Multi-source fusion feature generation method and device for transformer state evaluation

Abstract

The invention discloses a multi-source fusion feature generation method and device for transformer state evaluation, and the method comprises the steps: obtaining state data from a plurality of data collection assemblies related to a transformer according to a preset obtaining period; performing dimensionality reduction on each piece of state data based on a constructed double-sparse dictionary model to obtain dimensionality-reduced data; performing weighted filling on the dimension-reduced data according to a plurality of filling modes to obtain filled data; and according to a reconstruction error condition between the filling data and the state data, performing data cleaning on the filling data, and generating a multi-source fusion feature, thereby performing effective data preprocessing on the transformer data in time, and effectively improving accuracy and reliability of transformer evaluation.

Description

Technical Field

[0001] The present invention relates to the field of feature processing technology, and in particular to a multi-source fusion feature generation method and device for transformer state assessment. Background Art

[0002] In today's power system, transformers are key equipment, and their operating status plays a vital role in the normal operation of the power system. However, the existing transformer evaluation system has some obvious shortcomings.

[0003] At present, the input of characteristic parameters of transformer condition assessment system generally only considers electrical test data and oil chromatography data. At the same time, in the data preprocessing stage, for missing data, generally only one data filling method is used. For data dimensionality reduction, rough sets and deep belief networks are generally used for data sparse dimensionality reduction. For data cleaning, autoencoders are used for cleaning.

[0004] The above-mentioned single data source selection has great limitations and cannot fully reflect the actual status of the transformer under the overall operating conditions. Only considering a single filling method cannot fill in from the perspective of the entire data level, which may lead to overfitting in training. The existing dimensionality reduction method does not consider the impact of time series historical data. The data cleaning using autoencoders requires high computing resources, and it is difficult to perform effective data preprocessing on the transformer data in a timely manner, which can easily lead to reduced accuracy and reliability of transformer evaluation. Summary of the invention

[0005] The present invention provides a multi-source fusion feature generation method and device for transformer state assessment, which solves the technical problems that in the existing transformer state assessment process, the selection of a single data source has great limitations and cannot comprehensively reflect the actual state of the transformer under the overall operating conditions. Only considering a single filling method cannot fill in from the perspective of the entire data level, which may lead to overfitting in training. The existing dimensionality reduction method does not consider the influence of time series historical data. The data cleaning using an autoencoder has high requirements for computing resources, and it is difficult to perform effective data preprocessing on the transformer data in a timely manner, which easily leads to reduced accuracy and reliability of transformer assessment.

[0006] The present invention provides a multi-source fusion feature generation method for transformer state assessment, comprising:

[0007] According to a preset acquisition cycle, status data is obtained from various data acquisition components associated with the transformer respectively;

[0008] Based on the constructed double sparse dictionary model, the dimension of each state data is reduced to obtain reduced dimension data;

[0009] Performing weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data;

[0010] According to the reconstruction error between the filling data and the state data, the filling data is cleaned to generate multi-source fusion features.

[0011] Optionally, the state data is obtained from various data acquisition components associated with the transformer according to a preset acquisition cycle, including:

[0012] Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals;

[0013] When the acquisition time meets the preset acquisition period, the sound signal, the light signal, the ultrasonic signal and the vibration signal are respectively acquired from each of the data acquisition components;

[0014] The sound signal, the light signal, the ultrasonic signal and the vibration signal are sorted according to the acquisition time, and state data in sequence form are constructed respectively.

[0015] Optionally, the dimension reduction of each state data based on the constructed double sparse dictionary model to obtain the reduced dimension data includes:

[0016] The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model.

[0017] Projecting each of the state data to the dual sparse dictionary model;

[0018] An orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain dimension-reduced data;

[0019] The double sparse dictionary model is:

[0020]

[0021] in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, is a sparse representation of state data, i.e., dimensionality-reduced data, Represents the number of non-zero elements in the sparse coefficient vector.

