisecond-level inference at the edge while retaining core diagnostic capabilities.

[0125] In summary, the three modules in this embodiment work together to form a complete technology chain, realizing fully automated processing from raw signal acquisition, multi-dimensional feature extraction, intelligent feature filtering, and lightweight decision output, ultimately achieving high-precision real-time diagnosis of transformer fault types and location information.

[0126] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of equivalents of this invention, this invention is also intended to include these modifications and variations.

Claims

1. A method for intelligent fault diagnosis of transformers based on deep convolutional networks, characterized in that, Includes the following steps: A deep differentiable diagnostic decision tree is trained using a fault-sensitive feature topology graph, and key diagnostic nodes are located through backpropagation path saliency analysis. Micro-feature slices corresponding to key diagnostic nodes are extracted and input into a fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is then reconstructed into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information. The process of generating a binary diagnostic factor set containing only the core decision-making logic includes the following steps: Based on the fault association strength identified by the edge weights of the fault-sensitive feature topology graph, highly sensitive feature subgraphs with weight values ​​exceeding the material yield strength threshold are extracted. A causal decision manifold based on the fault energy transfer path is established between the nodes of the highly sensitive feature subgraphs, and the physical decision partial order relationship of each feature primitive on the output manifold is used as the decision tree splitting criterion. The initial set of feature primitives is input into the deep differentiable diagnostic decision tree. Its splitting nodes are dynamically adjusted according to the physical decision partial order relationship. Each decision node corresponds to: the feature subset that satisfies the material elastic deformation constraint on the left branch, and the feature subset that triggers the insulation breakdown condition on the right branch. Fault type labels are injected into the output of the deep differentiable diagnostic decision tree. The contribution of each decision node to the diagnostic result is calculated by the gradient flow tracing algorithm. The key decision path that causes gradient mutation is extracted, and the key diagnostic nodes in the path that exceed the equipment safety boundary are located. From the feature space corresponding to the key diagnostic node, extract feature slices that simultaneously meet the following conditions: the slice boundary contains the spatiotemporal coordinates of the deformation-induced breakdown marker, and the slice data distribution conforms to the statistical law of the energy flow conversion feature chain; input the micro-feature slices into the fault factor distiller and compress them through third-order physical compression. First-order energy threshold filtering: Removes feature points in the slice that are below the partial discharge initiation electric field; Second-order causal chain encoding: Rearranges the feature points into a decision sequence according to the spatiotemporal order of the energy flow conversion feature chain; Third-order binary mapping: Generates binary diagnostic factors based on the Boolean relationship between the spatial domain where the feature point is located and the material yield strength. The process of extracting the critical decision path that leads to gradient mutation includes the following steps: The prediction error of the output layer of the deep differentiable diagnostic decision tree is injected into the virtual deformation energy field generator. The difference in fault type labels is converted into virtual deformation energy under the constraint of energy conservation. The virtual deformation energy is distributed in reverse along the branch path of the deep differentiable diagnostic decision tree to generate a fault energy density field covering all nodes of the decision tree. In the fault energy density field, the energy density gradient at each decision node is solved. The gradient direction indicates the fault energy transfer path, and the gradient magnitude characterizes the importance of the node decision. Nodes whose energy density gradient exceeds the material's energy storage limit are identified. When the absolute value of the energy density gradient is greater than half of the product of the yield strength and the yield strain, a mutation marker is triggered. Tracing back the fault energy transfer path from the mutation node along the negative gradient direction, extracting nodes in the path whose splitting threshold satisfies any of the following conditions: the displacement gradient threshold is greater than the elastic limit of the winding material, or the local field strength threshold is greater than the critical breakdown field strength of the insulating medium; outputting the set of nodes that pass the verification is the critical decision path.

2. The intelligent transformer fault diagnosis method based on deep convolutional networks as described in claim 1, characterized in that, The process of tracing back the fault energy transfer path from the mutation node along the negative gradient direction includes the following steps: Starting from the mutation node, the fault energy transfer path tracking operator is activated along the negative direction of the energy density gradient. When traversing adjacent nodes, the preset conditions are forced to be met to constrain the continuity of energy transfer. Extract nodes that satisfy the displacement gradient threshold in the fault energy transfer path, and the spatial coordinates of the nodes coincide with the deformation-induced breakdown marker; extract nodes that satisfy the local field strength threshold in the fault energy transfer path, and the electromagnetic energy equipotential surface from the local field strength threshold reaches the boundary value, and the node timestamp matches the peak time of the energy flow conversion characteristic chain. The verification nodes are rearranged according to the direction of fault energy transmission, and each edge is assigned a multiple by which the energy of the quantized path segment exceeds the breakdown threshold.

