Transformer fault diagnosis method, device and system based on multi-modal information fusion

Abstract

The invention relates to a transformer fault diagnosis method, device and system based on multi-modal information fusion. The method comprises the following steps: collecting multi-modal information; selecting multi-modal information, and dividing and preprocessing faults; first-level decision training: inputting the full-mode external fault data into a training gating network, enabling the gating network to serve as a first-level decision, and recognizing and outputting one-hot codes through input data features; second-level decision training: establishing fusion strategies under different input conditions and encoding the fusion strategies, selecting corresponding training sets and test sets for training each fusion model, and performing fault diagnosis as a second-level decision; a primary diagnosis model; fusing the diagnosis model; testing the whole network to obtain a diagnosis result; trend prediction and hierarchical adaptive early warning are carried out; online incremental learning is carried out on cross-site model migration; the method has the advantages that multi-modal information is considered, a dynamic fusion strategy is introduced, and flexibility and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of transformer fault diagnosis, and particularly relates to a transformer fault diagnosis method, device and system based on multi-modal information fusion. BACKGROUND

[0002] The transformer plays an extremely key role in the power system, and it is particularly important to ensure its safe and stable operation. However, the transformer fault types are various. In order to timely and accurately find the potential defects and provide rich and effective information for subsequent operation and maintenance, it is necessary to improve the accuracy of the transformer fault diagnosis method. The traditional single-modal transformer fault diagnosis method mainly analyzes a certain specific monitoring information, such as the typical dissolved gas analysis (DGA) in oil, partial discharge (PD) monitoring, infrared (IR) identification technology and the like. Although these single-modal methods are relatively effective in specific scenarios, the diagnosis methods among different modalities lack coupling in the diagnosis process, the fault determination boundary is fuzzy, the diagnosis has certain limitations, and the accuracy of the diagnosis result is difficult to meet the actual working condition requirements. The current diagnosis methods are static, and the diagnosis mode is fixed, so the diagnosis flexibility is insufficient. For example, when a certain modality is missing, the diagnosis method is no longer applicable. When the data is abnormal, the diagnosis accuracy decreases, and it is difficult to meet the high monitoring requirements of the field. Therefore, it is necessary to provide a transformer fault diagnosis method, device and system based on multi-modal information fusion, which considers multi-modal information, introduces a dynamic fusion strategy, and improves the flexibility and accuracy. SUMMARY

[0003] The purpose of the present application is to overcome the deficiencies of the prior art and provide a transformer fault diagnosis method, device and system based on multi-modal information fusion, which considers multi-modal information, introduces a dynamic fusion strategy, and improves the flexibility and accuracy.

[0004] The purpose of the present application is achieved as follows: In a first aspect, a transformer fault diagnosis method based on multi-modal information fusion is provided, which comprises the following steps:

[0005] Multi-modal information collection: various types of sensors are deployed to collect transformer operation multi-source data information containing electrical quantities, thermal quantities and acoustic signals;

[0006] Multi-modal information selection, fault division and preprocessing: three modal data of dissolved gas data in oil, partial discharge data and infrared thermal image are selected for fusion, and the transformer fault is divided into internal fault and external fault, and the data is preprocessed;

[0007] First-level decision training: the full-modal external fault data is input into the training gate network, so that the gate network G(x) serves as the first-level decision, and the one-hot encoding is output through the input data feature recognition;

[0008] Second-level decision training: Establish fusion strategies under different input conditions and encode, select corresponding training sets and test sets for training of each fusion model, and make second-level decisions for fault diagnosis;

[0009] Primary diagnosis model: The preprocessed data is input into the primary diagnosis module to make primary diagnosis on internal faults, and the diagnosis results are converted into BPA in the recognition framework to construct multiple evidence bodies required for fusion;

[0010] Fusion diagnosis model: If the conclusions of multiple primary diagnoses are consistent, the decision conclusion is directly obtained, otherwise, evidence body synthesis is performed; if there is no hidden conflict among the evidence bodies, a discount method based on evidence classification is used, otherwise, a principal element-based evidence discount method is used for synthesis, and a decision conclusion is obtained;

[0011] Overall network test, get the diagnosis result;

[0012] Trend prediction and hierarchical adaptive early warning: For key state parameters in transformer operation, a trend prediction model combining sliding window regression and exponential smoothing is constructed, and an adaptive threshold adjustment mechanism is introduced to realize hierarchical alarm output;

[0013] Online incremental learning and cross-site model migration: Online incremental learning mechanism and transfer learning technology are introduced to maintain prediction and diagnosis accuracy with changes in operating conditions, seasons and equipment aging degree, and to migrate model parameters optimized in a substation to a new site.

[0014] The data preprocessing in the multi-modal information selection, fault division and preprocessing is specifically: for the normalization of H2 in oil dissolved gas, taking its volume fraction as the proportion of the total content of five gases, for other four hydrocarbon gases, taking its proportion of total hydrocarbon content; for infrared thermal image data, using an enhanced super-resolution generative adversarial network method for super-resolution reconstruction; for partial discharge data, using an adaptive weighted frame-based fast sparse representation denoising method to improve the quality of partial discharge data.

[0015] The first-level decision training is specifically: assuming that the modalities to be fused are m, the set formed by the m modalities is represented as the full modality X, X=(x1,x2,...,xm), considering all input conditions, designing a unique fusion strategy F for each network under each input condition, represented as F m (A), F i (A),(i=1,2,...,2 i m-1), in the formula, A is a subset of the full modality X; F m-1 (A) is a fusion strategy of the subset A; F iLet be the i-th fusion strategy; embed a trained gating network function G(x) at an appropriate position in the model, which can automatically select the appropriate mode to participate in the fusion based on the characteristics of the input data of each modality; that is, after inputting the full modality X into the gating network G(x), a discrete b-dimensional sparse vector is generated, denoted as g, where g = one-hot(argmax) i G(x) i The fusion strategy F determines whether to activate a branch by using the one-hot encoding of the output. i (A) will give you the final result.