[0022] Optionally, performing weighted filling on the dimension reduction data according to a plurality of filling modes to obtain filled data includes:

[0023] Calculate simulated filling values ​​corresponding to the dimensionality reduction data respectively according to multiple filling methods;

[0024] Calculate the error coefficients corresponding to the various filling methods by respectively using the simulated filling values ​​and the known values ​​in the dimension reduction data;

[0025] Setting the filling weights of the various filling methods according to the error coefficients;

[0026] Perform weighted averaging on the filling values ​​of each filling method according to the filling weight to obtain an actual filling value;

[0027] The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain filled data.

[0028] Optionally, performing data cleaning on the filling data according to a reconstruction error between the filling data and the state data to generate a multi-source fusion feature includes:

[0029] Calculating the loss error between the filling data and the status data in the acquisition cycle;

[0030] If the loss errors do not exceed the preset loss threshold, the filling data is determined as a lossless feature;

[0031] If the loss error exceeds the preset loss threshold, the error duration is counted;

[0032] If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature;

[0033] If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is called to reconstruct the padding data to obtain a reconstructed feature;

[0034] The lossless feature, the fluctuation feature and / or the reconstruction feature are used to generate a multi-source fusion feature.

[0035] The present invention also provides a multi-source fusion feature generation device for transformer state assessment, comprising:

[0036] A status data acquisition module, used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle;

[0037] A data dimension reduction module, used for reducing the dimension of each state data based on the constructed double sparse dictionary model to obtain reduced dimension data;

[0038] A data filling module, used for performing weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data;

[0039] The data cleaning module is used to clean the filling data according to the reconstruction error between the filling data and the state data to generate multi-source fusion features.

[0040] Optionally, the status data acquisition module is specifically used to:

[0041] Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals;

[0042] When the acquisition time meets the preset acquisition period, the sound signal, the light signal, the ultrasonic signal and the vibration signal are respectively acquired from each of the data acquisition components;

[0043] The sound signal, the light signal, the ultrasonic signal and the vibration signal are sorted according to the acquisition time, and state data in sequence form are constructed respectively.

[0044] Optionally, the data dimension reduction module is specifically used to:

[0045] The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model.

[0046] Projecting each of the state data to the dual sparse dictionary model;

[0047] An orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain dimension-reduced data;

[0048] The double sparse dictionary model is:

[0049]

[0050] in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, is a sparse representation of state data, i.e., dimensionality-reduced data, Represents the number of non-zero elements in the sparse coefficient vector.

[0051] Optionally, the data filling module is specifically used for:

[0052] Calculate simulated filling values ​​corresponding to the dimensionality reduction data respectively according to multiple filling methods;

[0053] Calculate the error coefficients corresponding to the various filling methods by respectively using the simulated filling values ​​and the known values ​​in the dimension reduction data;

[0054] Setting the filling weights of the various filling methods according to the error coefficients;

[0055] Perform weighted averaging on the filling values ​​of each filling method according to the filling weight to obtain an actual filling value;

[0056] The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain filled data.

[0057] Optionally, the data cleaning module is specifically used to:

[0058] Calculating the loss error between the filling data and the status data in the acquisition cycle;

[0059] If the loss errors do not exceed the preset loss threshold, the filling data is determined as a lossless feature;

[0060] If the loss error exceeds the preset loss threshold, the error duration is counted;

[0061] If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature;

[0062] If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is called to reconstruct the padding data to obtain a reconstructed feature;

[0063] The lossless feature, the fluctuation feature and / or the reconstruction feature are used to generate a multi-source fusion feature.

