Transformer oil chromatographic data analysis method and system based on deep belief network

Abstract

The invention discloses a transformer oil chromatographic data analysis method and system based on a deep belief network, and the method comprises the steps: determining a non-coding ratio as a characteristic parameter of a DBN model through analyzing the concentration and proportion of dissolved gas in transformer insulating oil, building the DBN model through employing a multilayer restricted Boltzmann machine, and carrying out the analysis of the chromatographic data of the transformer oil. And a strategy of layer-by-layer pre-training and overall fine tuning is adopted, and network weight is optimized through a contrast divergence algorithm, so that detailed feature differences of fault types are extracted. According to the invention, high-precision diagnosis of transformer fault types can be realized, including distinguishing of various fault types, such as low-temperature, medium-temperature and high-temperature thermal faults, partial discharge, low-energy discharge, high-energy discharge and the like. Compared with a traditional ratio method, the method has remarkable advantages in the aspects of diagnosis accuracy and stability. Particularly, when multiple fault types are processed, the DBN model can effectively identify and distinguish complex fault features, and the definition and accuracy of diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis of power equipment, and particularly to a method and system for analyzing transformer oil chromatogram data based on a deep belief network. Background Art

[0002] Power transformers are one of the key equipment in the power system, and their operating status determines whether the power grid can supply power reliably. Due to electrical, mechanical, and thermal stresses, transformers decompose small amounts of gases dissolved in the insulating oil. The concentration and relative proportion of the by-product gases are closely related to the insulation conditions of the transformer.

[0003] The gases present in the transformer insulating oil are usually hydrogen, methane, ethylene, ethane, and acetylene. The gas concentrations under early fault conditions (such as partial discharge, thermal heating, and arc) will increase according to the fault type. Detecting a certain level of gas generated by an operating oil-filled transformer can be used as the first sign of a fault.

[0004] Most traditional methods for analyzing oil chromatogram data rely on empirical rules and predefined gas ratios. Among these methods, the most commonly used ratio methods such as the Delphi method and the Combi method establish a connection between the concentration ratios of different gases in the oil and the fault types. However, these methods have the following problems: low accuracy: traditional methods rely on empirical rules and have limited detection capabilities for new or complex faults. Poor real-time performance: these methods are usually based on manual processing and periodic sampling, making it difficult to achieve real-time monitoring and immediate response. Poor data processing ability: traditional methods cannot fully mine the complex patterns contained in the oil chromatogram data, especially under large data volumes and high-dimensional features, with low analysis efficiency and high error rates.

[0005] Due to the lack of learning ability and processing efficiency, the current methods still have limitations in popularization and application. Due to differences in fault location, fault intensity, and transformer capacity, etc., transformer fault modes show diversity. Therefore, higher requirements are put forward for transformer fault diagnosis.

[0006] The deep belief network is an unsupervised learning model stacked by multiple autoencoders, which can be pre-trained on unlabeled data to extract high-order features of the data. The DBN first pre-trains the input data through unsupervised learning to automatically discover the potential structure in the data without manual annotation. For the transformer oil chromatographic data, this can avoid the cumbersome labeling process for different types of faults. The DBN can automatically learn the high-dimensional features of the transformer oil chromatographic data and compress the data into more discriminative low-dimensional features through dimensionality reduction operations, making subsequent fault diagnosis more accurate. The DBN can handle complex data with high non-linearity, such as the gas concentration changes in transformer oil chromatographic data, which gives it an advantage over traditional linear models in fault classification and prediction. After sufficient training, the DBN can automatically classify according to the features in the oil chromatographic data, perform fault type identification, fault time prediction and fault severity assessment, significantly improving the real-time monitoring ability of the transformer health status.

[0007] The present invention applies the deep belief network to transformer fault diagnosis and identification. A multi-hidden layer deep belief learning model is constructed, significantly improving the accuracy of power transformer fault diagnosis. Summary of the Invention

[0008] An object of the present invention is to provide a method for analyzing transformer oil chromatographic data based on a deep belief network on the one hand, and a system for analyzing transformer oil chromatographic data based on a deep belief network on the other hand. This system and method can apply the deep belief network to transformer fault diagnosis and identification, construct a multi-hidden layer deep belief learning model, and significantly improve the accuracy of power transformer fault diagnosis.

