Method and system for detecting looseness of transformer winding

Abstract

The invention provides a transformer winding looseness detection method and system, and the method comprises the steps: installing a plurality of acceleration sensors at each monitoring point position of a target transformer, so as to monitor an initial vibration signal of the target transformer in real time; de-noising the initial vibration signal collected by each monitoring point position by using a connection degree model, and fusing the de-noised information to obtain a pre-processing signal of the target transformer; performing normalization processing and time division on the preprocessed signal to obtain a sub-vibration time domain waveform signal; obtaining the maximum value of each sub-vibration time-domain waveform signal, and combining the maximum values to obtain a peak sequence signal; extracting characteristic parameters of the peak sequence signal by using a time-frequency analysis method, and inputting the characteristic parameters into the trained winding looseness identification model to obtain an identification result; and the structure of the winding looseness identification model is a CNN-LSTM network. The method realizes intelligent identification of winding looseness, can adapt to a complex vibration mode, and improves the identification efficiency and accuracy.

Description

Technical Field

[0001] The invention relates to the field of power equipment detection, and in particular to a transformer winding loosening detection method and system. Background Art

[0002] Transformers are key equipment in power systems, and the mechanical stability of their windings is crucial to the safe operation of transformers. Loose windings may cause local overheating, insulation damage, and even serious failures. Currently, commonly used detection methods include vibration analysis and frequency response analysis.

[0003] The invention patent with application number 202210641491.2 discloses a transformer winding loose fault diagnosis method, including step 1: collecting vibration signals on the surface of the transformer box; step 2: calculating the time domain eigenvalues ​​and frequency domain eigenvalues ​​of each vibration signal; step 3: constructing a feature matrix; step 4: normalization processing; step 5: constructing a GOA-KELM recognition model; step 6: fault classification and identification. Its vibration signal may be interfered by environmental noise and the noise of the equipment itself, affecting the effectiveness of feature extraction, and the effect of the GOA-KELM model may be limited by training data. Models trained under specific conditions may fail in new environments. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a transformer winding loose detection method and system, which utilizes the combination of CNN-LSTM network, fully utilizes the advantages of convolutional neural network (CNN) in feature extraction and the ability of long short-term memory network (LSTM) in time series data analysis, and realizes intelligent recognition of winding looseness. This intelligent detection scheme can adapt to complex vibration patterns and improve recognition efficiency and accuracy.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for detecting loose transformer windings, comprising:

[0007] A plurality of acceleration sensors are installed at each monitoring point of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time;

[0008] The initial vibration signal collected at each monitoring point is denoised using the connection degree model, and the denoised information is fused to obtain the preprocessed signal of the target transformer;

[0009] Performing normalization processing and time division on the preprocessed signal to obtain a sub-vibration time domain waveform signal;

[0010] Obtain the maximum value of each sub-vibration time domain waveform signal and combine them to obtain a peak sequence signal;

[0011] The characteristic parameters of the peak sequence signal are extracted by using a time-frequency analysis method, and the characteristic parameters are input into a trained winding looseness identification model to obtain an identification result; the structure of the winding looseness identification model is a CNN-LSTM network.

[0012] Preferably, the initial vibration signal collected at each monitoring point is denoised using a connection degree model, and the denoised information is fused to obtain a preprocessed signal of the target transformer, including:

[0013] A connection degree model is constructed according to the spatial distance between each acceleration sensor;

[0014] Constructing a vibration denoising function using the connection degree model;

[0015] Using the vibration denoising function to remove abnormal values ​​of vibration information at each monitoring point to generate denoised vibration information;

[0016] The fusion function is constructed using the denoised vibration information of each monitoring point;

[0017] Constructing a fusion matrix according to the fusion function;

[0018] Calculate the eigenvalues ​​corresponding to the fusion matrix;

[0019] The denoised vibration information of each monitoring point is fused using the eigenvalue to obtain a preprocessed signal of the target transformer.

