Transformer oil flow anomaly detection method, device and equipment

Abstract

The invention discloses a transformer oil flow anomaly detection method, device and equipment, and the method comprises the steps: collecting multi-modal data in the operation process of a transformer, and determining a flow velocity gradient matrix, a pressure fluctuation matrix and a temperature field through a mode of reconstructing a fluid field; performing feature fusion on the flow velocity gradient matrix, the pressure fluctuation matrix, the temperature field and the vibration spectrum through a preset time-space correlation model to obtain a fusion feature vector; performing anomaly analysis processing based on reconstruction comparison on the fusion feature vector by adopting a preset VAE, and determining an anomaly score; and performing anomaly detection analysis based on a dynamic threshold value and feedback optimization according to the anomaly score and the load power to obtain an anomaly detection result. The technical problems that the transformer oil flow anomaly detection efficiency and accuracy in an actual scene are low and application requirements are difficult to meet due to the fact that the prior art depends on a fixed threshold value and manual detection and it is difficult to capture multi-physical field coupling features and time-space correlation features of oil flow anomaly can be solved.

Description

Technical Field

[0001] This application relates to the field of transformer anomaly monitoring, and in particular to a method, device and equipment for detecting transformer oil flow anomalies. Background Technology

[0002] As a core piece of equipment in the power system, transformer oil flow anomaly detection technology mainly relies on fixed threshold method and single-mode sensor analysis, such as electromagnetic flowmeter or temperature sensor. However, the anomaly detection technology based on this still has obvious shortcomings.

[0003] Using preset fixed thresholds cannot handle dynamic operating conditions such as transformer load fluctuations and seasonal temperature changes, leading to frequent false alarms or missed alarms. Furthermore, existing technologies, mostly based on single sensors, struggle to capture the multi-physics coupling characteristics of oil flow anomalies. Additionally, most existing algorithms neglect spatiotemporal correlations and are highly dependent on human intervention; in particular, threshold calibration and fault diagnosis rely heavily on expert experience, making them ill-suited to the complex operating conditions of modern transformers. These technical shortcomings render existing transformer oil flow anomaly detection technologies inadequate for meeting the detection needs of practical applications. Summary of the Invention

[0004] This application provides a method, apparatus, and equipment for detecting abnormal transformer oil flow, which solves the technical problem that existing technologies rely on fixed thresholds and manual detection, and are difficult to capture the multi-physics coupling characteristics and spatiotemporal correlation characteristics of abnormal oil flow, resulting in low efficiency and accuracy of abnormal transformer oil flow in actual scenarios, making it difficult to meet application requirements.

[0005] In view of this, the first aspect of this application provides a method for detecting abnormal transformer oil flow, comprising: Multimodal data during transformer operation are collected, and the velocity gradient matrix, pressure fluctuation matrix, and temperature field are determined by reconstructing the fluid field. The multimodal data includes oil flow velocity, oil temperature distribution, and vibration spectrum. The velocity gradient matrix, pressure fluctuation matrix, temperature field, and vibration spectrum are fused using a preset spatiotemporal correlation model to obtain a fused feature vector. An anomaly analysis based on reconstruction comparison is performed on the fused feature vector using a preset VAE to determine the anomaly score; Anomaly detection analysis based on dynamic threshold and feedback optimization is performed according to the anomaly score and load power to obtain anomaly detection results, which include anomaly location and alarm signals.

[0006] Preferably, the step of acquiring multimodal data during transformer operation and determining the velocity gradient matrix, pressure fluctuation matrix, and temperature field by reconstructing the fluid field includes: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, the oil temperature distribution, and the vibration spectrum are standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and the oil temperature distribution. The fluid field is subjected to simplified difference calculation based on the central difference method to obtain the velocity gradient matrix, pressure fluctuation matrix and temperature field.

[0007] Preferably, the step of fusing features of the flow velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum through a preset spatiotemporal correlation model to obtain a fused feature vector includes: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a preset spatiotemporal correlation model is constructed through graph convolution calculation and LSTM temporal calculation. The velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum are input into the preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

[0008] Preferably, the step of using a preset VAE to perform anomaly analysis on the fused feature vector based on reconstruction comparison to determine anomaly scores includes: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on a preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on the encoded latent variables, the reconstruction error, and the preset KL divergence to determine the anomaly score, KL divergence loss, and update the latent variables.