[0064] It can be seen from the above technical solutions that the present invention has the following advantages:

[0065] The present invention obtains state data from various data acquisition components associated with the transformer according to a preset acquisition cycle; performs dimensionality reduction on each of the state data based on a constructed double sparse dictionary model to obtain reduced dimensionality data; performs weighted filling on the reduced dimensionality data according to multiple filling methods to obtain filled data; performs data cleaning on the filled data according to a reconstruction error between the filled data and the state data, generates multi-source fusion features, thereby timely performing effective data preprocessing on the transformer data, and effectively improving the accuracy and reliability of transformer evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0067] Figure 1 A flowchart of the steps of a multi-source fusion feature generation method for transformer state assessment provided by an embodiment of the present invention;

[0068] Figure 2 A structural block diagram of a multi-source fusion feature generation device for transformer state assessment provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The embodiment of the present invention provides a multi-source fusion feature generation method and device for transformer state assessment, which is used to solve the technical problems that in the existing transformer state assessment process, the selection of a single data source has great limitations and cannot comprehensively reflect the actual state of the transformer under the overall operating conditions. Only considering a single filling method cannot fill in from the perspective of the entire data level, which may lead to overfitting in training. The existing dimensionality reduction method does not consider the influence of time series historical data. The data cleaning using an autoencoder has high requirements for computing resources, and it is difficult to perform effective data preprocessing on the transformer data in a timely manner, which easily leads to reduced accuracy and reliability of transformer assessment.

[0070] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0071] See also Figure 1 , Figure 1 A flowchart of the steps of a multi-source fusion feature generation method for transformer state assessment provided in an embodiment of the present invention.

[0072] The present invention provides a multi-source fusion feature generation method for transformer state assessment, comprising:

[0073] Step 101, acquiring status data from various data acquisition components associated with the transformer according to a preset acquisition cycle;

[0074] The data acquisition component refers to a modular component for collecting various signal data, which can be powered independently or by the processor power supply of this embodiment, and this embodiment does not limit this. Its types may include but are not limited to sound signal acquisition components such as microphones and sound cards, optical signal acquisition components such as photodiodes and image sensors, ultrasonic signal acquisition components such as ultrasonic transducers and signal conditioning circuits, and vibration signal acquisition components such as acceleration sensors and data acquisition cards.

[0075] The status data refers to a variety of serialized data used to evaluate the operating status of the transformer, including but not limited to serialized data of sound signals, light signals, ultrasonic signals and vibration signals.

[0076] In this embodiment, the method can be applied to a central processing unit with communication and data processing functions. In order to obtain status data in the form of a time series, various characteristic data of the transformer can be collected in real time through a data acquisition component. This processor obtains status data from multiple data acquisition components associated with the transformer as a data basis according to a preset acquisition cycle.

[0077] In an example of the present application, step 101 may include the following sub-steps:

[0078] Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals;

[0079] When the acquisition time meets the preset acquisition period, the sound signal, light signal, ultrasonic signal and vibration signal are respectively acquired from each data acquisition component;

[0080] The sound signal, light signal, ultrasonic signal and vibration signal are sorted according to the acquisition time, and the state data in sequence form are constructed respectively.

[0081] Most existing transformer status assessments only consider electrical test data and oil chromatography data. In this embodiment, in order to achieve real-time monitoring and multi-source fusion of the transformer, multiple data acquisition components associated with the transformer can be synchronously called when the transformer is started to collect multiple feature data of the transformer in real time. When the acquisition time meets the preset acquisition cycle, or when it is necessary to generate multi-source fusion features, such as when an external request or trigger signal is received, the processor obtains sound signals, light signals, ultrasonic signals and vibration signals from each data acquisition component as the data basis. Since the above signals are collected in chronological order, the various signals can be sorted according to the acquisition time of the signal to obtain status data in sequence.