[0016] The second level of decision training specifically involves using DS evidence theory to perform decision fusion diagnosis on partial modality data, implemented through the following function: basic probability assignment function. In the formula, Θ represents the recognition frame; A represents a subset of the recognition frames; and the trust function... In the formula, B is all subsets of A; the likelihood function DS Evidence Theory Synthesis Calculation Rules In the formula, K is the conflict factor.

[0017] The primary diagnostic model includes the following steps:

[0018] Multi-evidence body construction method: Four intelligent methods, namely BP, RBF neural network, support vector machine and S_Kohonen network, are used to perform preliminary diagnosis of transformer faults and construct a multi-evidence body;

[0019] Evidence discounting synthesis method: The evidence discounting method calculates the discount factor based on the degree of conflict between evidence, and uses the obtained discount factor to modify the evidence body, thus overcoming the Dempster rule paradox.

[0020] The evidence discounting method is as follows: assuming the BPA of evidence A is m and its discount factor is α, then the discounted evidence should be... In the formula, Θ is the overall identification framework; m(A) is called the basic probability number of A, representing the precise trust in A.

[0021] The second aspect: a transformer fault diagnosis device based on multimodal information fusion, including a multi-parameter fusion detection device, wherein the transformer fault diagnosis device based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described above;

[0022] The multi-parameter fusion detection device includes an integrated detection terminal, an AD / DA module, a power coupling circuit, a detection module, a control chip, a data management module, a mechanical locking interface, and a mis-insertion protection device; the integrated detection terminal includes a voltage / current input interface and a pluggable modular unit.

[0023] The integrated detection terminal, AD / DA module, and power coupling circuit are combined to collect multi-source data information on transformer operation, including electrical quantities, thermal quantities, and acoustic signals.

[0024] The detection module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data.

[0025] The control chip is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition;

[0026] The control chip is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use it as a second-level decision for fault diagnosis.

[0027] The control chip is also used to input preprocessed data into the primary diagnostic module to perform primary diagnosis of internal faults. The diagnostic results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion.

[0028] The control chip is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there is no hidden conflict between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn.

[0029] The data management module is used for overall network testing and to obtain diagnostic results.

[0030] The third aspect is a transformer fault diagnosis system based on multimodal information fusion, which includes a multimodal signal acquisition and synchronization clock system, a preprocessing module, a multi-source fusion calculation module, a multimodal feature fusion diagnosis network module, a trend prediction and hierarchical adaptive early warning module, and an online incremental learning and cross-site model transfer module. The transformer fault diagnosis system based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described above.

[0031] The multimodal signal acquisition and synchronization clock system is used to acquire multi-source data information on transformer operation, including electrical, thermal, and acoustic signals, using various deployed sensors.

[0032] The preprocessing module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data.

[0033] The data preprocessing specifically involves: classifying real sample data, adding labels to samples according to the fault types confirmed in the historical operation and maintenance data reports, ensuring the time synchronization of multimodal data in the same group according to the specified time window and the actual data missing situation, and dividing the corresponding training set and test set according to different training requirements.

[0034] The multi-source fusion computing module is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition;

[0035] The multi-source fusion computing module is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use them as a second-level decision for fault diagnosis.

[0036] The multi-source fusion computing module is also used to input the preprocessed data into the primary diagnosis module to perform primary diagnosis of internal faults. The diagnosis results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion.

[0037] The multi-source fusion computing module is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there are no hidden conflicts between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn.

[0038] The multimodal feature fusion diagnostic network module is used for overall network testing to obtain diagnostic results;

[0039] The trend prediction and hierarchical adaptive early warning module is used to construct a trend prediction model that combines sliding window regression and exponential smoothing for key state parameters in transformer operation, and introduces an adaptive threshold adjustment mechanism to achieve hierarchical alarm output.

[0040] The online incremental learning and cross-site model transfer module is used to introduce an online incremental learning mechanism and transfer learning technology to maintain prediction and diagnostic accuracy as factors such as operating conditions, seasons, and equipment aging change, and to transfer model parameters that have been optimized in one substation to a new site.

[0041] The beneficial effects of this invention are as follows: This invention provides a transformer fault diagnosis method, device, and system based on multimodal information fusion. In use, this invention introduces a dynamic fusion strategy. First, a modal selection network is constructed, which can autonomously filter input monitoring information and dynamically select the diagnostic mode to adapt to different on-site monitoring conditions. Then, corresponding diagnostic models are established for different input modes. Under non-single-modal conditions, appropriate fusion methods are used for diagnosis, achieving full utilization of monitoring information from each modality. This invention effectively improves the flexibility and accuracy of transformer fault diagnosis results, adapts to different on-site monitoring conditions, has a higher identification accuracy, and lower false alarm and false negative rates. This invention has the advantages of considering multimodal information, introducing a dynamic fusion strategy, and improving flexibility and accuracy. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention.

[0043] Figure 2 This is an example diagram of the dynamic multimodal fusion of the present invention.

[0044] Figure 3 This is a block diagram of the information fusion fault diagnosis model of the present invention.

[0045] Figure 4 This is a schematic diagram of the structure of the device of the present invention.

[0046] Figure 5 This is a schematic diagram of the system composition of the present invention.

[0047] Figure 6 This is a schematic diagram of the AD / DA module resource reuse and power coupling circuit structure of the system of the present invention. Detailed Implementation

[0048] The present invention will now be further described with reference to the accompanying drawings.