[0009] To achieve this object, a system for analyzing transformer oil chromatographic data based on a deep belief network designed by the present invention is characterized in that it includes a data preprocessing module and a model construction module;

[0010] The data preprocessing module is used to preprocess the dissolved gas data in the transformer oil collected in real time and the historical dissolved gas data in the transformer oil to obtain the preprocessed historical dissolved gas data and the preprocessed real-time dissolved gas data;

[0011] The model construction module is used to input the preprocessed historical dissolved gas data into the deep belief network model, pre-train the deep belief network model and adjust the network weights of the deep belief network model, so as to obtain a trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains a gas concentration non-coding ratio associated with the transformer fault mode according to the extracted gas concentration features, and classifies the transformer faults according to the gas concentration non-coding ratio associated with the transformer fault mode.

[0012] Further, the dissolved gas data collected in real time from transformer oil includes: H in transformer oil 2 、CH 4 、C 2 H 4 、C 2 H 6 、C 2 H 2 concentrations.

[0013] Further, the non-coded ratio of gas concentrations associated with transformer fault modes includes: concentration of CH in transformer oil 4 / concentration of H in transformer oil 2 、concentration of C 2 H 4 / concentration of C 2 H 2 in transformer oil、concentration of C 2 H 4 / concentration of C 2 H 6 in transformer oil、concentration of C 2 H 2 / concentration of (C1 + C2) in transformer oil、concentration of H in transformer oil 2 / concentration of (H 2 + C1 + C2) in transformer oil、concentration of C 2 H 4 / concentration of (C1 + C2)、concentration of CH in transformer oil 4 / concentration of (C1 + C2) in transformer oil、concentration of C 2 H 6 / concentration of (C1 + C2) in transformer oil and concentration of (CH 4 + C 2 H 4 ) / concentration of (C1 + C2) in transformer oil, where C1 represents CH 4 ,C2 represents the sum of C 2 H 4 、C 2 H 6 and C 2 H 2 .

[0014] Further, the method for preprocessing the collected dissolved gas data includes: denoising the collected dissolved gas data, standardizing it, filling in the missing values in the data by the mean interpolation method, and smoothing the dissolved gas data.

[0015] Further, the method for denoising the collected dissolved gas data specifically includes: using a low-pass filter to remove high-frequency noise, and the processing formula is:

[0016]

[0017] where y(t) is the historical dissolved gas data after denoising, x(t) is the historical dissolved gas data before denoising, h(k) is the filter coefficient, and K is the order of the filter;

[0018] The method for normalizing the collected dissolved gas data specifically includes:

[0019]

[0020] where x i is the historical dissolved gas data before normalization, x′ i is the historical dissolved gas data after normalization, μ i is the mean value of the concentration of gas i, and σ i is the standard deviation of the concentration of gas i;

[0021] The method for filling the missing values in the collected dissolved gas data by mean interpolation specifically includes: using the mean value of a column of data to fill the missing values in a column of data,

[0022]

[0023] where x missing is the missing historical dissolved gas data, N is the number of known historical dissolved gas data points, and x i is the value of the known historical dissolved gas data points;

[0024] The method for smoothing the collected dissolved gas data specifically includes: using the moving average method, and at each historical dissolved gas data point, calculating the average value of multiple surrounding historical dissolved gas data points:

[0025]

[0026] where y(t) is the value of the historical dissolved gas data after smoothing, the smoothing result at time t, x(t-k) is the value of the data at time t-k, k represents the offset of time, and M represents the number of historical dissolved gas data points used in calculating the average value.

[0027] Further, the method for pre-training the deep belief network model includes: inputting the preprocessed dissolved gas data into the deep belief network model, and pre-training the deep belief network model through unsupervised learning to extract potential features from the dissolved gas data:

[0028] h = f(Wx + b)

[0029] Among them, h is the output of the feature extraction result of historical dissolved gas data, W is the classification layer weight matrix set according to the importance of gas concentration, x is the historical dissolved gas data, b is a preset constant bias term, and the bias term is used to adjust the activation value of the hidden layer to ensure the non-linear representation ability of the network. f is a preset activation function, and the activation function introduces non-linear transformation to enable the deep belief network model to learn complex patterns and relationships;

[0030] When pre-training the deep belief network model, the deep belief network model will reconstruct the preprocessed dissolved gas data and use the mean square error to reduce the difference between the data distribution and the model generation distribution:

[0031]

[0032] Among them, E reconstruction is the mean square error, x i is the preprocessed historical dissolved gas data, is the reconstructed historical dissolved gas data, and N is the number of historical dissolved gas data samples.