[0009] Preferably, the step of performing anomaly detection analysis based on dynamic thresholds and feedback optimization according to the anomaly score and load power to obtain anomaly detection results includes: A threshold update formula is constructed based on the KL divergence loss, the updated latent variables, and the load power, and feedback optimization calculation is performed to obtain the dynamic threshold. If the anomaly score exceeds the dynamic threshold, an alarm signal is generated, and the anomaly is located based on the flow velocity gradient to obtain the anomaly detection result.

[0010] A second aspect of this application provides a transformer oil flow anomaly detection device, comprising: The reconstruction calculation unit is used to collect multimodal data during transformer operation and determine the velocity gradient matrix, pressure fluctuation matrix and temperature field by reconstructing the fluid field. The multimodal data includes oil flow velocity, oil temperature distribution and vibration spectrum. The feature fusion unit is used to fuse the flow velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum through a preset spatiotemporal correlation model to obtain a fused feature vector; An anomaly analysis unit is used to perform anomaly analysis processing on the fused feature vector based on reconstruction comparison using a preset VAE to determine anomaly scores; An anomaly detection unit is used to perform anomaly detection analysis based on dynamic threshold and feedback optimization according to the anomaly score and load power, and obtain anomaly detection results, which include anomaly location and alarm signals.

[0011] Preferably, the reconstruction computing unit is specifically used for: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, the oil temperature distribution, and the vibration spectrum are standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and the oil temperature distribution. The fluid field is subjected to simplified difference calculation based on the central difference method to obtain the velocity gradient matrix, pressure fluctuation matrix and temperature field.

[0012] Preferably, the feature fusion unit is specifically used for: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a preset spatiotemporal correlation model is constructed through graph convolution calculation and LSTM temporal calculation. The velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum are input into the preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

[0013] Preferably, the anomaly analysis unit is specifically used for: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on a preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on the encoded latent variables, the reconstruction error, and the preset KL divergence to determine the anomaly score, KL divergence loss, and update the latent variables.

[0014] A third aspect of this application provides a transformer oil flow anomaly detection device, the device including a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the transformer oil flow anomaly detection method described in the first aspect according to the instructions in the program code.

[0015] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This application provides a method for detecting abnormal oil flow in transformers, comprising: acquiring multimodal data during transformer operation, and determining the velocity gradient matrix, pressure fluctuation matrix, and temperature field by reconstructing the fluid field; the multimodal data including oil flow velocity, oil temperature distribution, and vibration spectrum; performing feature fusion on the velocity gradient matrix, pressure fluctuation matrix, temperature field, and vibration spectrum using a preset spatiotemporal correlation model to obtain a fused feature vector; performing anomaly analysis processing on the fused feature vector based on reconstruction comparison using a preset VAE to determine anomaly scores; and performing anomaly detection analysis based on dynamic thresholds and feedback optimization according to the anomaly scores and load power to obtain anomaly detection results, which include anomaly location and alarm signals.

[0016] The transformer oil flow anomaly detection method provided in this application constructs a temperature field using multimodal data, performs feature fusion based on a spatiotemporal graph network, and conducts anomaly analysis based on a preset VAE. This enables adaptive dynamic anomaly detection without relying on manual intervention. Furthermore, this process jointly analyzes the spatiotemporal coupling characteristics of the oil flow velocity field, temperature field, and vibration spectrum, improving anomaly detection sensitivity and efficiency. It simultaneously considers the coupling characteristics analysis of multiple physics fields and spatiotemporal correlation analysis, thereby ensuring the accuracy and reliability of the detection results. Therefore, this application solves the technical problem of existing technologies relying on fixed thresholds and manual detection, which struggles to capture the multiphysics coupling characteristics and spatiotemporal correlation characteristics of oil flow anomalies, resulting in low efficiency and accuracy in transformer oil flow anomaly detection in practical scenarios, failing to meet application requirements. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for detecting abnormal transformer oil flow provided in an embodiment of this application; Figure 2 This is a schematic diagram of a transformer oil flow anomaly detection device provided in an embodiment of this application. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0019] For easier understanding, please refer to Figure 1 An embodiment of a transformer oil flow anomaly detection method provided in this application includes: Step 101: Collect multimodal data during transformer operation and determine the velocity gradient matrix, pressure fluctuation matrix, and temperature field by reconstructing the fluid field. The multimodal data includes oil flow velocity, oil temperature distribution, and vibration spectrum.