[0049] Example 1

[0050] like Figures 1-3 As shown, a transformer fault diagnosis method based on multimodal information fusion includes the following steps:

[0051] Multimodal information collection: Utilize various deployed sensors to collect multi-source data information on transformer operation, including electrical quantities (such as current waveform and harmonic distribution), thermal quantities (such as hot spot temperature distribution and temperature rise rate), and acoustic signals (such as partial discharge pulse characteristics and spectral energy distribution), including parameters such as DC resistance, short-circuit impedance, insulation resistance, and dielectric loss.

[0052] Multimodal information selection, fault classification and preprocessing: Three types of data, namely dissolved gas data in oil, partial discharge data and infrared thermal image, are selected and fused. Transformer faults are classified into internal faults and external faults. The data are then preprocessed.

[0053] In this invention, ① transformer faults can be divided into internal and external faults according to their location. Internal faults mainly occur in components such as windings and cores, while external faults mainly occur in bushings. Therefore, three modal data can be fused: dissolved gas data in oil, partial discharge data, and infrared thermal images. In this invention, transformer fault types are reclassified according to fault type and location, as shown in Table 1 below.

[0054] Table 1. Classification of Transformer Fault Types Based on Multimodal Information

[0055]

[0056] ② Combining the types of internal transformer faults and the collected samples, electrical test fault characteristic data were selected. Referring to the results of existing fault mode classifications, the identification framework Θ was finally determined as follows: {F1 winding fault; F2 tap changer and lead fault; F3 core fault; F4 magnetic shielding and structural component fault; F5 enclosure and support bar fault; F6 floating potential body fault; F7 insulating oil fault; F8 normal}. Based on the determined identification framework and the collected fault samples, the selected electrical test data included winding DC resistance imbalance coefficient, winding dielectric loss, transformation ratio, core grounding current, and micro-water. Oil chromatographic data of H2, CH4, C2H2, C2H4, and C2H6 were also used.

[0057] ③ The dissolved gas analysis (DGA) data acquisition frequency is 1 data point / hour. When DGA data is available, a time window of 1 hour is set with the DGA data sampling time as the center, that is, 0.5 hours before and after the DGA data acquisition time. Other modal data within this time period are considered to belong to the same time period as DGA data. When DGA data is missing, infrared (IR) / partial discharge (PD) is used for single-mode fault diagnosis.

[0058] Secondly, since the external and internal components are monitored independently and their data are not shared, the model will perform diagnosis separately. To enable the model to distinguish between internal and external data, the internal data will be labeled "0" and the external data will be labeled "1".

[0059] Finally, to improve the model training effect, the raw data needs to be preprocessed. In the analysis of dissolved gases in oil, this invention selects the volume fractions of five characteristic gases—H2, CH4, C2H2, C2H4, and C2H6—as the model input to optimize the training effect. Specifically, for the normalization of H2, its volume fraction (10... -6The percentage of the total content of the five gases is taken as the percentage of the total hydrocarbon content of the other four hydrocarbon gases.

[0060] The enhanced super-resolution generative adversarial network (ESRGAN) method is used to perform super-resolution reconstruction of infrared images to improve image clarity and facilitate subsequent identification.

[0061] An adaptive weighted frame-based fast sparse representation denoising method is employed to improve the quality of partial discharge data. Each partial discharge sample is derived from a 5-second raw signal (approximately 250 cycles). After denoising, the pulse and phase are extracted, and the phase from 0-360° and 0-V are combined. max The discharge amplitude (maximum discharge amplitude within 5 seconds) is discretized into 90 interval spectra. Using the number of discharges as matrix elements, a 90×90 matrix is ​​formed as the PRPD spectrum. Alternatively, the sigmoid function can be used as the membership function for fuzzification. Where x represents the actual value of the data to be diagnosed; x a y represents the attention value; y represents the preprocessed output value.

[0062] First-level decision training: The gating network is trained by inputting full-modal external fault data, so that the gating network G(x) serves as the first-level decision and outputs one-hot encoding through feature recognition of the input data;

[0063] Second-level decision training: Establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use them as second-level decisions for fault diagnosis;

[0064] In this invention, ① it is assumed that there are a total of m modes to be fused, and the set formed by the m modes is represented as the full mode X, X = (x1, x2, ..., x...). m Considering all input scenarios, a unique fusion strategy F is designed for the network under each input condition, denoted as Fi. i (A), F i (A), (i = 1, 2, ..., 2) m-1 In the formula, A is a subset of the full-modal X; F i Let i be the fusion strategy.

[0065] A trained gating network function G(x) is embedded at an appropriate location in the model. This function can automatically select the appropriate modalities to participate in the fusion based on the characteristics of the input data from each modality. That is, after inputting the full modality X into the gating network G(x), a discrete b-dimensional sparse vector (one-hot encoded) is generated, denoted as g, where g = one-hot(argmax) i G(x) iThe fusion strategy F determines whether to activate a branch by using the one-hot encoding of the output. i (A) yields the final result; taking m=2 as an example, the specific decision-making process is as follows: Figure 2 As shown in the figure, w is the weighting coefficient; y is the output result.

[0066] Each fusion strategy is a pre-defined and trained model. Since all possible modality combinations have been considered, each fusion selects only the most suitable branch strategy, maximizing computational simplicity. Decision fusion diagnosis of partial modality data is performed using DS evidence theory, implemented through the following function:

[0067] (1) Basic probability assignment function, In the formula, Θ is the recognition frame; A is a subset of the recognition frame.

[0068] (2) Trust function, In the formula, B is all subsets of A.

[0069] (3) Likelihood function

[0070] (4) DS evidence theory synthesis calculation rules In the formula, K is the conflict factor.