[0082] It should be noted that the normal working noise of power equipment such as transformers is low-frequency audible sound, while the energy of discharge sound signals is mostly concentrated in the high-frequency band. The normal working sound of the transformer has a relatively uniform energy distribution in the range of 0-6kHz, but the main energy is also concentrated below 1kHz, while the high-energy part of the spark discharge sound is distributed around 10kHz, and the sound characteristics are significantly different. The light signal generated during corona discharge has both continuous and linear spectra, and the wavelength is mainly distributed between 280-400nm, belonging to the ultraviolet band, and a small part of the wavelength is distributed between 230-280nm. The energy of the corona current can be reflected by the size of the ultraviolet radiation power, so the size of the ultraviolet radiation power can be used as a basis for evaluating the strength of local discharge. When local discharge occurs, since the bubbles are charged, they will be squeezed by the electric field force, causing changes in the local volume of the bubbles. The change in bubble volume will form sparse and dense waves around it. This wave is ultrasonic wave. The process of local discharge generating ultrasonic wave is actually the process of electric-force-sound conversion. The amplitude of core vibration acceleration is proportional to the square of voltage, and its vibration frequency is twice the voltage frequency. Because the voltage frequency is 50Hz, the acceleration of core vibration is 100Hz. The deformation of core will increase its vibration intensity, which will make the amplitude, frequency and trend of vibration signal different from the core vibration signal during normal operation. Operators can extract characteristic values ​​from the collected vibration signal and distinguish it from normal vibration signal, so as to diagnose transformer core vibration.

[0083] Step 102, reducing the dimension of each state data based on the constructed double sparse dictionary model to obtain reduced dimension data;

[0084] The double sparse dictionary model combines a fixed dictionary with a learning dictionary. This new parameter structure will bring a concise and flexible dictionary representation, effectively reduce the dictionary training time, and will be more adaptable and effective. In this embodiment, the double sparse dictionary model uses a sparse base dictionary as a fixed dictionary and an overcomplete dictionary as a learning dictionary. A sparse base dictionary is a dictionary with a specific structure and properties, and its purpose is to be able to represent a given signal with as few non-zero coefficients as possible. It usually has a certain mathematical structure, such as orthogonality, tight frame, etc. For example, discrete cosine transform (DCT) dictionary, discrete wavelet transform (DWT) dictionary, etc. are all common sparse base dictionaries. An overcomplete dictionary means that the number of elements in the dictionary is much larger than the dimension of the signal, that is, the dictionary is overcomplete.

[0085] In this embodiment, after obtaining the state data in the form of a sequence, since the amount of data collected in real time is large, in order to improve the efficiency of subsequent data processing, it is necessary to reduce the dimension of the state data. Specifically, after constructing a double sparse dictionary model, the double sparse dictionary model is used to reduce the dimension of each state data, thereby obtaining a sparse representation of each state under the double sparse dictionary model as the reduced dimension data.

[0086] In an example of the present application, step 102 may include the following sub-steps:

[0087] The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model.

[0088] Project each state data to the double sparse dictionary model;

[0089] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain the reduced dimension data;

[0090] The double sparse dictionary model is:

[0091]

[0092] in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, It is a sparse representation of state data, that is, dimensionality-reduced data. Represents the number of non-zero elements in the sparse coefficient vector.

[0093] In an embodiment of the present application, a fixed dictionary and a learning dictionary can be used simultaneously in the signal representation, and the final sparse representation is obtained by linear combination or weighted summation, and then the weight distribution strategy of the fixed dictionary and the learning dictionary in the superposition is determined, which can be adjusted according to prior knowledge or through experiments. Joint optimization is performed to adjust the parameters of the fixed dictionary and the learning dictionary at the same time, so that the dual sparse dictionary model achieves the best performance on a specific task. The method of alternating optimization can be used to optimize the fixed dictionary and the learning dictionary respectively, and then alternate until convergence. During the optimization process, consider using appropriate regularization techniques to prevent overfitting. After completing the construction of the dual sparse dictionary model, by projecting each state data onto the dual sparse dictionary model, the state data is first subjected to feature extraction through a fixed dictionary to extract general prior features, and then adaptively extracted according to the specific type of the above-mentioned signal through the learning dictionary model to obtain more targeted features.