[0020] Further, step 101 includes: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, oil temperature distribution, and vibration spectrum were standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and oil temperature distribution. A simplified difference calculation based on the central difference method is performed on the fluid field to obtain the velocity gradient matrix, pressure fluctuation matrix, and temperature field.

[0021] It should be noted that electromagnetic flowmeters can obtain oil flow velocity. Distributed fiber optic temperature sensors extract oil temperature distribution. ,in, , , , These represent the horizontal, vertical, and axial coordinates of the oil temperature distribution space, as well as the time point, respectively, while the MEMS vibration sensor can collect the vibration spectrum. ;in, Indicates frequency.

[0022] In order to perform comprehensive analysis and processing of heterogeneous data, this embodiment will also standardize the data:

[0023] in, The mean, Standard deviation The data before standardization can be any of the following: oil flow velocity, oil temperature distribution, or vibration spectrum. This is the data obtained after standardization.

[0024] The parameters used to reconstruct the fluid field are the coefficient of thermal expansion, the normalized oil flow velocity, and the oil temperature distribution. It is generated using the unsteady Navier-Stokes equations, which can be specifically expressed as:

[0025] in, The velocity vector of the oil flow. It is the gravitational acceleration vector. For oil density, The coefficient of thermal expansion is 1 / 3. For localized temperature rise, Represents the vector differential operator. For hydrostatic pressure, This represents the kinematic viscosity. The reconstructed fluid field in this embodiment includes a term based on the thermal expansion effect, namely... This item models the buoyancy effect caused by the temperature gradient due to the thermal expansion effect; introducing the temperature expansion coefficient can realize thermodynamic-fluid dynamic coupling, which is a comprehensive analysis of multiple physics fields.

[0026] Generally, under conditions of low oil flow velocity, i.e., transformer oil flow velocity less than 1 m / s, the convection phase... The impact on the fluid field is minimal and can be ignored. The ingress velocity gradient can be calculated using the central difference method.

[0027] in, , For grid spacing, , For grid indexing.

[0028] Then the discretized reconstructed fluid field can be obtained:

[0029] The velocity gradient matrix, pressure fluctuation matrix, and temperature field can be obtained from the reconstructed fluid field, and are respectively expressed as: , , ;in, This indicates the number of discrete points in space.

[0030] Step 102: Use a preset spatiotemporal correlation model to fuse the velocity gradient matrix, pressure fluctuation matrix, temperature field and vibration spectrum to obtain a fused feature vector.

[0031] Further, step 102 includes: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a pre-defined spatiotemporal correlation model is constructed through graph convolution computation and LSTM temporal computation. The velocity gradient matrix, pressure fluctuation matrix, temperature field, and vibration spectrum are input into a preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

[0032] Feature fusion is implemented based on a pre-defined spatiotemporal correlation model, specifically by fusing the flow velocity gradient matrix. Pressure fluctuation matrix Temperature field and vibration spectrum Data from heterogeneous sensors are uniformly encoded into spatiotemporal correlation features. Furthermore, the graph structure can capture the physical connections and temporal dynamics between oil passage nodes, thereby helping to identify local anomalies such as blockages and eddies.

[0033] The pre-defined spatiotemporal correlation model is a graph structure model. Based on graph convolutional encoding, the topology of the transformer oil passages can be modeled, resulting in an oil passage topology graph model. On this basis, graph convolutional computation and LSTM time-series computation can be used to jointly capture time-varying features such as flow pulsation and temperature drift, thereby constructing the pre-defined spatiotemporal correlation model. This embodiment combines a dynamic graph structure with normalized spatial convolution, which can significantly improve the detection accuracy of local anomalies in the oil passages, with an actual improvement of approximately 22%.