[0071] ② Dynamic Multimodal Fusion Fault Diagnosis Method: Considering different data monitoring conditions on site, this invention proposes three diagnostic strategies for different input conditions:

[0072] Strategy 1: Retain the ability of a single modality model to solve specific problems and use it when only a single piece of available data exists. That is, each of the three modalities uses its appropriate model for diagnosis.

[0073] Strategy 2: When there is anomaly or missing data in a certain modality, use oil chromatography data as the main data source and adopt a decision fusion method with partial discharge monitoring data or infrared image data to improve diagnostic accuracy; if oil chromatography data is missing, partial discharge monitoring data and infrared image data are diagnosed using single-modality models respectively.

[0074] Strategy 3: When there is no missing data and no anomalies, a hybrid fusion approach is adopted to make full use of the complementarity between various modes. The diagnostic output after fusing the features of oil chromatography and partial discharge data is fused with the recognition output of the infrared image model using evidence theory to obtain the final diagnostic result.

[0075] Primary diagnostic model: Preprocessed data is input into the primary diagnostic module to perform a primary diagnosis of internal faults. The diagnostic results are converted into BPA (based on probability allocation) under the identification framework to construct multiple evidence bodies required for fusion.

[0076] In this invention, a method for constructing and synthesizing multiple pieces of evidence is provided.

[0077] ①Multiple Evidence Body Construction Method: This invention uses four intelligent methods—BP, RBF neural network, support vector machine, and S-Kohonen network—to perform preliminary diagnosis of transformer faults and construct multiple evidence bodies;

[0078] ② Evidence Discounting and Composition Method: When using Dempster's rule for evidence reasoning, the rule may fail or conclusions may arise that contradict common sense; this is collectively referred to as the Dempster paradox. Paradoxes directly affect the credibility of the evidence composition result. The evidence discounting method calculates a discount factor based on the degree of conflict between pieces of evidence, and uses this discount factor to modify the evidence body, thus overcoming the Dempster paradox. Assuming evidence A has a BPA of m and a discount factor of α, the discounted evidence should be: In the formula, Θ is the overall identification framework; m(A) is called the basic probability number of A, representing the exact trust in A; the determination of the discount factor α is the core of the evidence discounting method.

[0079] The classification-based evidence discounting method uses a modified Jousselme distance and conjunction conflict to identify credible and non-conflicting evidence, assigning a discount factor of 1 to these two types of evidence. This greatly preserves the support for credible and non-conflicting evidence. For conflicting evidence, a discount factor of appropriate magnitude is assigned based on the degree of conflict. This retains some information about the evidence while reducing the negative impact of conflicting evidence on the fusion result, which is quite reasonable. This evidence discounting method can effectively classify evidence, and the synthesized evidence can focus on the correct hypothesis to a great extent, exhibiting good convergence. However, this method classifies and discounts evidence based on the average modified distance and the average magnitude of conjunction conflict. It can fuse general conflicting evidence well, but if there are hidden conflicts between the evidence, it will not be able to achieve fusion. The purpose of the evidence principal component-based method is to propose a method for judging hidden conflicts in evidence sets. The principal component-based discounting method itself can determine whether there are hidden conflicts between evidence, so this method is still effective when hidden conflicts exist.

[0080] To effectively overcome various conflict paradoxes in the transformer fusion decision-making process, this invention first determines whether there are hidden conflicts between the obtained evidence bodies. If there are no hidden conflicts, a classification-based evidence discounting method is preferred to ensure the credibility of the decision. The principal component-based evidence discounting algorithm is used as a supplementary fusion method when there are hidden conflicts between the evidence bodies.

[0081] As a specific achievable embodiment: the present invention collected 321 sets of samples containing normal data, selected 180 sets as training samples, and the rest as test samples; the test samples were substituted into the four sub-modules of the primary diagnosis to obtain the primary diagnosis results; let the j-th output value of the i-th intelligent diagnosis algorithm be 0. i (j) Calculate the accuracy δ of the four intelligent diagnostic algorithms respectively. i (i = 1, 2, 3, 4), then the basic probability allocation of the intelligent diagnostic algorithm for judging j is as follows: The BPA allocation for each piece of evidence is shown in Table 2. In the table, m1, m2, m3, and m4 correspond to BP neural network, FBF neural network, support vector machine, and S-Kohonen clustering, respectively.

[0082] Table 2. Basic probability allocation results for each piece of evidence.

[0083] body

[0084]

[0085] Fusion diagnostic model: If the conclusions of multiple primary diagnoses are consistent, a decision conclusion is directly drawn; otherwise, evidence bodies are synthesized. If there are no hidden conflicts between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn.

[0086] In this invention: ① If the conclusions given by the four evidence bodies in the primary diagnosis are consistent, a decision conclusion is given directly without further fusion of the evidence bodies; ② If condition ① is not met, the evidence bodies are first classified. After classification, it is first determined whether there is a hidden conflict between credible evidence and non-conflicting evidence. If there is no hidden conflict, that is, the intersection of the principal components of the evidence set composed of credible evidence and non-conflicting evidence is not empty, then the evidence body fusion decision is made using the evidence classification method to give a conclusion; ③ In the evidence set composed of credible evidence and non-conflicting evidence, if the intersection of the principal components of each evidence body is empty, that is, there are no identical principal components, then the evidence body fusion decision is made using the principal component discount method to give a conclusion, and the fused confidence interval is obtained. and uncertainty m i After (Θ), the diagnostic conclusion can be determined by the following rules:

[0087] 1) The fault type identified should have the highest confidence function value.

[0088] 2) The confidence function value of the identified fault type must be greater than the uncertain confidence function value.