[0094] In a specific implementation, after projecting each data to the double sparse dictionary model respectively, the orthogonal matching pursuit algorithm can be used to solve the sparse representation of each state data under the double sparse dictionary model to obtain the reduced-dimensional data.

[0095] It should be noted that the Orthogonal Matching Pursuit (OMP) algorithm is used to find the sparse representation of the signal under an overcomplete dictionary. The basic idea is to select an atom from the dictionary that best matches the current residual each time, then project the signal onto the atom, update the residual and repeat the process until a certain stopping condition is met.

[0096] In this embodiment, the data dimensionality reduction in data preprocessing is different from the existing rough set and deep belief network. The simple data dimensionality reduction that only considers cannot consider the time series historical data and has poor processing effect on a large amount of data. A double sparse dictionary is used to perform feature sparse dimensionality reduction on a large amount of online time series data of the transformer.

[0097] Step 103, performing weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data;

[0098] In this embodiment, in order to account for the differences in signal types in different dimensionality reduction data, after the dimensionality reduction of the data is completed, multiple missing items appear in the data due to its sparse representation. In order to maximize the accuracy of subsequent data processing, the error coefficients of various filling methods can be weighted to combine the advantages of different filling methods to perform weighted filling on the dimensionality reduction data, thereby obtaining filled data.

[0099] In an example of the present application, step 103 may include the following sub-steps:

[0100] Calculate simulated filling values ​​corresponding to the dimension reduction data according to multiple filling methods;

[0101] The error coefficients corresponding to various filling methods are calculated by using each simulated filling value and the known values ​​in the dimension reduction data;

[0102] Set the filling weights of various filling methods according to each error coefficient;

[0103] According to the filling weight, the filling values ​​of each filling method are weighted averaged to obtain the actual filling value;

[0104] The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain the filled data.

[0105] In an embodiment of the present application, the simulated filling values ​​corresponding to the dimension reduction data are calculated respectively according to a variety of filling methods, and the error coefficients corresponding to the various filling methods are calculated by using the simulated filling values ​​and the known values ​​in the dimension reduction data. Among them, the error coefficient can be calculated in a variety of ways, such as calculating the mean absolute error, mean square error, etc. between the filling value and the true value. After the error coefficient is calculated, the filling weights of various filling methods are set according to the size of the error coefficient. The smaller the error coefficient, the greater the weight, and the sum of the weights is 1. The filling weight can be assigned by calculating the ratio of the error coefficient and taking the inverse.

[0106] After completing the filling weight setting, the filling values ​​of each filling method are weighted averaged according to the filling weight to calculate the actual filling value at each position. Then, the missing values ​​in the dimensionality reduction data are filled according to the actual filling value to generate the filled data.

[0107] It should be noted that the filling methods may include but are not limited to KNN filling, mean filling, zero value filling, median filling, forward filling and backward filling.

[0108] Step 104, according to the reconstruction error between the filling data and the state data, the filling data is cleaned to generate multi-source fusion features.

[0109] In this embodiment, in order to reduce the resource consumption of data cleaning, after generating the filling data, the loss error and error duration between the filling data and the state data are calculated, so as to obtain the reconstruction error between the filling data and the state data. The filling data is cleaned according to the reconstruction error to generate multi-source fusion features.

[0110] In an example of the present application, step 104 may include the following sub-steps:

[0111] Calculate the loss error between filling data and status data during the acquisition cycle;

[0112] If the loss errors do not exceed the preset loss threshold, the filled data is determined as lossless features;

[0113] If the loss error exceeds the preset loss threshold, the error duration is counted;

[0114] If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature;

[0115] If the error duration exceeds the preset duration threshold, the preset stacked denoising autoencoder model is called to reconstruct the padded data to obtain the reconstructed features;

[0116] Multi-source fusion features are generated by using lossless features, fluctuation features and / or reconstruction features.