[0034] Specifically, the transformer oil duct diagram construction process includes defining the coordinate mapping positions of sensors as nodes, and using the similarity calculated based on oil flow velocity similarity as edge weights:

[0035] in, The edge weight represents the similarity of oil flow velocities between two nodes. The larger the edge weight, the more synchronized the fluid motion. , Representing nodes respectively and nodes The oil flow velocity vector is Three-dimensional components ,like ,but ,if and If the difference is large, then If a section of the oil passage is blocked, then It will suddenly drop. Related features will also decrease, as graph convolution automatically weakens the feature propagation of abnormal nodes; generally, the oil flow velocity fluctuation is small in the straight section of the main oil channel, while the oil flow velocity fluctuation is large in the vortex region of the bend. The normalization coefficient is typically the standard deviation of the velocity differences across all nodes.

[0036] in, This represents the number of samples.

[0037] The graph convolution calculation in the model can be expressed as:

[0038] in, It is an adjacency matrix with self-connections. The edge weights directly constitute the non-zero elements of the adjacency matrix. This determines the aggregation range of node features; for example, high-weight edges will enhance the feature transfer between adjacent nodes. For degree matrix, Represents the weighted number of connections between nodes, an off-diagonal matrix. ,Right now ; For the first l Spatial pathway characteristics of the layer For the first l The spatial convolution weights of a layer are trainable weight matrices; This is the activation function.

[0039] The LSTM timing computation process of the model in this embodiment is a computation process combining LSTM and Conv1D, which can be specifically expressed as follows:

[0040] in, For the first l The temporal characteristics of the layer For feature concatenation, the dimension of the concatenated feature vector is the sum of the dimensions of the two original vectors. For example, if Conv1D outputs 64 dimensions and LSTM outputs 64 dimensions, the concatenated output is 128 dimensions. Here, the 64-dimensional Conv1D output consists of 64 convolutional kernels, each sliding along the time axis to extract local features; the 64-dimensional LSTM output is the same time-series data as Conv1D. 64 dimensions are sufficient to encode complex spatiotemporal features. In practical engineering, it has been found that with dimensions below 32, the F1 score decreases by 15%, and with dimensions above 128, the risk of overfitting increases. In GPU parallel computing, multiples of 64 can optimize memory alignment and improve computation speed by more than 20%.

[0041] The feature fusion calculation process is as follows:

[0042] in, , They represent the first l Spatial pathway characteristics and temporal characteristics of the layer splicing operation, This is the fused feature vector.

[0043] Step 103: Use a preset VAE to perform anomaly analysis on the fused feature vector based on reconstruction comparison, and determine the anomaly score.

[0044] Further, step 103 includes: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on the preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on encoded latent variables, reconstruction error, and preset KL divergence to determine anomaly scores, KL divergence loss, and updated latent variables.

[0045] The degree of anomaly is quantified by comparing the fused feature vector with the VAE reconstructed vector. VAE can learn the data distribution under normal operating conditions and is particularly sensitive to unknown anomalies, such as localized blockage of oil flow. Deviating from the normal distribution, VAE jointly labels anomalies using reconstruction error E and latent space KL divergence. Furthermore, this process can be used for feature decoupling; its latent variables can compress high-dimensional features into a low-dimensional latent space, achieving separation of the coupling effects of modes such as oil flow velocity, temperature, and vibration, thereby improving anomaly localization accuracy. In practical engineering measurements, the accuracy can reach [percentage missing]. Within.

[0046] Specifically, the latent variables are encoded as follows:

[0047] in, The Gaussian distribution function is used to force the latent variables. Obey structured priors, H Input the feature vector to the current layer. , These are the mean and variance of the encoder output, respectively. This process involves converting the input feature vector... H Mapped into the latent space.

[0048] The reconstruction error calculation process is as follows:

[0049] in, This represents the reconstructed features obtained from the latent variable space. These are the decoder parameters; the ability of a VAE to reconstruct the input is measured by how outlier samples, due to their distribution, can cause reconstruction errors. ESignificantly increased.

[0050] The process of performing anomaly analysis and calculating anomaly scores based on encoded latent variables, reconstruction errors, and preset KL divergence is expressed as follows:

[0051] in, To balance the weights, a value of 0.5 is typically used; KL divergence is used to measure... and differences For encoder parameters, Given the standard normal prior probability, the consistency of the KL term on the right-hand side of the joint optimization formula with the second term in the latent space is... Input fidelity.