[0033] Furthermore, the method for adjusting the network weights of the deep belief network model to obtain the trained deep belief network model includes: further training the pre-trained deep belief network model using the contrastive divergence algorithm:

[0034] ΔW = ∈(<Hx T > data - <Hx T > model )

[0035] Among them, ΔW is the value after the network weight is updated, representing the relationship between the gas concentration data and the hidden layer features. H is the input of the dissolved gas concentration in the visible layer, x T is the activation value of the hidden layer, ∈ is the learning rate, <Hx T > data is the expectation of the data distribution, and the expectation of the data distribution represents the relationship between the visible layer and the hidden layer, reflecting the distribution characteristics of the real data; <Hx T > model is the expectation of the model distribution, and the expectation of the model distribution represents the relationship between the data generated by the current model and the hidden layer, reflecting the generation distribution characteristics of the model. Finally, the trained deep belief network model is obtained.

[0036] Furthermore, the method for the trained deep belief network model to extract gas concentration features from the preprocessed dissolved gas data and perform fault classification includes: the trained deep belief network model automatically extracts high-dimensional features from the oil chromatographic data, and the high-dimensional features include gas concentration, change trend, and the non-coding ratio of gas concentration of the gas. The trained deep belief network model predicts the fault type according to the gas concentration, change trend, and the non-coding ratio of gas concentration. The fault classification formula is: where is the predicted fault type, W is the classification layer weight matrix set according to the importance of gas concentration, h is the output of the feature extraction result of the dissolved gas data, b is a preset constant bias term, and the bias term is used to adjust the activation value of the hidden layer to ensure the non-linear representation ability of the network; and through regression analysis, predict the time of fault occurrence: where is the predicted time of fault occurrence.

[0037] Furthermore, a method for analyzing transformer oil chromatogram data based on the transformer oil chromatogram data analysis system includes:

[0038] Preprocess the real-time collected dissolved gas data in transformer oil and the historical dissolved gas data in transformer oil to obtain the preprocessed historical dissolved gas data and the preprocessed real-time dissolved gas data;

[0039] Input the preprocessed historical dissolved gas data into the deep belief network model, pre-train the deep belief network model and adjust the network weights of the deep belief network model to obtain the trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains the non-coding ratio of gas concentration associated with the transformer fault mode according to the extracted gas concentration features, and classifies the transformer faults according to the non-coding ratio of gas concentration associated with the transformer fault mode.

[0040] Advantages of the present invention: By analyzing the concentration and ratio of dissolved gases in transformer insulating oil, the present invention determines the non-coding ratio as the characteristic parameter of the DBN model. A DBN model is constructed using a multi-layer restricted Boltzmann machine, and a strategy of layer-by-layer pre-training and overall fine-tuning is adopted. The network weights are optimized by the contrastive divergence algorithm to extract the detailed feature differences of the fault types. The present invention realizes high-precision diagnosis of transformer fault types, including the distinction of various fault types, such as low, medium, and high-temperature thermal faults, partial discharge, low-energy discharge, and high-energy discharge, etc. Compared with the traditional ratio method, the present invention shows significant advantages in terms of diagnostic accuracy and stability. Especially when dealing with multiple fault types, the DBN model can effectively identify and distinguish complex fault features, improving the clarity and accuracy of diagnosis. Description of the Drawings

[0041] Figure 1 This is a structural diagram of an application program for a method of analyzing transformer oil chromatogram data based on a deep belief network according to the present invention.

[0042] Figure 2 This is a block diagram of a system for analyzing transformer oil chromatogram data based on a deep belief network according to the present invention.

[0043] Figure 3 This is a block diagram of a modeling process for analyzing transformer oil chromatogram data based on a deep belief network according to the present invention;

[0044] Figure 4 This is a schematic structural diagram of the present invention. Detailed Description of the Invention

[0045] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0046] Embodiment 1

[0047] As Figure 4 shown, a system for analyzing transformer oil chromatogram data based on a deep belief network includes a data preprocessing module and a model construction module;

[0048] The data preprocessing module is used to preprocess the dissolved gas data in transformer oil collected in real time and the historical dissolved gas data in transformer oil, so as to obtain the preprocessed historical dissolved gas data and the preprocessed real-time dissolved gas data;

[0049] The model construction module is used to input the preprocessed historical dissolved gas data into the deep belief network model, pre-train the deep belief network model and adjust the network weights of the deep belief network model, so as to obtain a trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains a gas concentration non-coding ratio associated with the transformer fault mode according to the extracted gas concentration features, and classifies the transformer faults according to the gas concentration non-coding ratio associated with the transformer fault mode.