[0117] The stacked denoising autoencoder model SDAE is composed of multiple denoising autoencoders (DAEs) stacked together. Each DAE contains an encoder and a decoder. The encoder maps the input data to a low-dimensional latent representation, while the decoder attempts to reconstruct the original input data from this latent representation. By stacking multiple DAEs, more advanced feature representations can be gradually learned.

[0118] In this embodiment, the loss error between the filling data and the status data during the acquisition cycle can be calculated by the following formula:

[0119]

[0120] in, is the loss error of filling data and status data during the acquisition cycle, represents a loss function based on a specific function H, whose input is state data x and filling data z, is the entropy function to measure uncertainty, To normalize the variable x, Same reason.

[0121] In the embodiment of the present application, the loss threshold is Thd, and the reconstruction error of the state data Re≤Thd is called lossless data; if Re>Thd and Et≤Tw of the partial data, it indicates that these monitoring parameters are affected at the same time, and the transformer abnormality or failure cannot be ruled out; if Re>Thd and Et>Tw of the partial data, the state data is the state quantity of the normal operation of the transformer, and the "dirty" data is a missing value; if Re>Thd and Et>Tw of the partial data, the state data is the abnormal operation state data of the transformer. For the abnormal data defined in the above steps, the reconstruction representation of the "dirty" number by the SDAE model is used as the repair value, and the lossless data and the repair values ​​of the singular points and missing values ​​constitute the valid information of the normal state data.

[0122] In the specific implementation, since the current data cleaning uses an autoencoder, which results in a large occupation of training resources and poor training effect, in an embodiment of the present application, the filling data whose error duration exceeds the preset duration threshold can be input into the trained SDAE model. The SDAE model processes the input data according to its trained structure and parameters. First, the underlying denoising autoencoder (DAE) encodes the input data and maps it to a low-dimensional potential representation space. In this process, the model tries to remove noise and abnormal information from the data. Then, through the decoder, the data is gradually reconstructed from the low-dimensional potential representation. Since the SDAE model learns the characteristic patterns of normal data during the training process, it will reconstruct a data representation close to the normal state as much as possible based on the characteristics of the input "dirty" data, thereby obtaining reconstructed data.

[0123] It should be noted that the error duration is Et, and Tw is the maximum time period or duration allowed for the error to exist. Within this time range, even if the system shows some error or deviation, it is considered acceptable and no immediate action is required.

[0124] In an embodiment of the present application, state data is obtained from various data acquisition components associated with the transformer according to a preset acquisition cycle; each state data is reduced in dimension based on a constructed double sparse dictionary model to obtain reduced-dimensional data; the reduced-dimensional data is weightedly filled in accordance with a plurality of filling methods to obtain filled data; the filled data is cleaned according to the reconstruction error between the filled data and the state data to generate multi-source fusion features, thereby timely and effectively preprocessing the transformer data, and effectively improving the accuracy and reliability of transformer evaluation. For the construction of characteristic parameters of the transformer state evaluation system, a construction of multi-source data characteristic parameters based on the fusion of non-electrical data acquisition is proposed. For the missing data, the noise of the data, and the sparse dimensionality reduction of the data in the data preprocessing, a weighted preprocessing method integrating multiple filling methods, a data cleaning method of a stacked denoising autoencoder, and a multi-source data signal compression sensing method of a double sparse dictionary model are proposed respectively, and a multi-system fusion method for overall data preprocessing is realized.

[0125] See also Figure 2 , Figure 2 A structural block diagram of a multi-source fusion feature generation device for transformer state assessment in an embodiment of the present application is shown.

[0126] The embodiment of the present invention provides a multi-source fusion feature generation device for transformer state assessment, comprising:

[0127] The status data acquisition module 201 is used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle;

[0128] A data dimension reduction module 202 is used to reduce the dimension of each state data based on the constructed double sparse dictionary model to obtain reduced dimension data;

[0129] The data filling module 203 is used to perform weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data;

[0130] The data cleaning module 204 is used to clean the filling data according to the reconstruction error between the filling data and the state data to generate multi-source fusion features.