3. The intelligent transformer fault diagnosis method based on deep convolutional networks as described in claim 2, characterized in that, The process of extracting nodes that satisfy the local field strength threshold in the fault energy transfer path includes the following steps: Extract the position-dependent critical breakdown field strength from the generated electromagnetic energy equipotential surface; dynamically generate time-varying field strength thresholds at each node of the fault energy transfer path. The peak moment of discharge energy conversion rate of the energy flow conversion feature chain at the node spatial coordinates is obtained only if the path node simultaneously satisfies the decision tree splitting threshold and the node timestamp. Obtain node puncture risk factors, bind node puncture risk factors to spatial coordinates, and generate insulation failure early warning nodes carrying spatiotemporal risk identifiers.

4. The intelligent transformer fault diagnosis method based on deep convolutional networks as described in claim 3, characterized in that, The process of obtaining node penetration risk factors includes the following steps: Input the decision tree splitting threshold at the node into the effective insulation barrier calculator, call the critical breakdown field strength of the electromagnetic energy equipotential surface corresponding to the node's spatial coordinates; introduce the discharge energy conversion rate pulse waveform of the differential characteristics of the energy flow conversion characteristic chain at the peak moment; Outputs the dynamic reference field strength margin modulated by pulse waveform; Calculate the absolute difference between the decision threshold and the dynamic reference field strength, obtain the discharge energy conversion rate pulse phase with the neighborhood derivative sign at the peak time, mark the energy accumulation period with the positive phase and the energy release period with the negative phase; Scalar risk values ​​are synthesized to obtain node penetration risk factors; Spatiotemporal feature coding: A node breakdown risk factor greater than 0 indicates insufficient insulation margin during the energy accumulation period; a node breakdown risk factor not greater than 0 indicates breakdown backstroke effect during the energy release period; the node breakdown risk factor quantifies the risk intensity level.

5. The intelligent transformer fault diagnosis method based on deep convolutional networks as described in claim 1, characterized in that, The entire transient signal of the transformer is captured by a spatially discretized heterogeneous physical field synchronous sampling array, and the three heterogeneous data streams of vibration spectrum, electromagnetic leakage wave and partial discharge pulse are input into a multi-scale feature decoupling program. The multi-scale feature decoupling program uses asymmetric deep convolution kernel groups to strip features along the three-dimensional directions of time, space, and frequency, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds to form an initial feature primitive set for fault diagnosis.

6. The intelligent transformer fault diagnosis method based on deep convolutional networks as described in claim 5, characterized in that, The generated vibration mode tensor, electromagnetic eddy current characteristic spectrum, and discharge pulse characteristic cloud are input into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression of redundant features based on a dynamic attenuation function. The calibration results of the information contribution intensity drive the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types.

7. A transformer fault intelligent diagnosis system based on deep convolutional networks, used to implement the transformer fault intelligent diagnosis method based on deep convolutional networks as described in any one of claims 1-6, characterized in that, Include: The multimodal feature generation module is used to capture the full-domain transient signal of the transformer through a spatially discretized heterogeneous physical field synchronous sampling array, and inputs three types of heterogeneous data streams, namely vibration spectrum, electromagnetic leakage wave and partial discharge pulse, into the multi-scale feature decoupling program; The multi-scale feature decoupling procedure uses asymmetric deep convolution kernel groups to strip features along the three-dimensional directions of time, space, and frequency domain, and outputs physically meaningful vibration mode tensors, electromagnetic eddy current feature spectra, and discharge pulse feature clouds to form an initial feature primitive set for fault diagnosis. The adaptive fusion module is used to input the generated vibration mode tensor, electromagnetic eddy current feature spectrum, and discharge pulse feature cloud into the feature interaction contribution analysis program. The feature interaction contribution analysis program calculates the information contribution intensity of each feature element in the decision space in real time through a two-way feature entropy weight evaluation mechanism, and performs nonlinear suppression of redundant features based on a dynamic decay function. The calibration result of the information contribution intensity drives the cross-modal adaptive fusion gate to generate a fault-sensitive feature topology map with weight labels. Its edge weights directly characterize the correlation strength between feature elements and fault types. The lightweight diagnostic module is used to train a deep differentiable diagnostic decision tree using a fault-sensitive feature topology graph, and to locate key diagnostic nodes through backpropagation path saliency analysis. The micro-feature slices corresponding to the key diagnostic nodes are extracted and input into the fault factor distiller. An iterative knowledge compression algorithm generates a binary diagnostic factor set containing only the core decision logic. The binary diagnostic factor set is then reconstructed into a lightweight diagnostic model in real time on the terminal device through an edge-end dynamic reconstruction engine, achieving millisecond-level synchronous output of fault type and location information.

Images