[0050] As Figure 1As shown in the figure, the typical application structure of a deep belief network is a deep neural network composed of several restricted Boltzmann machines (RBMs) and a classification output layer. The restricted Boltzmann machines with hierarchical pre-training construct the network structure of the deep belief network. Each restricted Boltzmann machine consists of a visible layer and a hidden layer, which are used for feature extraction and representation. They are used to represent higher-level and multi-level features of the input pattern. The classification output is used to integrate the neural network layers and perform classification using the learned features. The deep belief network obtains weights layer by layer through CD training. Since each restricted Boltzmann machine is trained independently, the obtained weights can only optimize the features layer by layer and cannot guarantee the performance of the entire deep belief network. Therefore, after training each layer by the CD algorithm, the weights of the deep belief network will be fine-tuned to inversely fit the entire deep belief network model. The fine-tuning is to use the stochastic gradient descent method to search for the minimum error between the network output label and the standard label.

[0051] As Figure 2 shown, a block diagram of a transformer oil chromatographic data analysis system based on a deep belief network includes:

[0052] (1) Data acquisition module:

[0053] It is used to collect the dissolved gas data in the transformer oil in real time and transmit the collected data to the data preprocessing module.

[0054] (2) Data preprocessing module:

[0055] It is used to preprocess the collected dissolved gas data, such as denoising, standardizing, filling missing values, and data smoothing, to ensure the quality and usability of the data.

[0056] (3) Deep belief network (DBN) model module:

[0057] It uses the preprocessed dissolved gas data to construct a deep belief network (DBN) model for data feature extraction and fault classification.

[0058] (4) Fault diagnosis module:

[0059] Based on the results output by the DBN model, it judges the fault type, location, and severity of the transformer.

[0060] (5) Alarm and health assessment module:

[0061] It is used to monitor the health status of the transformer in real time, automatically trigger an alarm when a fault is detected, and generate a health assessment report to assist maintenance personnel in dealing with the fault in a timely manner.

[0062] In the above technical solution, the dissolved gas data in the transformer oil collected in real time includes: H in the transformer oil 2, CH 4 , C 2 H 4 , C 2 H 6 , C 2 H 2 concentration.

[0063] Real-time collection of dissolved gas data in transformer oil includes hydrogen (H 2 ), methane (CH 4 ), ethylene (C 2 H 4 ), ethane (C 2 H 6 ), acetylene (C 2 H 2 ), carbon monoxide (CO) and carbon dioxide (CO 2 ), etc. The concentration data of common gases are expressed as: X = {x 1 , x 2 , …, x n}, where x i is the concentration of the i-th gas, and n is the number of gas types. The collected oil chromatogram data can be collected regularly (such as every hour, daily), or the sampling frequency can be dynamically adjusted according to the operating status of the transformer, load changes, etc.

[0064] Based on a large number of transformer simulation experiments, this invention mainly studies the laws of the following four transformer faults and the concentration of dissolved gases in transformer oil:

[0065] 1. Thermal faults in transformer insulating oil: Insulating mineral oil will release a large amount of CH 4 , some C 2 H 4 , C 2 H 6 and a small amount of H 2 . Taking local heating at 600 °C as an example, each gram of transformer oil will produce 5.848 mg of CH 4 , 3.247 mg of C 2 H 4 , 2.601 mg of C 2 H 6 , 0.320 mg of H 2 .

[0066] 2. Arc discharge in insulating oil: Insulating mineral oil will produce gases mainly H 2 and C 2 H 2 , with a small amount of CH 4 and C 2 H 4 . For example, when 100 L of insulating oil is under arc discharge, 54 - 74 L of H2 ,12 to 24 L of C 2 H 2 ,0 to 3 L of CH 4 ,0 to 1 L of C 2 H 2 。

[0067] 3. Partial discharge: Generates relatively more H 2 and CH 4 ,without C 2 H 2 。For example, 50% of the gas generated by partial discharge is H 2 and 45% is CH 4 。

[0068] 4. Spark discharge: Predominantly H 2 and C 2 H 2 ,with a small amount of CH 4 。For example, 77% of the gas generated by spark discharge is H 2 ,18% is C 2 H 2 ,4% is CH 4 。

[0069] Preprocessing the collected dissolved gas data can significantly improve the quality and reliability of the data, providing a more accurate and stable data basis for subsequent data analysis and processing.

[0070] According to the actual transformer fault statistics data, nine different combinations of DGA ratios related to the fault mode can be determined, and the non - code ratios are determined as the characteristic parameters of the deep belief network model.