[0131] Optionally, the status data acquisition module 201 is specifically used for:

[0132] Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals;

[0133] When the acquisition time meets the preset acquisition period, the sound signal, light signal, ultrasonic signal and vibration signal are respectively acquired from each data acquisition component;

[0134] The sound signal, light signal, ultrasonic signal and vibration signal are sorted according to the acquisition time, and the state data in sequence form are constructed respectively.

[0135] Optionally, the data dimension reduction module 202 is specifically used for:

[0136] The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model.

[0137] Project each state data to the double sparse dictionary model;

[0138] The orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain the reduced dimension data;

[0139] The double sparse dictionary model is:

[0140]

[0141] in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, is a sparse representation of state data, i.e., dimensionality-reduced data, Represents the number of non-zero elements in the sparse coefficient vector.

[0142] Optionally, the data filling module 203 is specifically used for:

[0143] Calculate simulated filling values ​​corresponding to the dimension reduction data according to multiple filling methods;

[0144] The error coefficients corresponding to various filling methods are calculated by using each simulated filling value and the known values ​​in the dimension reduction data;

[0145] Set the filling weights of various filling methods according to each error coefficient;

[0146] According to the filling weight, the filling values ​​of each filling method are weighted averaged to obtain the actual filling value;

[0147] The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain the filled data.

[0148] Optionally, the data cleaning module 204 is specifically used for:

[0149] Calculate the loss error between filling data and status data during the acquisition cycle;

[0150] If the loss errors do not exceed the preset loss threshold, the filled data is determined as lossless features;

[0151] If the loss error exceeds the preset loss threshold, the error duration is counted;

[0152] If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature;

[0153] If the error duration exceeds the preset duration threshold, the preset stacked denoising autoencoder model is called to reconstruct the padded data to obtain the reconstructed features;

[0154] Multi-source fusion features are generated by using lossless features, fluctuation features and / or reconstruction features.

[0155] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0156] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0157] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-source fusion feature generation method for transformer condition assessment, characterized in that: include: According to a preset acquisition cycle, status data is obtained from various data acquisition components associated with the transformer respectively; Based on the constructed double sparse dictionary model, the dimension of each state data is reduced to obtain reduced dimension data; Performing weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data; According to the reconstruction error between the filling data and the state data, the filling data is cleaned to generate multi-source fusion features.

2. The method according to claim 1, characterized in that According to the preset acquisition cycle, the status data is respectively acquired from various data acquisition components associated with the transformer, including: Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals; When the acquisition time meets the preset acquisition period, the sound signal, the light signal, the ultrasonic signal and the vibration signal are respectively acquired from each of the data acquisition components; The sound signal, the light signal, the ultrasonic signal and the vibration signal are sorted according to the acquisition time, and state data in sequence form are constructed respectively.

3. The method according to claim 1, characterized in that The dual sparse dictionary model constructed based on the state data is used to reduce the dimension of each state data to obtain reduced dimension data, including: The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model. Projecting each of the state data to the dual sparse dictionary model; An orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain dimension-reduced data; The double sparse dictionary model is: in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, is a sparse representation of state data, i.e., dimensionality-reduced data, Represents the number of non-zero elements in the sparse coefficient vector.

4. The method according to claim 1, characterized in that: The weighted filling of the dimension reduction data according to a plurality of filling modes to obtain the filled data includes: Calculate simulated filling values ​​corresponding to the dimensionality reduction data respectively according to multiple filling methods; Calculate the error coefficients corresponding to the various filling methods by respectively using the simulated filling values ​​and the known values ​​in the dimension reduction data; Setting the filling weights of the various filling methods according to the error coefficients; Perform weighted averaging on the filling values ​​of each filling method according to the filling weight to obtain an actual filling value; The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain filled data.

