Power transformer fault diagnosis method based on dynamic multi-modal information fusion

Abstract

The invention relates to the technical field of power transformer fault diagnosis, and discloses a power transformer fault diagnosis method based on dynamic multi-modal information fusion. According to the method, firstly, a layer of modal selection network is constructed at the front end of a model to serve as a first layer of decision, monitoring information input on site can be automatically screened, and a diagnosis mode is dynamically selected; and secondly, diagnosis models under different mode input conditions are established and trained, fault diagnosis is carried out by adopting corresponding fusion strategy models aiming at various inputs, and improvement of information utilization efficiency and diagnosis accuracy under different data monitoring conditions is realized. According to the method, the flexibility of a traditional multi-source information or multi-mode diagnosis method can be improved, and normal use of the diagnosis method can be ensured as much as possible under the condition of lack of data.

Description

Technical Field

[0001] The present invention relates to the technical field of power transformer fault diagnosis, and in particular to a power transformer fault diagnosis method based on dynamic multi-modal information fusion. Background Art

[0002] Traditional single-mode transformer fault diagnosis methods mainly analyze certain specific monitoring information, such as typical methods such as dissolved gas analysis in oil, partial discharge monitoring, infrared recognition technology, etc. Although these single-mode methods are effective in specific scenarios, there is a lack of coupling between diagnostic methods of different modes during the diagnosis process, the fault judgment boundary is fuzzy, the diagnosis has certain limitations, and the accuracy of the diagnostic results is difficult to meet the requirements of actual working conditions.

[0003] In recent years, in order to solve the above problems, many scholars have applied multi-source information fusion technology to transformer fault diagnosis. For example, Li Gang et al. used oil chromatography data as the main body, combined with transformer partial discharge, historical fault information, family quality history and other data, and adopted a multi-level decision fusion method to diagnose transformer faults; Huang Xinbo et al. combined electrical test data and oil chromatography data as tuples, performed feature-level fusion, input them into the same model for multiple diagnoses, and fused all outputs using improved evidence theory to obtain the final diagnosis result; Zhan Zhongqiang et al. used extended fuzzy logic to map and calculate oil chromatography data and core grounding current data, and then used evidence theory to perform decision fusion to complete fault diagnosis. The above studies can effectively prove that information from different sources is complementary. Compared with the fault diagnosis method of single information, the multi-type information fusion diagnosis method can effectively improve the accuracy of transformer fault diagnosis.

[0004] However, the information used by the above multi-source information fusion diagnosis technology is mostly structured data based on the content of dissolved gas in oil, and unstructured data such as infrared images are not considered. The diagnosis of thermal property faults only relies on the dissolved gas data in oil, and the transformer information is still not fully utilized, and the accuracy has room for improvement. In addition, under the actual on-site acquisition conditions, the types of deployed sensors and the data acquisition frequency are different, and the monitoring data has problems of abnormality or loss, resulting in the actual available multimodal information types being not fixed in the same period. However, the current diagnostic methods are all static, the model diagnosis mode is fixed, and the diagnostic flexibility is insufficient: when a certain mode is missing, the diagnostic method is no longer applicable; when the data is abnormal, the diagnostic accuracy decreases, and it is difficult to meet the high requirements of the site. Therefore, it is necessary to fully consider various types of multimodal information while improving the flexibility and accuracy of the diagnostic method and its adaptability to the diverse data acquisition conditions on site. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned traditional static diagnosis model, the present invention adapts to different on-site monitoring conditions, makes full use of multimodal monitoring information, and improves the flexibility and accuracy of the transformer fault diagnosis method. A power transformer fault diagnosis method based on dynamic multimodal information fusion is proposed.

[0006] In order to achieve the above object, the present invention provides a method for power transformer fault diagnosis based on dynamic multimodal information fusion, the method comprising the following steps:

[0007] (1) The data collected under different conditions on site are grouped according to data anomalies or missing data, and aligned in time;

[0008] (2) Establish a dynamic decision network. First, consider the three types of modal inputs that exist on site: single, dual, and full. Establish a unique diagnostic strategy for each situation. Second, perform one-hot encoding on each strategy so that the model can distinguish different strategies. Finally, train the model so that it can automatically select the modality suitable for fusion based on the characteristics of each modal input data, activate the corresponding diagnostic strategy, and output the correct one-hot encoding.