[0071] In the above - mentioned technical solution, the non - code ratio of the gas concentration associated with the transformer fault mode includes: the concentration of CH 4 in transformer oil / the concentration of H 2 in transformer oil, the concentration of C 2 H 4 in transformer oil / the concentration of C 2 H 2 in transformer oil, the concentration of C 2 H 4 in transformer oil / the concentration of C 2 H 6 in transformer oil, the concentration of C 2 H 2 in transformer oil / the concentration of (C1 + C2) in transformer oil, the concentration of H 2 in transformer oil / the concentration of (H 2 + C1 + C2) in transformer oil, the concentration of C 2 H4 Concentration of / Concentration of (C1 + C2), CH in transformer oil 4 Concentration of / Concentration of (C1 + C2) in transformer oil, C in transformer oil 2 H 6 Concentration of / Concentration of (C1 + C2) in transformer oil and (CH 4 + C 2 H 4 ) Concentration of / Concentration of (C1 + C2) in transformer oil, where C1 represents CH 4 , C2 represents C 2 H 4 , C 2 H 6 and C 2 H 2 The sum of.

[0072] In the above technical solution, the method for preprocessing the collected dissolved gas data includes: denoising the collected dissolved gas data, standardizing the data, filling in the missing values in the data by the mean interpolation method, and smoothing the dissolved gas data.

[0073] As Figure 3 shown, a flow chart of data analysis and modeling for transformer oil chromatography based on a deep belief network includes: real-time collection of dissolved gas data in transformer oil through a gas chromatograph or an on-line sensor; preprocessing the data: using a low-pass filter to remove high-frequency noise to make the change of oil chromatography data more stable, standardizing each gas concentration data, filling in the missing values in the data by an interpolation method (such as linear interpolation or mean interpolation), and smoothing the time series data by the moving average method; then using the preprocessed dissolved gas data to construct and train a deep belief network model. After the deep belief network is trained, feature extraction and fault diagnosis of the preprocessed dissolved gas data are carried out. The deep belief network model classifies the transformer faults according to the concentration characteristics of the preprocessed dissolved gas data, and conducts fault prediction and evaluation, predicts the time of fault occurrence, and the deep belief network model monitors the transformer faults in real time. Once a transformer fails, the deep belief network model immediately triggers an alarm module and outputs the fault type, location and its severity.

[0074] In the above technical solution, the method for denoising the collected dissolved gas data specifically includes: using a low-pass filter to remove high-frequency noise, and the processing formula is:

[0075]

[0076] Among them, y(t) is the historical dissolved gas data after denoising, x(t) is the historical dissolved gas data before denoising, h(k) is the filter coefficient, and K is the order of the filter; the chromatographic data of transformer oil are often easily affected by errors of external environmental factors, resulting in unnecessary noise in the data, which will interfere with subsequent data analysis and fault diagnosis. Through data denoising processing, the accuracy and reliability of the dissolved gas data can be improved, providing a more accurate data basis for subsequent data analysis.

[0077] The method for standardizing the collected dissolved gas data specifically includes:

[0078]

[0079] Among them, x i is the historical dissolved gas data before standardization, x′ i is the historical dissolved gas data after standardization, μ i is the mean value of the concentration of gas i, and σ i is the standard deviation of the concentration of gas i;

[0080] In the chromatographic data of transformer oil, there are significant differences in the concentration units and orders of magnitude of gases. If directly input into the model, it may cause the influence of some gases on the model to be overamplified while the influence of other gases is ignored. Therefore, it is necessary to standardize the concentration data of each gas.

[0081] The method for filling the missing values in the collected dissolved gas data by the mean value interpolation method specifically includes: using the mean value of a column of data to fill the missing values in a column of data,

[0082]

[0083] Among them, x missing is the missing historical dissolved gas data, N is the number of known historical dissolved gas data points, and x i is the value of the known historical dissolved gas data points; the mean value interpolation method is a commonly used data interpolation method. It uses the average value of known data to estimate the value of unknown data points. By filling the missing values with the mean value interpolation method, the continuity of the data is maintained, and data breaks or jumps at the missing values are avoided. After filling the missing values, the data integrity is improved, which is conducive to subsequent data analysis and processing.

[0084] The method for smoothing the collected dissolved gas data specifically includes: using the moving average method, calculating the average value of multiple surrounding historical dissolved gas data points at each historical dissolved gas data point:

[0085]

[0086] Among them, y(t) is the value of the smoothed historical dissolved gas data, the smoothing result at time t, x(t - k) is the value of the data at time t - k, k represents the offset of time, and M represents the number of historical dissolved gas data points used when calculating the average value. During the acquisition process of oil chromatographic data, some short-term fluctuations or mutations may occur. These fluctuations sometimes do not represent actual faults or changes, but may be manifestations of noise. Smoothing processing can reduce the interference of such fluctuations on data analysis and prediction models.