[0020] Preferably, the calculation formula of the connection degree model is:

[0021]

[0022] Wherein, σ represents the average value of the vibration difference between the acceleration sensor at point m and the acceleration sensor at point n in the historical time period, dis(m,n) represents the spatial distance between the acceleration sensor at point m and the acceleration sensor at point n, and R represents the preset parameter.

[0023] Preferably, the vibration denoising function is constructed using the connection degree model, including:

[0024] The denoising threshold is constructed using the weighted value of the vibration value of each monitoring point; wherein the denoising threshold is: Among them, Y i represents the i-th vibration value collected by the acceleration sensor, W i Represents Y i The weighted value of

[0025] A vibration denoising function is constructed based on the denoising threshold and the connection degree model; the calculation formula of the vibration denoising function is: in, represents the denoised vibration information at time t, x(m,t) represents the measurement value of the acceleration sensor at point m at time t, x(n,t) represents the measurement value of the acceleration sensor at point n at time t, ρ represents the adjustable threshold, |N m | represents the number of acceleration sensors at the same monitoring point,

[0026] Preferably, the calculation formula of the fusion function is:

[0027]

[0028] Among them, q ij represents the fusion function, X i represents the vibration information of the ith monitoring point after denoising, X j Represents the vibration information of the jth monitoring point after denoising.

[0029] Preferably, the fusion matrix is:

[0030]

[0031] Preferably, the denoised vibration information of each monitoring point is fused using the characteristic value to obtain the preprocessed signal of the target transformer, including:

[0032] Using formula Get the fusion weighted coefficient; where L represents the denoised wavelet coefficient, r i represents the i-th eigenvalue of the fusion matrix;

[0033] The fusion weighted coefficient is used to perform weighted averaging on the denoised vibration values ​​of all monitoring points to obtain the preprocessed signal of the target transformer.

[0034] Preferably, the characteristic parameters include frequency and amplitude.

[0035] Preferably, the method for constructing the winding looseness identification model includes:

[0036] Obtaining a preset frequency sample data set and an amplitude sample data set;

[0037] Construct a convolutional neural network for multi-object segmentation;

[0038] The convolutional neural network is trained according to the frequency sample data set, the amplitude sample data set and the knee joint ligament data set to obtain a trained first classifier and a second classifier;

[0039] Using the first classifier to identify loose results in the frequency sample data set to obtain a frequency confidence set, and using the second classifier to identify loose results in the amplitude sample data set to obtain an amplitude confidence set;

[0040] Summing the confidence levels of the loosening results in the bone confidence set and the cartilage confidence set, respectively obtaining the loosening confidence sums of the frequency confidence set and the amplitude confidence set;

[0041] Select the data with the largest sum of loose confidence as the training threshold data, and retrain the classifiers corresponding to the other sums of loose confidence so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network;

[0042] The trained LSTM neural network is connected after the classification network to obtain the winding looseness identification model.

[0043] A transformer winding loosening detection system, comprising:

[0044] A sensor installation module is used to install a plurality of acceleration sensors at various monitoring points of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time;

[0045] A preprocessing module, used to denoise the initial vibration signal collected at each monitoring point using a connection degree model, and fuse the denoised information to obtain a preprocessing signal of the target transformer;

[0046] A signal acquisition module, used for normalizing and time-dividing the pre-processed signal to obtain a sub-vibration time-domain waveform signal;

[0047] A sequence acquisition module is used to obtain the maximum value of each sub-vibration time domain waveform signal and combine them to obtain a peak sequence signal;

[0048] The identification module is used to extract the characteristic parameters of the peak sequence signal by using a time-frequency analysis method, and input the characteristic parameters into a trained winding loose identification model to obtain an identification result; the structure of the winding loose identification model is a CNN-LSTM network.