[0009] (3) Establish multiple fusion diagnostic models, establish corresponding diagnostic models according to all diagnostic strategies, and use their corresponding data for training;

[0010] (4) Construct a complete dynamic multimodal fusion diagnosis network. First, use the dynamic decision network as the first-layer decision network to identify the data and output the corresponding diagnosis strategy code. Secondly, correspond all diagnostic models to the codes one by one, and construct the second-layer decision diagnosis network. According to the code output by the first layer, enable the corresponding diagnosis model to perform multimodal fault diagnosis.

[0011] Preferably, the processing process for multimodal data includes:

[0012] Taking a certain municipal bureau in the database as an example, the DGA data collection frequency is 1 / h. When DGA data exists, it is taken as the center and the time window is set to 1h, that is, 0.5h before and after the DGA data collection time. It is considered that other modal data and DGA data within this period belong to the same period of data; when DGA is missing, infrared / partial discharge is used for single-mode fault diagnosis;

[0013] Since the external and internal components are monitored independently, the data of the two are not common, and the model will diagnose them separately. In order to make the model distinguish between internal and external data, the internal data is labeled "0" and the external data is labeled "1";

[0014] The data are aligned in time according to the above divisions, and the data are grouped according to the data missing situation.

[0015] Preferably, in order to achieve better model training effect, the original data needs to be preprocessed, including:

[0016] In the analysis of dissolved gas in oil, directly inputting the characteristic gas content will lead to poor training results due to the large dispersion of values;

[0017] The percentages of five characteristic gases, H2, CH4, C2H6, C2H4, and C2H2, were selected as model inputs to optimize the training effect;

[0018] Since infrared field data is complex and is limited by factors such as equipment and environment, low-pixel images will affect the training effect of the deep learning model. Super-resolution reconstruction of infrared images can improve image clarity and facilitate subsequent identification.

[0019] Preferably, in order to denoise the raw data of partial discharge and improve the data quality, the data source of each partial discharge sample is a 5s raw signal. After denoising, its pulse and phase are extracted, and the phase of 0 to 360° and the discharge amplitude of 0 to vmax are discretized into 90 interval spectra. The number of discharges is used as the matrix element to form a 90×90 matrix as the PRPD spectrum.

[0020] Preferably, the process of establishing a dynamic decision network includes:

[0021] Assume that there are m modalities to be fused, expressed as X:

[0022] X=(x1,x2,...,x m ) (1)

[0023] Considering all input conditions, a unique fusion strategy F is designed for the network under each input condition, denoted as F i (A):

[0024] F i (A)(i=1,2,...,2 m-1 ) (2)

[0025] Among them, A is a subset of the full mode X, F i is the i-th fusion strategy.

[0026] Preferably, a trained gating network function G(x) is embedded in a suitable position in the model, which can automatically select the modality suitable for fusion according to 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 (one-hot encoding) will be generated, expressed as g:

[0027] g = one-hot(argmax i G(x) i ) (3)

[0028] The fusion strategy F of a branch is determined by outputting one-hot encoding. i (A) and get the final result.

[0029] Preferably, the gating network selects a multi-layer perceptron MLP.

[0030] Preferably, considering different data monitoring situations on site, the second-level fusion diagnosis system is established according to the following three strategies:

[0031] Strategy 1: Retain the ability of a single modal model to address specific problems and use it when only a single type of data is available, i.e., use the appropriate models for each of the three modalities for diagnosis;

[0032] Strategy 2: When a certain modality data is abnormal or missing, the oil chromatography data is mainly used, and a decision-level fusion method is adopted with either partial discharge or infrared modality to improve the diagnostic accuracy. If the oil chromatography data is missing, the partial discharge and infrared are diagnosed using single-modality models respectively.

[0033] Strategy 3: When there is no missing data or abnormality, a hybrid fusion method is adopted to make full use of the complementarity between the modes. The diagnostic output after the fusion of oil chromatography and partial discharge data features and the infrared model recognition output are fused for diagnosis using evidence theory to obtain the final diagnostic result.