[0052] The optimization loss function constructed based on anomaly scores can be expressed as:

[0053] Among them, maximizing the reconstruction term can force the encoder to reconstruct the input data as accurately as possible, thereby ensuring that it can sensitively capture small deviations in the input features during anomaly detection; This indicates the distributed nature of the encoder output. Distribution The expectation, that is, for all possible latent variables Find the mean; In other words, the posterior probability distribution defined by the encoder represents the input data. The probability distribution mapped to the latent space is usually assumed to be a Gaussian distribution. , , For encoders based on input data The output means and standard deviation; The likelihood distribution is defined by the decoder, representing the distribution from the latent variables. Reconstructing data The probability is usually assumed to follow a Gaussian or Bernoulli distribution. Log-likelihood is used to measure the degree of matching between the reconstructed data and the original data.

[0054] Minimizing the regularization term can avoid excessive complexity in the latent space and ensure the model's generalization ability; in the anomaly detection process, normal data... We should closely follow the prior distribution, while outlier data... It will deviate significantly. This solution process can yield outlier scores. It can also determine the KL divergence loss. and updating latent variables .

[0055] Understandably, the higher the anomaly score, the more severe the anomaly. Inputting the anomaly detection results into the transformer cooling system controller allows for real-time adjustment of the pump speed and accumulation of potential variables. z Historical offsets are used to construct equipment degradation indicators.

[0056] Step 104: Perform anomaly detection analysis based on dynamic threshold and feedback optimization according to the anomaly score and load power to obtain anomaly detection results, including anomaly location and alarm signals.

[0057] Further, step 104 includes: A threshold update formula is constructed based on KL divergence loss, updated latent variables, and load power, and feedback optimization calculation is performed to obtain the dynamic threshold. If the abnormal score exceeds the dynamic threshold, an alarm signal is triggered, and the abnormality is located based on the flow velocity gradient to obtain the abnormality detection result.

[0058] This embodiment achieves dynamic anomaly detection based on dynamic thresholds and a feedback mechanism, by adjusting the dynamic thresholds in real time. It adapts to operating disturbances such as transformer load fluctuations and oil temperature changes, avoiding high-frequency false alarms caused by traditional fixed thresholds, such as the normal acceleration of oil flow under high load in summer being mistaken for an anomaly. In actual scenario testing, a converter station's measurements showed that the dynamic threshold reduced the false alarm rate from 12.3% to 1.7%.

[0059] Specifically, the dynamic threshold update and adjustment process in this embodiment is as follows:

[0060] in, For learning rate, This is the load impact factor, with a default value of 0.1. The load power is collected from the SCADA system; This is a dynamic threshold, also known as the threshold baseline. It inherits the threshold from the previous time step to ensure continuity, and it is used for recent abnormal score sets. Stable, then Fine-tuning; if If the fluctuations are violent, then Rapid response; in addition, ,in, , These are the mean and standard deviation of the outlier scores, respectively. The term is the VAE gradient feedback term, based on the VAE reconstruction error. E With KL divergence loss Backpropagation updates use a chain rule to pass feedback to the VAE encoder, optimizing the latent variable generation strategy. This term represents the load change term, indicating the lag in oil flow velocity response to sudden load changes. This term is used to provide forward-looking compensation; for example, if the load rate increases by 10%, then... Temporarily increase by 8%-12%.

[0061] like This will trigger an alarm and locate the abnormal node:

[0062] in, Given the oil flow velocity gradient, the anomaly localization process involves finding the location corresponding to the maximum oil flow velocity gradient.

[0063] The anomaly detection results obtained from the anomaly detection analysis in this embodiment include real-time dynamic thresholds. Abnormal alarm signals, abnormal location information, and updated potential variables z The real-time dynamic threshold is an anomaly detection boundary value that can adaptively adjust over time, integrating equipment operating conditions, load changes, and model learning results, making it more consistent with the characteristics of real-world scenarios. The anomaly alarm signal can be a binary judgment result; if the anomaly score exceeds the dynamic threshold, an anomaly alarm signal is generated and triggered. The anomaly location information in this embodiment can clearly identify the specific location of the anomaly, with an accuracy reaching [percentage missing]. The updated latent variables are 64-dimensional low-dimensional codes, which can be used for visualization of device health status or analysis of degradation trends. Details will not be elaborated here.