[0087] In the above technical solution, the method for pre-training the deep belief network model includes: inputting the preprocessed dissolved gas data into the deep belief network model, and pre-training the deep belief network model through unsupervised learning to extract potential features from the dissolved gas data: h = f(Wx + b)

[0088] Among them, h is the output of the feature extraction result of the historical dissolved gas data, W is the classification layer weight matrix set according to the importance of gas concentration, x is the historical dissolved gas data, b is a preset constant bias term, the bias term is used to adjust the activation value of the hidden layer to ensure the non-linear representation ability of the network, f is a preset activation function, and the activation function introduces non-linear transformation to enable the deep belief network model to learn complex patterns and relationships;

[0089] When pre-training the deep belief network model, the deep belief network model will reconstruct the preprocessed dissolved gas data, and use the mean square error to reduce the difference between the data distribution and the model-generated distribution:

[0090]

[0091] Among them, E reconstruction is the mean square error, x i is the preprocessed historical dissolved gas data, is the reconstructed historical dissolved gas data, and N is the number of historical dissolved gas data samples. The mean square error minimizes the reconstruction error, enabling the network to gradually adjust the weights and biases, thereby learning the effective features of the data and improving the accuracy of prediction or reconstruction.

[0092] The deep belief network is stacked by multiple autoencoder layers (AE). Each autoencoder uses unsupervised learning methods to pre-train the data. The pre-training is carried out on each layer of the autoencoder in the deep belief network. The deep belief network learns the latent structure and patterns of the input data by training the autoencoders layer by layer. The goal of each layer of the autoencoder is to compress the input data into a low-dimensional representation (i.e., low-dimensional features), and be able to reconstruct the original data from this compressed representation. For transformer oil chromatographic data, each autoencoder can learn the corresponding patterns of gas concentration and compress them into low-dimensional feature representations.

[0093] When pre-training a deep belief network, first train the first-layer autoencoder, and then use the weights obtained from the training of the first layer as initialization to train the second layer, and so on. What each layer of the autoencoder learns is the low-dimensional features input from the upper layer, and maps them to a higher abstract level again, finally forming the multi-layer feature representation of the entire deep belief network.

[0094] In the above technical solution, the method for adjusting the network weights of the deep belief network model to obtain the trained deep belief network model includes: further training the pre-trained deep belief network model using the contrastive divergence algorithm:

[0095] △W = ∈(<Hx T > data - <Hx T > model )

[0096] Among them, ΔW is the value after the network weights are updated, representing the relationship between the gas concentration data and the hidden layer features, H is the input of the dissolved gas concentration of the transformer oil in the visible layer, x T is the activation value of the hidden layer, ∈ is the learning rate, <Hx T > data is the expectation of the data distribution, and the expectation of the data distribution represents the relationship between the visible layer and the hidden layer, reflecting the distribution characteristics of the real data; <Hx T > model is the expectation of the model distribution. The expectation of the model distribution represents the relationship between the data generated by the current model and the hidden layer, reflecting the generation distribution characteristics of the model. Finally, the trained deep belief network model is obtained. The contrastive divergence algorithm optimizes the weights by calculating the difference between the data distribution and the model distribution, and finally updates the weights of the network so that the data distribution generated by the model is closer to the real data distribution.

[0097] Use the deep belief network to model the preprocessed dissolved gas data and automatically extract multi-level features. The deep belief network automatically extracts high-order features from the input data through multiple autoencoder layers (AE). Each layer of the autoencoder is responsible for further learning more abstract and more compact representations from the features extracted from the previous layer, and inputs the extracted features into the classification layer of the deep belief network model. After the model training is completed, the deep belief network can automatically extract the most diagnostically valuable features from the transformer oil chromatogram data, and classify faults according to these features; make judgments based on the changes in different gas components and output the fault types.

[0098] In the above technical solution, the method for the trained deep belief network model to extract gas concentration features from the preprocessed dissolved gas data and perform fault classification includes: the trained deep belief network model automatically extracts high-dimensional features from the oil chromatographic data, and the high-dimensional features include gas concentration, change trend, and gas concentration non-coding ratio. The trained deep belief network model predicts the fault type according to the gas concentration, change trend, and gas concentration non-coding ratio. The fault classification formula is: Where, is the predicted fault type, W is the classification layer weight matrix set according to the importance of gas concentration, h is the output of the feature extraction result of the dissolved gas data, b is a preset constant bias term, and the bias term is used to adjust the activation value of the hidden layer to ensure the non-linear representation ability of the network; and through regression analysis, the time of fault occurrence is predicted: Where, is the predicted time of fault occurrence.