[0089] 3) The difference between the confidence function value of the identified fault type and other types must be greater than a certain threshold. If this value is too large, a diagnostic conclusion cannot be drawn; if it is too small, the fault type cannot be effectively distinguished. Therefore, based on the actual diagnostic situation and considering the requirement to effectively distinguish faults, this invention has determined it to be 0.4 after extensive debugging.

[0090] 4) The uncertainty confidence function value must be less than a certain threshold; otherwise, the corresponding evidence will be too unreliable to provide a diagnostic conclusion. Therefore, this threshold must be greater than the value of equation m. i (Θ)=1-δ i The uncertainty confidence function values ​​for each piece of evidence shown were determined to be 0.26 after repeated adjustments based on the selection of reliability coefficients for each diagnostic network.

[0091] Overall network testing yielded diagnostic results;

[0092] Trend prediction and hierarchical adaptive early warning: For key state parameters in transformer operation, a trend prediction model combining sliding window regression and exponential smoothing is constructed, and an adaptive threshold adjustment mechanism is introduced to realize hierarchical alarm output.

[0093] Online incremental learning for cross-site model transfer: Introducing online incremental learning mechanisms and transfer learning techniques enables the maintenance of prediction and diagnostic accuracy as factors such as operating conditions, seasons, and equipment aging change, and allows the transfer of model parameters that have been optimized in one substation to a new site.

[0094] In summary, this invention can dynamically filter data and select appropriate fusion strategies for fault diagnosis based on input monitoring information, adapting to different on-site acquisition conditions and offering greater flexibility compared to traditional diagnostic models. This invention considers more comprehensive information, and through the complementarity between modes, it further improves the accuracy of transformer fault diagnosis, especially for thermal faults. This invention combines neural networks and S-Kohonen clustering methods to form a novel fault diagnosis method based on information fusion within multiple evidence bodies. It automatically determines whether there are hidden conflicts between evidence bodies. If no hidden conflicts exist, it prioritizes the evidence classification discount algorithm; otherwise, it uses the evidence principal component discount algorithm to modify the evidence source and synthesizes the fault diagnosis conclusion using Dempster rules. It reasonably coordinates the information conflicts between evidence bodies and effectively integrates the judgment results of various intelligent algorithms. Due to the increased information richness and the introduction of dynamic fusion strategies, this invention can automatically remove abnormal data before diagnosis and effectively reduce the false positive and false negative rates of faults by dynamically selecting appropriate diagnostic modes. It provides more detailed and accurate information for transformer operation and maintenance, improves on-site operation and maintenance efficiency, and ensures the safe and reliable operation of the power system.

[0095] This invention relates to a transformer fault diagnosis method, device, and system based on multimodal information fusion. In use, this invention introduces a dynamic fusion strategy. First, a modal selection network is constructed as the first-level decision layer, which can autonomously filter input monitoring information and dynamically select the diagnostic mode to adapt to different on-site monitoring conditions. This invention uses neural networks, SVM, and S-Kohone clustering algorithms to form multiple evidence bodies when there are discrepancies in the initial diagnostic results. It judges whether there are hidden conflicts between the evidence bodies. If there are no hidden conflicts, the evidence classification discount algorithm is used first; otherwise, the evidence source is modified using the evidence principal component evidence discount algorithm. The fault diagnosis conclusion is then synthesized using Dempster rules. Then, diagnostic models corresponding to different input modes are established and trained for different input conditions. Under non-single-modal conditions, corresponding fusion methods are used for diagnosis, realizing the full utilization of monitoring information from each modality. This invention can effectively improve the flexibility and accuracy of transformer fault diagnosis results, adapt to different on-site monitoring conditions, have higher recognition accuracy, and lower false alarm and false negative rates. This invention has the advantages of considering multimodal information, introducing a dynamic fusion strategy, and improving flexibility and accuracy.

[0096] Example 2

[0097] like Figure 4 As shown, the transformer fault diagnosis device based on multimodal information fusion includes a multi-parameter fusion detection device. The transformer fault diagnosis device based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described above.

[0098] The multi-parameter fusion detection device includes an integrated detection terminal, an AD / DA module, a power coupling circuit, a detection module, a control chip, a data management module, a mechanical locking interface, and a mis-insertion protection device; the integrated detection terminal includes a voltage / current input interface and a pluggable modular unit.

[0099] In this embodiment, the multi-parameter fusion detection device adopts a one-time wiring and multi-parameter sequential detection architecture, equipped with a highly reliable mechanical locking interface and a mis-insertion protection device. Combined with a channel self-identification chip and a dynamic sequence optimization algorithm, it significantly reduces the on-site wiring and switching operation time. To ensure the reliability of the equipment under continuous high load operation, finite element simulation is used to optimize thermal management and internal layout, so that the core components can still work stably under complex working conditions.

[0100] The pluggable modular unit encapsulates signal conditioning, power drive, data acquisition, and edge computing units in layers. Inter-module communication is achieved through a high-speed backplane bus (such as PCIe Gen3 or high-speed LVDS), and FPGA programmable logic units are reserved in the interface layer to facilitate field expansion or upgrade of functions.

[0101] The signal conditioning unit employs a combination of intelligent gain control and anti-aliasing filtering. It automatically adjusts the gain coefficient G based on sensor feedback, maintaining the signal dynamic range within its optimal range and ensuring a high signal-to-noise ratio for both strong and weak signals. This avoids clipping or distortion caused by fixed-gain circuits in complex environments. The anti-aliasing filtering section optimizes the cutoff frequency and stopband characteristics, suppressing high-frequency noise and aliasing, providing stable input for subsequent digital processing. The pluggable modular units not only enhance the system's adaptability and maintainability but also allow for flexible addition or removal of functional modules according to task requirements, enabling rapid configuration and efficient detection in various scenarios.