[0064] The anomaly detection process in this embodiment achieves a multi-physical quantity coupled response. The load change term incorporates the load change rate into the threshold adjustment process, fully considering the nonlinear relationship between oil flow rate and load. Furthermore, this detection process can be used for long-term stability maintenance. Specifically, it continuously optimizes the threshold generation mechanism through gradient feedback of the KL divergence loss of the VAE, addressing parameter drift caused by equipment aging and ensuring the algorithm's long-term effectiveness.

[0065] The transformer oil flow anomaly detection method provided in this application constructs a temperature field using multimodal data, performs feature fusion based on a spatiotemporal graph network, and conducts anomaly analysis based on a preset VAE. This enables adaptive dynamic anomaly detection without relying on manual intervention. Furthermore, this process jointly analyzes the spatiotemporal coupling characteristics of the oil flow velocity field, temperature field, and vibration spectrum, improving anomaly detection sensitivity and efficiency. It simultaneously considers the coupling characteristic analysis of multiple physics fields and spatiotemporal correlation analysis, thereby ensuring the accuracy and reliability of the detection results. Therefore, this application embodiment solves the technical problem of existing technologies relying on fixed thresholds and manual detection, which struggles to capture the multiphysics coupling characteristics and spatiotemporal correlation characteristics of oil flow anomalies, resulting in low efficiency and accuracy in transformer oil flow anomaly detection in practical scenarios, failing to meet application requirements.

[0066] For easier understanding, please refer to Figure 2 This application provides an embodiment of a transformer oil flow anomaly detection device, comprising: The reconstruction calculation unit 201 is used to collect multi-modal data during the operation of the transformer and determine the velocity gradient matrix, pressure fluctuation matrix and temperature field by reconstructing the fluid field. The multi-modal data includes oil flow velocity, oil temperature distribution and vibration spectrum. The feature fusion unit 202 is used to fuse the velocity gradient matrix, pressure fluctuation matrix, temperature field and vibration spectrum through a preset spatiotemporal correlation model to obtain a fused feature vector. Anomaly analysis unit 203 is used to perform anomaly analysis processing on the fused feature vector based on reconstruction comparison using a preset VAE to determine anomaly scores; The anomaly detection unit 204 is used to perform anomaly detection analysis based on dynamic threshold and feedback optimization according to the anomaly score and load power to obtain anomaly detection results, which include anomaly location and alarm signals.

[0067] Furthermore, the reconfigured computing unit 201 is specifically used for: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, oil temperature distribution, and vibration spectrum were standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and oil temperature distribution. A simplified difference calculation based on the central difference method is performed on the fluid field to obtain the velocity gradient matrix, pressure fluctuation matrix, and temperature field.

[0068] Furthermore, the feature fusion unit 202 is specifically used for: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a pre-defined spatiotemporal correlation model is constructed through graph convolution computation and LSTM temporal computation. The velocity gradient matrix, pressure fluctuation matrix, temperature field, and vibration spectrum are input into a preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

[0069] Furthermore, the anomaly analysis unit 203 is specifically used for: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on the preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on encoded latent variables, reconstruction error, and preset KL divergence to determine anomaly scores, KL divergence loss, and updated latent variables.

[0070] This application also provides a transformer oil flow anomaly detection device, the device including a processor and a memory; The memory is used to store program code and transfer the program code to the processor; The processor is used to execute the transformer oil flow anomaly detection method in the above method embodiment according to the instructions in the program code.

[0071] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0072] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0074] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of this application through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0075] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting abnormal transformer oil flow, characterized in that, include: Multimodal data during transformer operation are collected, and the velocity gradient matrix, pressure fluctuation matrix, and temperature field are determined by reconstructing the fluid field. The multimodal data includes oil flow velocity, oil temperature distribution, and vibration spectrum. The velocity gradient matrix, pressure fluctuation matrix, temperature field, and vibration spectrum are fused using a preset spatiotemporal correlation model to obtain a fused feature vector. An anomaly analysis based on reconstruction comparison is performed on the fused feature vector using a preset VAE to determine the anomaly score; Anomaly detection analysis based on dynamic threshold and feedback optimization is performed according to the anomaly score and load power to obtain anomaly detection results, which include anomaly location and alarm signals.