[0099] The method for the trained deep belief network model to predict the fault type according to the gas concentration, change trend, and gas concentration non-coding ratio is: first, collect the gas concentration in the transformer oil, analyze the change trend of the gas concentration. If the concentration of some gases increases abnormally within a period, it indicates that the transformer may have a fault. After initially judging that the transformer has a fault, then judge the transformer fault type by calculating the gas concentration non-coding ratio. Combining the gas concentration change trend and the gas concentration non-coding ratio can more accurately predict the fault type. Taking the partial discharge fault as an example, the concentrations of methane and hydrogen increase, and the non-coding ratios CH 4 / H 2 、C 2 H 2 / CH 4 are relatively low.

[0100] The trained deep belief network model can monitor the transformer status in real time, input the real-time collected oil chromatographic data into the DBN model of the deep belief network for fault diagnosis. If an anomaly or fault is detected, the trained deep belief network will immediately trigger an alarm and notify the operation and maintenance personnel.

[0101] Embodiment 2

[0102] According to the transformer oil chromatographic data analysis method of the transformer oil chromatographic data analysis system based on the deep belief network, it is characterized in that it includes:

[0103] Preprocess the real-time collected dissolved gas data in the transformer oil and the historical dissolved gas data in the transformer oil to obtain the preprocessed historical dissolved gas data and the preprocessed real-time dissolved gas data;

[0104] Input the preprocessed historical dissolved gas data into the deep belief network model, pre-train the deep belief network model, and adjust the network weights of the deep belief network model to obtain a trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains a gas concentration non-coding ratio associated with the transformer fault mode according to the extracted gas concentration features, and classifies the transformer faults according to the gas concentration non-coding ratio associated with the transformer fault mode.

[0105] Embodiment 3

[0106] A computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method described in Embodiment 2 are implemented.

[0107] The above is only the preferred embodiment of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0108] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0109] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a system for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0110] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction system, and the instruction system implements the functions in the processFigure 1 one process or multiple processes and / or blocks Figure 1 the functions specified in one block or multiple blocks.

[0111] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide for implementing in the process Figure 1 one process or multiple processes and / or blocks Figure 1 the steps of the functions specified in one block or multiple blocks.

[0112] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

Claims

1. A transformer oil chromatographic data analysis system based on deep belief network, characterized in that: It includes data preprocessing module and model building module; The data preprocessing module is used to preprocess the real-time collected data of dissolved gas in transformer oil and the historical data of dissolved gas in transformer oil to obtain the preprocessed historical data of dissolved gas and the preprocessed real-time data of dissolved gas; The model building module is used to input the preprocessed historical dissolved gas data into the deep belief network model, pre-train the deep belief network model and adjust the network weights of the deep belief network model to obtain a trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains the gas concentration non-coding ratio associated with the transformer fault mode based on the extracted gas concentration features, and classifies the transformer fault according to the gas concentration non-coding ratio associated with the transformer fault mode.

2. A transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The dissolved gas data in transformer oil collected in real time include: the concentrations of H2, CH4, C2H4, C2H6, and C2H2 in transformer oil.

3. A transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The non-coded ratios of gas concentrations associated with the transformer failure mode include: the concentration of CH4 in transformer oil / the concentration of H2 in transformer oil, the concentration of C2H4 in transformer oil / the concentration of C2H2 in transformer oil, the concentration of C2H4 in transformer oil / the concentration of C2H6 in transformer oil, the concentration of C2H2 in transformer oil / the concentration of (C1+C2) in transformer oil, the concentration of H2 in transformer oil / the concentration of (H2+C1+C2) in transformer oil, the concentration of C2H4 in transformer oil / the concentration of (C1+C2), the concentration of CH4 in transformer oil / the concentration of (C1+C2) in transformer oil, the concentration of C2H6 in transformer oil / the concentration of (C1+C2) in transformer oil and the concentration of (CH4+C2H4) in transformer oil / the concentration of (C1+C2) in transformer oil, wherein C1 represents CH4 and C2 represents the sum of C2H4, C2H6 and C2H2.

4. A transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The method for preprocessing the collected dissolved gas data includes: denoising and standardizing the collected dissolved gas data, filling missing values ​​in the data by the mean interpolation method, and smoothing the dissolved gas data.