[0102] The integrated detection terminal, AD / DA module, and power coupling circuit are combined to collect multi-source data information on transformer operation, including electrical quantities, thermal quantities, and acoustic signals.

[0103] In this embodiment, by reusing AD / DA module resources and integrating power coupling circuits, unified detection of multiple parameters such as DC resistance, short-circuit impedance, insulation resistance, and dielectric loss is achieved, and functional modules are expanded without increasing size and weight.

[0104] The integrated testing terminal is based on a high-resolution multi-channel synchronous sampling system, combined with hardware protection measures such as optical isolation, differential input, and independent power supply partitioning wiring, and integrates algorithms such as adaptive bandpass filtering, Kalman filtering and dynamic moving average to achieve stable and reliable data acquisition.

[0105] The detection module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data.

[0106] The control chip is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition;

[0107] The control chip is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use it as a second-level decision for fault diagnosis.

[0108] The control chip is also used to input preprocessed data into the primary diagnostic module to perform primary diagnosis of internal faults. The diagnostic results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion.

[0109] The control chip is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there is no hidden conflict between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn.

[0110] The data management module is used for overall network testing and to obtain diagnostic results.

[0111] In this embodiment, the data management module constructs a hierarchical state criterion library and introduces a shallow neural network and a random forest model to achieve multi-indicator combination diagnosis, significantly enhancing the ability to identify latent defects such as mild degradation and potential overheating.

[0112] Example 3

[0113] like Figures 5-6 As shown, the transformer fault diagnosis system based on multimodal information fusion includes a multimodal signal acquisition and synchronization clock system, a preprocessing module, a multi-source fusion calculation module, a multimodal feature fusion diagnosis network module, a trend prediction and hierarchical adaptive early warning module, and an online incremental learning and cross-site model transfer module. The transformer fault diagnosis system based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described above.

[0114] The multimodal signal acquisition and synchronization clock system is used to acquire multi-source data information on transformer operation, including electrical, thermal, and acoustic signals, using various deployed sensors.

[0115] In this embodiment, to ensure the timing consistency of multi-source data such as electrical quantities, thermal quantities, and acoustic signals, a synchronous acquisition system based on the IEEE 1588PTP precision clock is constructed to control the sampling delay within an acceptable range; AD / DA module resource reuse and power coupling circuit integration methods are utilized, such as... Figure 6 As shown, a prototype multifunctional detection module with a universal voltage / current input interface is constructed to realize the real-time acquisition of voltage U(t), current I(t), temperature T(t), and acoustic emission signal S(t). Dual clock domain control technology is introduced to ensure that the Nyquist condition is met at the high-speed sampling frequency, and to support different types of sensors to operate at their respective optimal sampling frequencies. Cross-modal alignment is achieved through timestamp marking, thereby improving the effective bandwidth and anti-interference capability of the fused data.

[0116] The preprocessing module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data.

[0117] The data preprocessing specifically involves: classifying real sample data, adding labels to samples according to the fault types confirmed in the historical operation and maintenance data reports, ensuring the time synchronization of multimodal data in the same group according to the specified time window and the actual data missing situation, and dividing the corresponding training set and test set according to different training requirements.

[0118] The multi-source fusion computing module is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition;

[0119] The multi-source fusion computing module is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use them as a second-level decision for fault diagnosis.

[0120] The multi-source fusion computing module is also used to input the preprocessed data into the primary diagnosis module to perform primary diagnosis of internal faults. The diagnosis results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion.

[0121] The multi-source fusion computing module is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there are no hidden conflicts between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn.

[0122] In this embodiment, a multi-source fusion algorithm is embedded on the FPGA chip to achieve edge computing and data compression, reducing uplink transmission pressure. The core of the algorithm is a weighted fusion model: X f =∑(w i *(x i -μ i ) / σ i ),∑w i =1, where μ i σ i For standardized parameters; w i The accuracy is dynamically adjusted based on historical data; the fusion computing not only integrates multimodal inputs such as electrical quantities, thermal quantities, and acoustic signals, but also enables preliminary discrimination and data filtering at the hardware level.

[0123] In the feature extraction stage, the fused data directly outputs key indicators such as harmonic components, temperature rise rate, and acoustic emission envelope energy, achieving "initial judgment at the hardware end." This method can effectively filter out invalid or redundant data at the data acquisition source, reducing the computational burden of subsequent algorithms and reducing network transmission bandwidth usage. By completing part of the feature extraction and data compression at the FPGA end, the system not only shortens the overall analysis latency but also improves fault tolerance and stability under complex working conditions, providing high-value, low-latency input data for upper-level intelligent diagnostic algorithms and significantly enhancing the overall system's rapid response and real-time decision-making capabilities.

[0124] The multimodal feature fusion diagnostic network module is used for overall network testing to obtain diagnostic results;

[0125] In this embodiment, electrical quantities (such as current waveforms and harmonic distributions), thermal quantities (such as hot spot temperature distributions and temperature rise rates), and acoustic signals (such as partial discharge pulse characteristics and spectral energy distributions) are fused in a multimodal manner to form a comprehensive feature vector, which is then input into a hybrid deep learning network. The network architecture combines the local feature extraction capability of convolutional neural networks with the temporal dependency capture capability of long short-term memory networks, enabling simultaneous identification of short-term abrupt changes and long-term trend anomalies. To improve the interpretability of the diagnosis, the system introduces feature contribution visualization technology (such as SHAP or Grad-CAM), allowing the algorithm to not only provide diagnostic results but also clearly display the judgment criteria. Through a multi-task learning framework, dedicated sub-models are established for different types of defects (partial discharge, winding deformation, insulation aging, etc.), enabling accurate diagnosis of multiple types of faults with a single model. Furthermore, the feature weights can be dynamically adjusted according to the equipment operating environment to improve adaptability and robustness.