2. The transformer oil flow anomaly detection method according to claim 1, characterized in that, The process involves collecting multimodal data during transformer operation and determining the velocity gradient matrix, pressure fluctuation matrix, and temperature field by reconstructing the fluid field, including: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, the oil temperature distribution, and the vibration spectrum are standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and the oil temperature distribution. The fluid field is subjected to simplified difference calculation based on the central difference method to obtain the velocity gradient matrix, pressure fluctuation matrix and temperature field.

3. The transformer oil flow anomaly detection method according to claim 1, characterized in that, The process involves fusing features from the flow velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum using a preset spatiotemporal correlation model to obtain a fused feature vector, including: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a preset spatiotemporal correlation model is constructed through graph convolution calculation and LSTM temporal calculation. The velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum are input into the preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

4. The transformer oil flow anomaly detection method according to claim 1, characterized in that, The step of using a preset VAE to perform anomaly analysis on the fused feature vector based on reconstruction comparison to determine anomaly scores includes: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on a preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on the encoded latent variables, the reconstruction error, and the preset KL divergence to determine the anomaly score, KL divergence loss, and update the latent variables.

5. The transformer oil flow anomaly detection method according to claim 4, characterized in that, The anomaly detection analysis based on the anomaly score and load power, using dynamic thresholding and feedback optimization, yields the anomaly detection results, including: A threshold update formula is constructed based on the KL divergence loss, the updated latent variables, and the load power, and feedback optimization calculation is performed to obtain the dynamic threshold. If the anomaly score exceeds the dynamic threshold, an alarm signal is generated, and the anomaly is located based on the flow velocity gradient to obtain the anomaly detection result.

6. A transformer oil flow anomaly detection device, characterized in that, include: The reconstruction calculation unit is used to collect multimodal data during transformer operation and determine the velocity gradient matrix, pressure fluctuation matrix and temperature field by reconstructing the fluid field. The multimodal data includes oil flow velocity, oil temperature distribution and vibration spectrum. The feature fusion unit is used to fuse the flow velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum through a preset spatiotemporal correlation model to obtain a fused feature vector; An anomaly analysis unit is used to perform anomaly analysis processing on the fused feature vector based on reconstruction comparison using a preset VAE to determine anomaly scores; An anomaly detection unit is used to perform anomaly detection analysis based on dynamic threshold and feedback optimization according to the anomaly score and load power, and obtain anomaly detection results, which include anomaly location and alarm signals.

7. The transformer oil flow anomaly detection device according to claim 6, characterized in that, The reconstruction calculation unit is specifically used for: Multimodal data were obtained by collecting oil flow velocity, oil temperature distribution and vibration spectrum during transformer operation using electromagnetic flow, distributed fiber optic temperature sensors and MEMS vibration sensors, respectively. The oil flow velocity, the oil temperature distribution, and the vibration spectrum are standardized respectively. Based on the coefficient of thermal expansion, the fluid field is reconstructed according to the standardized oil flow velocity and the oil temperature distribution. The fluid field is subjected to simplified difference calculation based on the central difference method to obtain the velocity gradient matrix, pressure fluctuation matrix and temperature field.

8. The transformer oil flow anomaly detection device according to claim 6, characterized in that, The feature fusion unit is specifically used for: The transformer oil passage topology is modeled using graph convolution coding operations to obtain an oil passage topology graph model. Based on the oil duct topology graph model, a preset spatiotemporal correlation model is constructed through graph convolution calculation and LSTM temporal calculation. The velocity gradient matrix, the pressure fluctuation matrix, the temperature field, and the vibration spectrum are input into the preset spatiotemporal correlation model for feature fusion to obtain a fused feature vector.

9. The transformer oil flow anomaly detection device according to claim 6, characterized in that, The anomaly analysis unit is specifically used for: The fused feature vector is subjected to feature encoding based on latent variables to obtain encoded latent variables; Based on a preset VAE, the reconstruction error is calculated according to the fused feature vector and the reconstructed feature vector; Anomaly analysis is performed based on the encoded latent variables, the reconstruction error, and the preset KL divergence to determine the anomaly score, KL divergence loss, and update the latent variables.

10. A transformer oil flow anomaly detection device, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the transformer oil flow anomaly detection method according to any one of claims 1-5 according to the instructions in the program code.

Images