5. A transformer oil chromatographic data analysis system based on deep belief network according to claim 4, characterized in that: The method for denoising the collected dissolved gas data specifically includes: using a low-pass filter to remove high-frequency noise, and the processing formula is: Among them, y(t) is the historical dissolved gas data after denoising, x(t) is the historical dissolved gas data before denoising, h(k) is the filter coefficient, and K is the order of the filter; The method for standardizing the collected dissolved gas data specifically includes: Among them, x i is the historical dissolved gas data before normalization, x′ i is the standardized historical dissolved gas data, μ i is the mean concentration of gas i, σ i is the standard deviation of the concentration of gas i; The method of filling missing values ​​in the collected dissolved gas data by the mean interpolation method specifically includes: using the mean of a column of data to fill missing values ​​in a column of data, Among them, x missing is the missing historical dissolved gas data, N is the number of known historical dissolved gas data points, and x i is the value of a known historical dissolved gas data point; The method for smoothing the collected dissolved gas data specifically includes: using the moving average method, at each historical dissolved gas data point, calculating the average value of multiple historical dissolved gas data points around it: Among them, y(t) is the smoothed historical dissolved gas data value, the smoothing result at time t, x(tk) is the value of the data at time tk, k represents the offset at the time, and M represents the number of historical dissolved gas data points used when calculating the average value.

6. A transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The method for pre-training the deep belief network model includes: inputting the pre-processed dissolved gas data into the deep belief network model, pre-training the deep belief network model through unsupervised learning, and extracting potential features of the dissolved gas data: h = f(Wx+b) Among them, h is the output of the feature extraction result of the historical dissolved gas data, W is the classification layer weight matrix set according to the importance of gas concentration, x is the historical dissolved gas data, b is a preset constant bias term, the bias term is used to adjust the activation value of the hidden layer to ensure the nonlinear representation ability of the network, and f is the preset activation function. The activation function introduces nonlinear transformation, so that the deep belief network model can learn complex patterns and relationships; When pre-training the deep belief network model, the deep belief network model reconstructs the pre-processed dissolved gas data, using the mean square error to reduce the difference between the data distribution and the model-generated distribution: Among them, E reconstruction is the mean square error, x i is the preprocessed historical dissolved gas data, is the reconstructed historical dissolved gas data, and N is the number of historical dissolved gas data samples.

7. A transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The method of adjusting the network weights of the deep belief network model to obtain the trained deep belief network model includes: further training the pre-trained deep belief network model using a contrastive divergence algorithm: △W=∈(<Hx T > data -<Hx T > model ) Among them, ΔW is the updated value of the network weight, which represents the relationship between the gas concentration data and the hidden layer features, H is the transformer oil dissolved gas concentration input of the visible layer, and x T is the activation value of the hidden layer, ∈ is the learning rate, <Hx T > data is the expectation of data distribution, which represents the relationship between the visible layer and the hidden layer, reflecting the distribution characteristics of the real data; <Hx T > model The expectation of the model distribution represents the relationship between the data generated by the current model and the hidden layer, reflects the generation distribution characteristics of the model, and finally obtains the trained deep belief network model.

8. The transformer oil chromatographic data analysis system based on deep belief network according to claim 1, characterized in that: The method for extracting gas concentration features and classifying faults using a trained deep belief network model on preprocessed real-time dissolved gas data includes: the trained deep belief network model automatically extracts high-dimensional features from oil chromatography data, the high-dimensional features include gas concentration, change trend, and gas concentration non-coding ratio of the gas; the trained deep belief network model predicts the fault type according to the gas concentration, change trend, and gas concentration non-coding ratio; the fault classification formula is: in, is the predicted fault type, W is the classification layer weight matrix set according to the importance of gas concentration, h is the output of the feature extraction result of dissolved gas data, b is a preset constant bias term, and the bias term is used to adjust the activation value of the hidden layer to ensure the nonlinear representation ability of the network; and through regression analysis, the time of fault occurrence is predicted: in, The predicted time of failure occurrence.

9. A transformer oil chromatographic data analysis method based on deep belief network, characterized in that: It includes: Preprocessing the real-time collected data of dissolved gas in transformer oil and the historical data of dissolved gas in transformer oil to obtain the preprocessed historical data of dissolved gas and the preprocessed real-time data of dissolved gas; The preprocessed historical dissolved gas data is input into the deep belief network model, the deep belief network model is pre-trained and the network weights of the deep belief network model are adjusted to obtain a trained deep belief network model. The trained deep belief network model extracts gas concentration features from the preprocessed real-time dissolved gas data, and obtains a gas concentration non-coding ratio associated with the transformer fault mode based on the extracted gas concentration features, and classifies the transformer fault according to the gas concentration non-coding ratio associated with the transformer fault mode.

10. A computer program product, comprising a computer program / instruction, which implements the steps of the method according to claim 9 when executed by a processor.

Images