[0034] Preferably, Yolov5 is used for infrared technology, and yolov5s is used as the pre-training parameter weight; repVGG is used for partial discharge monitoring, and the learning rate is 1×10 -4 , the moment estimation factor is (0.9, 0.999); SVM is used for the analysis of dissolved gas in oil, the RBF kernel function is used, the kernel function parameter is 0.5, and the regularization coefficient is 500.

[0035] Preferably, the feature fusion diagnosis of dissolved gas analysis in oil and partial discharge monitoring adopts a deep belief network DBN, a classic 5-layer model, a maximum number of iterations of RBM is 100, a learning rate is 0.1, and a momentum parameter is 0.9.

[0036] Compared with the prior art, the technical solution of the present invention has the following advantages:

[0037] (1) The dynamic multimodal fusion diagnosis model constructed by the present invention screens all data and can dynamically select appropriate diagnosis methods for situations where each data is missing or abnormal, thereby ensuring the normal use of the diagnosis method as much as possible. Compared with the traditional static model, it is more flexible and reduces the high dependence on data.

[0038] (2) The dynamic multimodal fusion diagnosis model constructed by the present invention takes into account more online monitoring information (infrared) compared with the traditional method, and improves the accuracy of thermal property fault diagnosis of power transformers. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the first layer dynamic decision network. DETAILED DESCRIPTION

[0040] The specific embodiments of the present invention are described in detail below. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0041] The present invention proposes a method for diagnosing power transformer faults based on dynamic multimodal fusion. First, a modal selection network is constructed at the front end of the model as the first layer of decision-making, which can autonomously screen the monitoring information input on site and dynamically select the diagnosis mode; secondly, a diagnosis model under different mode input conditions is established and trained, and corresponding fusion strategy models are used for fault diagnosis for various inputs, so as to improve the information utilization efficiency and diagnosis accuracy under different data monitoring conditions. The present invention can not only improve the flexibility of traditional multi-source information or multi-modal diagnosis methods, but also ensure that the diagnosis method can still be used normally as much as possible in the absence of data.

[0042] Processing of multimodal data:

[0043] Taking a certain municipal bureau in the database as an example, the DGA data collection frequency is 1 / h. When DGA data exists, it is taken as the center and the time window is set to 1h, that is, 0.5h before and after the DGA data collection time. It is considered that other modal data and DGA belong to the same period of data during this period. When DGA is missing, infrared / partial discharge is used for single-mode fault diagnosis.

[0044] Secondly, since the external and internal components are monitored independently, the data of the two are not common, and the model will diagnose them separately. In order to make the model distinguish between internal and external data, the internal data is labeled "0" and the external data is labeled "1".

[0045] The data are temporally aligned according to the above divisions and grouped according to the data missing situation (input is single / dual / full mode).

[0046] Finally, in order to make the model training more effective, the original data needs to be preprocessed:

[0047] In the analysis of dissolved gas in oil, directly inputting the characteristic gas content will lead to poor training effect due to the large dispersion of values. The present invention selects the percentage of the five characteristic gas contents of H2, CH4, C2H6, C2H4, and C2H2 as model input to optimize the training effect.

[0048] Due to the complexity of infrared field data and the limitations of equipment, environment and other factors, low-pixel images will affect the training effect of deep learning models. Super-resolution reconstruction of infrared images can improve image clarity and facilitate subsequent recognition.

[0049] Denoising the raw data of partial discharge to improve data quality. The data source of each partial discharge sample is a 5s raw signal (about 250 cycles). After denoising, its pulse and phase are extracted, and the phase of 0-360° and the discharge amplitude of 0-vmax (maximum discharge amplitude within 5s) are discretized into 90 interval spectra. The number of discharges is used as the matrix element to form a 90×90 matrix as the PRPD spectrum.

[0050] 2. Establish a dynamic decision-making network

[0051] Assume that there are m modalities to be fused, expressed as X:

[0052] X=(x1,x2,...,x m ) (1)

[0053] Considering all input conditions, a unique fusion strategy F is designed for the network under each input condition, denoted as F i (A):

[0054] F i (A)(i=1,2,...,2 m-1 ) (2)

[0055] Among them, A is a subset of the full mode X, F i is the i-th fusion strategy.