[0126] The trend prediction and hierarchical adaptive early warning module is used to construct a trend prediction model that combines sliding window regression and exponential smoothing for key state parameters in transformer operation, and introduces an adaptive threshold adjustment mechanism to achieve hierarchical alarm output.

[0127] In this embodiment, a multi-period trend prediction model combining sliding window regression and exponential smoothing is constructed for key state parameters during transformer operation. in This is the model's predicted value at time t+1; These are the base values ​​or constants of the model, representing the baseline level for prediction; This represents the sliding window regression part. The i-th regression coefficient represents the influence of the data from the i-th time step on the current prediction; y(ti) is the observed value at time ti; ε tThe error term represents the prediction error or noise at time t. This model can predict short-term and medium-term operating trends with the support of multi-source data, and adaptively adjust the model parameters through historical samples to adapt to different working conditions. The prediction results are compared with the real-time collected values. Combined with the anomaly scoring system, the warning level (slight, moderate, severe) is automatically determined, and an intuitive graded alarm output is generated.

[0128] An adaptive threshold adjustment mechanism is introduced into the early warning strategy to dynamically adjust the alarm threshold based on external conditions such as seasonal temperature, equipment load fluctuations, and years of operation. When the equipment operating environment is in a high-risk state such as high temperature, heavy load, or increased vibration, the system automatically tightens the threshold and increases the monitoring frequency. Under stable environmental and load conditions, the threshold is appropriately relaxed to reduce false alarm rate and invalid alarms. Compared with the traditional fixed threshold early warning method, this method can significantly reduce the probability of missed and false alarms under variable operating conditions, while maintaining high early warning sensitivity and stability. Especially in the early stage of latent faults, it can capture small deviations of key parameters in advance through trend prediction, reserve intervention time window for operation and maintenance decisions, and improve the system's foresight and reliability.

[0129] The online incremental learning and cross-site model transfer module is used to introduce an online incremental learning mechanism and transfer learning technology to maintain prediction and diagnostic accuracy as factors such as operating conditions, seasons, and equipment aging change, and to transfer model parameters that have been optimized in one substation to a new site.

[0130] In this embodiment, to ensure that the intelligent analysis algorithm can adapt to changes in the field environment over a long period of time, an online incremental learning mechanism is introduced. This mechanism supports updating parameters only for newly collected data samples without retraining the entire model, thereby significantly shortening the optimization cycle and reducing computing power while ensuring model stability. The online incremental learning process can be carried out continuously during the daily operation of the system, enabling the model to maintain prediction and diagnostic accuracy as factors such as working conditions, seasons, and equipment aging change.

[0131] In terms of cross-site promotion, transfer learning technology is adopted to transfer the model parameters that have been optimized in a certain substation to a new site. After the transfer, the system only needs to use a small amount of local field data for fine-tuning to achieve a high diagnostic accuracy in the new environment. This avoids large-scale training from scratch, significantly reduces the data collection cycle and training computation, and reduces the dependence on large samples and high computing power.

Claims

1. A transformer fault diagnosis method based on multimodal information fusion, characterized in that: The method includes the following steps: Multimodal information collection: Utilizing various deployed sensors to collect multi-source data information on transformer operation, including electrical quantities, thermal quantities, and acoustic signals; Multimodal information selection, fault classification and preprocessing: Three types of data, namely dissolved gas data in oil, partial discharge data and infrared thermal image, are selected and fused. Transformer faults are classified into internal faults and external faults. The data are then preprocessed. First-level decision training: The gating network is trained by inputting full-modal external fault data, so that the gating network G(x) serves as the first-level decision and outputs one-hot encoding through feature recognition of the input data; Second-level decision training: Establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use them as second-level decisions for fault diagnosis; Primary diagnostic model: Preprocessed data is input into the primary diagnostic module to perform a primary diagnosis of internal faults. The diagnostic results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion. Fusion diagnostic model: If the conclusions of multiple primary diagnoses are consistent, a decision conclusion is directly drawn; otherwise, evidence bodies are synthesized. If there are no hidden conflicts between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn. Overall network testing yielded diagnostic results; Trend prediction and hierarchical adaptive early warning: For key state parameters in transformer operation, a trend prediction model combining sliding window regression and exponential smoothing is constructed, and an adaptive threshold adjustment mechanism is introduced to realize hierarchical alarm output. Online incremental learning for cross-site model transfer: Introducing online incremental learning mechanisms and transfer learning techniques enables the maintenance of prediction and diagnostic accuracy as factors such as operating conditions, seasons, and equipment aging change, and allows the transfer of model parameters that have been optimized in one substation to a new site.

2. The transformer fault diagnosis method based on multimodal information fusion as described in claim 1, characterized in that: The data preprocessing in the multimodal information selection, fault division, and preprocessing is specifically as follows: for the standardization of dissolved gas H2 in oil, its volume fraction is taken as the proportion of the total content of the five gases; for the other four hydrocarbon gases, their proportion of the total hydrocarbon content is taken; for infrared thermal image data, super-resolution reconstruction is performed using an enhanced super-resolution generative adversarial network method; for partial discharge data, an adaptive weighted frame-based fast sparse representation denoising method is used to improve the quality of partial discharge data.