5. The method according to claim 1, characterized in that The step of performing data cleaning on the filling data according to the reconstruction error between the filling data and the state data to generate multi-source fusion features includes: Calculating the loss error between the filling data and the status data in the acquisition cycle; If the loss errors do not exceed the preset loss threshold, the filling data is determined as a lossless feature; If the loss error exceeds the preset loss threshold, the error duration is counted; If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature; If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is called to reconstruct the padding data to obtain a reconstructed feature; The lossless feature, the fluctuation feature and / or the reconstruction feature are used to generate a multi-source fusion feature.

6. A multi-source fusion feature generation device for transformer state assessment, characterized in that: include: A status data acquisition module, used to acquire status data from various data acquisition components associated with the transformer according to a preset acquisition cycle; A data dimension reduction module, used for reducing the dimension of each state data based on the constructed double sparse dictionary model to obtain reduced dimension data; A data filling module, used for performing weighted filling on the dimension reduction data according to a plurality of filling methods to obtain filled data; The data cleaning module is used to clean the filling data according to the reconstruction error between the filling data and the state data to generate multi-source fusion features.

7. The device according to claim 6, characterized in that The status data acquisition module is specifically used for: Calling various data acquisition components associated with the transformer to collect characteristic data in real time; wherein the characteristic data includes sound signals, light signals, ultrasonic signals and vibration signals; When the acquisition time meets the preset acquisition period, the sound signal, the light signal, the ultrasonic signal and the vibration signal are respectively acquired from each of the data acquisition components; The sound signal, the light signal, the ultrasonic signal and the vibration signal are sorted according to the acquisition time, and state data in sequence form are constructed respectively.

8. The device according to claim 6, characterized in that The data dimension reduction module is specifically used for: The fixed dictionary model is used as the base layer to superimpose the learning dictionary model to construct a dual sparse dictionary model. Projecting each of the state data to the dual sparse dictionary model; An orthogonal matching pursuit algorithm is used to solve the sparse representation of each state data under the double sparse dictionary model to obtain dimension-reduced data; The double sparse dictionary model is: in, is the coefficient matrix Middle The coefficient column vector corresponding to the dictionary elements, is the state data in the learning dictionary model The sparse representation coefficient matrix under It is a learning dictionary model The coefficient matrix after sparse representation, is a fixed dictionary model predefined for each dictionary element. is the i-th sample in the state data, is a sparse coefficient vector, is a sparse representation of state data, i.e., dimensionality-reduced data, Represents the number of non-zero elements in the sparse coefficient vector.

9. The device according to claim 6, characterized in that The data filling module is specifically used for: Calculate simulated filling values ​​corresponding to the dimensionality reduction data respectively according to multiple filling methods; Calculate the error coefficients corresponding to the various filling methods by respectively using the simulated filling values ​​and the known values ​​in the dimension reduction data; Setting the filling weights of the various filling methods according to the error coefficients; Perform weighted averaging on the filling values ​​of each filling method according to the filling weight to obtain an actual filling value; The missing values ​​of the dimensionality reduction data are filled according to the actual filling values ​​to obtain filled data.

10. The device according to claim 6, characterized in that The data cleaning module is specifically used for: Calculating the loss error between the filling data and the status data in the acquisition cycle; If the loss errors do not exceed the preset loss threshold, the filling data is determined as a lossless feature; If the loss error exceeds the preset loss threshold, the error duration is counted; If the error duration does not exceed the preset duration threshold, the filling data is determined as a fluctuation feature; If the error duration exceeds a preset duration threshold, a preset stacked denoising autoencoder model is called to reconstruct the padding data to obtain a reconstructed feature; The lossless feature, the fluctuation feature and / or the reconstruction feature are used to generate a multi-source fusion feature.

Images