[0056] A trained gating network function G(x) is embedded in the appropriate position of the model, which can automatically select the modality suitable for 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 (one-hot encoding) will be generated, represented as g:

[0057] g = one-hot(argmax i G(x) i ) (3)

[0058] The fusion strategy F of a branch is determined by outputting one-hot encoding. i (A) and get the final result.

[0059] Taking m=2 as an example, the specific decision process is as follows Figure 1 shown.

[0060] Each fusion strategy is a pre-developed and trained model that has taken into account all possible modal combinations. Then, each fusion only selects the most appropriate branch strategy to maximize the simplification of calculations.

[0061] In actual field monitoring conditions, abnormal data may occasionally appear due to the influence of the environment or equipment, such as the presence of obstructions in infrared images, excessive local discharge noise, and still not being able to be used normally after denoising, and abnormal DGA gas content collection, etc. These abnormal data and normally collected data are used to train the network to ensure that the model can correctly identify and select the corresponding decision under the input conditions of missing data and abnormal data.

[0062] 3. Establish a second-tier integrated diagnostic system

[0063] Considering different data monitoring conditions on site, the present invention proposes three dynamic fusion strategies based on transformer fault diagnosis, as shown below:

[0064] Strategy 1: Retain the ability of a single modal model to address specific problems and use it when only a single type of data is available, i.e., use the appropriate models for each of the three modalities for diagnosis;

[0065] One-hot encoding of the output under this strategy

[0066] DGA: 100

[0067] PD: 010

[0068] IR:001

[0069] Strategy 2: When a certain modality data is abnormal or missing, the oil chromatography data is mainly used, and a decision-level fusion method is adopted with either partial discharge or infrared modality to improve the diagnostic accuracy. If the oil chromatography data is missing, the partial discharge and infrared are diagnosed using single-modality models respectively.

[0070] One-hot encoding of the output under this strategy

[0071] DGA / PD: 110

[0072] DGA / IR:101

[0073] Strategy 3: When there is no missing data or abnormality, a hybrid fusion method is adopted to make full use of the complementarity between the modes. The diagnostic output after the fusion of oil chromatography and partial discharge data features and the infrared model recognition output are fused for diagnosis using evidence theory to obtain the final diagnostic result.

[0074] One-hot encoding of the output under this strategy

[0075] DGA / PD / IR: 111

[0076] The data of each group are used to train each diagnostic model.

[0077] 4. Build a complete dynamic multimodal fusion diagnosis network

[0078] In the present invention, Yolov5 is selected as infrared technology, and yolov5s is used as the pre-training parameter weight; repVGG is selected for partial discharge monitoring, the learning rate is 1×10-4, and the moment estimation factor is (0.9, 0.999); SVM is selected for dissolved gas analysis in oil, RBF kernel function is used, the kernel function parameter is 0.5, and the regularization coefficient is 500. The deep belief network DBN is used for the feature fusion diagnosis of dissolved gas analysis in oil and partial discharge monitoring, and the classic 5-layer model is used. The maximum number of iterations of RBM is 100, the learning rate is 0.1, and the momentum parameter is 0.9. The gated network selects the multi-layer perceptron MLP.

[0079] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various technical features in any other suitable manner, and these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention.

Claims

1. A power transformer fault diagnosis method based on dynamic multimodal information fusion, It is characterized in that The method comprises the following steps: (1) The data collected under different conditions on site are grouped according to data anomalies or missing data, and aligned in time; (2) Establish a dynamic decision network. First, consider the three types of modal inputs that exist on site: single, dual, and full. Establish a unique diagnostic strategy for each situation. Second, perform one-hot encoding on each strategy so that the model can distinguish different strategies. Finally, train the model so that it can automatically select the modality suitable for fusion based on the characteristics of each modal input data, activate the corresponding diagnostic strategy, and output the correct one-hot encoding. (3) Establish multiple fusion diagnostic models, establish corresponding diagnostic models according to all diagnostic strategies, and use their corresponding data for training; (4) Construct a complete dynamic multimodal fusion diagnosis network. First, use the dynamic decision network as the first-layer decision network to identify the data and output the corresponding diagnosis strategy code. Secondly, correspond all diagnostic models to the codes one by one, and construct the second-layer decision diagnosis network. According to the code output by the first layer, enable the corresponding diagnosis model to perform multimodal fault diagnosis.

2. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 1, It is characterized in that The processing of multimodal data includes: Taking a certain municipal bureau in the database as an example, the DGA data collection frequency is 1 / h. When DGA data exists, it is taken as the center and the time window is set to 1h, that is, 0.5h before and after the DGA data collection time. It is considered that other modal data and DGA data within this period belong to the same period of data; when DGA is missing, infrared / partial discharge is used for single-mode fault diagnosis; Since the external and internal components are monitored independently, the data of the two are not common, and the model will diagnose them separately. In order to distinguish the internal and external data, the internal data is labeled "0" and the external data is labeled "1"; The data are aligned in time according to the above divisions, and the data are grouped according to the data missing situation.

3. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 2, It is characterized in that In order to achieve better model training results, the original data needs to be preprocessed, including: In the analysis of dissolved gas in oil, directly inputting the characteristic gas content will lead to poor training results due to the large dispersion of values; Select H 2 , CH 4 , C 2 H 6 , C 2 H 4 , C 2 H 2 The percentages of five characteristic gas contents were used as model input to optimize the training effect; Since infrared field data is complex and is limited by factors such as equipment and environment, low-pixel images will affect the training effect of the deep learning model. Super-resolution reconstruction of infrared images can improve image clarity and facilitate subsequent identification.

4. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 3, It is characterized in that In order to denoise the raw data of partial discharge and improve the data quality, the data source of each partial discharge sample is a 5s original signal. After denoising, its pulse and phase are extracted, and the phase of 0-360° and the discharge amplitude of 0-vmax are discretized into 90 interval spectra. The number of discharges is used as the matrix element to form a 90×90 matrix as the PRPD spectrum.

5. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 1, It is characterized in that The process of building a dynamic decision network includes: Assume that there are m modalities to be fused, expressed as X: X=(x 1 ,x 2 ,...,x m ) (1) Considering all input conditions, a unique fusion strategy F is designed for the network under each input condition, denoted as F i (A): F i (A)(i=1,2,...,2 m-1 ) (2) Among them, A is a subset of the full mode X, F i is the i-th fusion strategy.

6. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 5, It is characterized in that A trained gating network function G(x) is embedded in the appropriate position of the model. It can automatically select the modality suitable for 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 (one-hot encoding) will be generated, expressed as g: g=one-hot(argmax i G(x) i ) (3) The fusion strategy F of a branch is determined by outputting one-hot encoding. i (A) and get the final result.

7. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 6, It is characterized in that The gating network selects a multi-layer perceptron MLP.

8. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 1, It is characterized in that Considering the different data monitoring situations on site, the second-level fusion diagnosis system is established according to the following three strategies: Strategy 1: Retain the ability of a single modal model to address specific problems and use it when only a single type of data is available, i.e., use the appropriate models for each of the three modalities for diagnosis; Strategy 2: When a certain modality data is abnormal or missing, the oil chromatography data is mainly used, and a decision-level fusion method is adopted with either partial discharge or infrared modality to improve the diagnostic accuracy. If the oil chromatography data is missing, the partial discharge and infrared are diagnosed using single-modality models respectively. Strategy 3: When there is no missing data or abnormality, a hybrid fusion method is adopted to make full use of the complementarity between the modes. The diagnostic output after the fusion of oil chromatography and partial discharge data features and the infrared model recognition output are fused for diagnosis using evidence theory to obtain the final diagnostic result.

9. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 1, It is characterized in that Infrared technology uses Yolov5, and yolov5s is used as the pre-trained parameter weight; partial discharge monitoring uses repVGG, and the learning rate is 1×10 -4 , the moment estimation factor is (0.9, 0.999); SVM was used for the analysis of dissolved gas in oil, with the RBF kernel function, a kernel function parameter of 0.5, and a regularization coefficient of 500.

10. The power transformer fault diagnosis method based on dynamic multimodal information fusion according to claim 9, It is characterized in that The deep belief network DBN is used for the feature fusion diagnosis of dissolved gas analysis and partial discharge monitoring in oil. The classic 5-layer model is adopted, the maximum number of iterations of RBM is 100, the learning rate is 0.1, and the momentum parameter is 0.9.

Images