3. The transformer fault diagnosis method based on multimodal information fusion as described in claim 2, characterized in that: The first-level decision training is as follows: Assume there are m modes that need to be fused, and the set formed by the m modes is represented as the full mode X, X = (x1, x2, ..., x...). m Considering all input scenarios, a unique fusion strategy F is designed for the network under each input condition, denoted as Fi. i (A), F i (A), (i = 1, 2, ..., 2) m-1 In the formula, A is a subset of the full-modal X; F i Let be the i-th fusion strategy; embed a trained gating network function G(x) at an appropriate position in the model, which can automatically select the appropriate mode to participate in the fusion based on the characteristics of the input data of each modality; that is, after inputting the full modality X into the gating network G(x), a discrete b-dimensional sparse vector is generated, denoted as g, where g = one-hot(argmax) i G(x) i The fusion strategy F determines whether to activate a branch by using the one-hot encoding of the output. i (A) will give you the final result.

4. The transformer fault diagnosis method based on multimodal information fusion as described in claim 3, characterized in that: The second level of decision training specifically involves using DS evidence theory to perform decision fusion diagnosis on partial modality data, implemented through the following function: basic probability assignment function. In the formula, Θ represents the recognition frame; A represents a subset of the recognition frames; Trust function In the formula, B is all subsets of A; the likelihood function DS Evidence Theory Synthesis Calculation Rules In the formula, K is the conflict factor.

5. The transformer fault diagnosis method based on multimodal information fusion as described in claim 2, characterized in that: The primary diagnostic model includes the following steps: Multi-evidence body construction method: Four intelligent methods, namely BP, RBF neural network, support vector machine and S_Kohonen network, are used to perform preliminary diagnosis of transformer faults and construct a multi-evidence body; Evidence discounting synthesis method: The evidence discounting method calculates the discount factor based on the degree of conflict between evidence, and uses the obtained discount factor to modify the evidence body, thus overcoming the Dempster rule paradox.

6. The transformer fault diagnosis method based on multimodal information fusion as described in claim 5, characterized in that: The evidence discounting method is as follows: assuming the BPA of evidence A is m and its discount factor is α, then the discounted evidence should be... In the formula, Θ is the overall identification framework; m(A) is called the basic probability number of A, representing the precise trust in A.

7. A transformer fault diagnosis device based on multimodal information fusion, including a multi-parameter fusion detection device, characterized in that: The transformer fault diagnosis device based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described in any one of claims 1-6; The multi-parameter fusion detection device includes an integrated detection terminal, an AD / DA module, a power coupling circuit, a detection module, a control chip, a data management module, a mechanical locking interface, and a mis-insertion protection device; the integrated detection terminal includes a voltage / current input interface and a pluggable modular unit.

8. The transformer fault diagnosis device based on multimodal information fusion as described in claim 7, characterized in that: The integrated detection terminal, AD / DA module, and power coupling circuit are combined to collect multi-source data information on transformer operation, including electrical quantities, thermal quantities, and acoustic signals. The detection module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data. The control chip is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition; The control chip is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use it as a second-level decision for fault diagnosis. The control chip is also used to input preprocessed data into the primary diagnostic module to perform primary diagnosis of internal faults. The diagnostic results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion. The control chip is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there is no hidden conflict between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn. The data management module is used for overall network testing and to obtain diagnostic results.

9. A transformer fault diagnosis system based on multimodal information fusion, comprising a multimodal signal acquisition and synchronization clock system, a preprocessing module, a multi-source fusion calculation module, a multimodal feature fusion diagnostic network module, a trend prediction and hierarchical adaptive early warning module, and an online incremental learning and cross-site model transfer module, characterized in that: The transformer fault diagnosis system based on multimodal information fusion is used to execute the transformer fault diagnosis method based on multimodal information fusion as described in any one of claims 1-6.

10. The transformer fault diagnosis system based on multimodal information fusion as described in claim 1, characterized in that: The multimodal signal acquisition and synchronization clock system is used to acquire multi-source data information on transformer operation, including electrical, thermal, and acoustic signals, using various deployed sensors. The preprocessing module is used to select three modal data—dissolved gas data in oil, partial discharge data, and infrared thermal image—for fusion, and to classify transformer faults into internal and external faults, and to preprocess the data. The data preprocessing specifically involves: classifying real sample data, adding labels to samples according to the fault types confirmed in the historical operation and maintenance data reports, ensuring the time synchronization of multimodal data in the same group according to the specified time window and the actual data missing situation, and dividing the corresponding training set and test set according to different training requirements. The multi-source fusion computing module is used to input full-modal external fault data into the training gating network, so that the gating network G(x) is used as the first-level decision and outputs one-hot encoding through input data feature recognition; The multi-source fusion computing module is also used to establish and encode fusion strategies under different input conditions, select corresponding training and test sets for training each fusion model, and use them as a second-level decision for fault diagnosis. The multi-source fusion computing module is also used to input the preprocessed data into the primary diagnosis module to perform primary diagnosis of internal faults. The diagnosis results are converted into BPA under the identification framework to construct multiple evidence bodies required for fusion. The multi-source fusion computing module is also used to determine whether a decision conclusion is directly drawn if the conclusions of multiple primary diagnoses are consistent; otherwise, evidence body synthesis is performed. If there are no hidden conflicts between evidence bodies, a discount method based on evidence classification is used; otherwise, a principal component-based evidence discount method is used for synthesis, and a decision conclusion is drawn. The multimodal feature fusion diagnostic network module is used for overall network testing to obtain diagnostic results; The trend prediction and hierarchical adaptive early warning module is used to construct a trend prediction model that combines sliding window regression and exponential smoothing for key state parameters in transformer operation, and introduces an adaptive threshold adjustment mechanism to achieve hierarchical alarm output. The online incremental learning and cross-site model transfer module is used to introduce an online incremental learning mechanism and transfer learning technology to maintain prediction and diagnostic accuracy as factors such as operating conditions, seasons, and equipment aging change, and to transfer model parameters that have been optimized in one substation to a new site.

Images