[0049] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0050] The present invention provides a transformer winding loose detection method and system, the method comprising: installing a plurality of acceleration sensors at each monitoring point of a target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time; using a connection degree model to denoise the initial vibration signal collected at each monitoring point, and fusing the denoised information to obtain a preprocessing signal of the target transformer; normalizing and time-dividing the preprocessing signal to obtain a sub-vibration time domain waveform signal; obtaining the maximum value of each sub-vibration time domain waveform signal, and combining to obtain a peak sequence signal; extracting characteristic parameters of the peak sequence signal using a time-frequency analysis method, and inputting the characteristic parameters into a trained winding loose identification model to obtain an identification result; the structure of the winding loose identification model is a CNN-LSTM network. The present invention uses a connection degree model to denoise the initial vibration signal, which can effectively remove the noise generated in the environment and during operation and improve the signal quality. Fusing the denoised information helps to form a more reliable preprocessing signal and enhance the effectiveness of subsequent analysis. The present invention utilizes the combination of CNN-LSTM networks, fully utilizing the advantages of convolutional neural networks (CNN) in feature extraction and the capabilities of long short-term memory networks (LSTM) in time series data analysis to achieve intelligent recognition of winding looseness, adapt to complex vibration modes, and improve recognition efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] 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 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.

[0052] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of the system structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments 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.

[0055] The purpose of the present invention is to provide a transformer winding loose detection method and system, which utilizes the combination of CNN-LSTM network, fully utilizes the advantages of convolutional neural network (CNN) in feature extraction and the ability of long short-term memory network (LSTM) in time series data analysis, and realizes intelligent recognition of winding looseness. This intelligent detection scheme can adapt to complex vibration modes and improve recognition efficiency and accuracy.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a method for detecting transformer winding looseness, comprising:

[0058] Step 100: installing a plurality of acceleration sensors at various monitoring points of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time;

[0059] Step 200: De-noising the initial vibration signal collected at each monitoring point using a connection degree model, and fusing the de-noised information to obtain a pre-processed signal of the target transformer;

[0060] Step 300: normalizing and time-dividing the preprocessed signal to obtain a sub-vibration time domain waveform signal;

[0061] Step 400: obtaining the maximum value of each sub-vibration time domain waveform signal, and combining them to obtain a peak sequence signal;

[0062] Step 500: extract characteristic parameters of the peak sequence signal using a time-frequency analysis method, and input the characteristic parameters into a trained winding looseness identification model to obtain an identification result; the structure of the winding looseness identification model is a CNN-LSTM network.

[0063] Specifically, this embodiment first requires the reasonable selection and arrangement of monitoring points. The structure of the transformer is complex, and the vibration characteristics may be different in different parts. Therefore, this embodiment requires the installation of sensors at key parts of the transformer, such as the top, bottom, and sides of the winding. These locations are usually the areas where the vibration signals are most obvious, and can effectively capture the vibration changes caused by loose windings. During installation, it is necessary to ensure that the sensor is in close contact with the surface of the transformer to improve the accuracy and stability of signal acquisition.

[0064] Furthermore, selecting a suitable acceleration sensor is the key to achieving efficient monitoring in this embodiment. The sensor should have high sensitivity and wide-band characteristics to capture small vibration changes and signals within a wide frequency range. In addition, the sensor needs to have good anti-interference capabilities to reduce the impact of external environmental noise on the signal. During installation, the fixing method of the sensor also needs to be considered to avoid signal distortion caused by loose or unstable installation. Bolt fixing or strong glue bonding is usually used to ensure long-term stable signal acquisition.

[0065] Optionally, in this embodiment, multiple sensors can be set at the same monitoring site. Multiple sensors can collect vibration signals at the same location at the same time. By comparing and comprehensively analyzing these signals, the impact of errors or failures that may occur in a single sensor can be effectively reduced. This redundant design of this embodiment can improve the accuracy and reliability of the data and ensure the credibility of the monitoring results. Different sensors may have different sensitivities and frequency response characteristics. By using them in combination, a wider frequency range and vibration amplitude can be covered. This multi-sensor configuration can capture more subtle vibration changes and improve the ability to resolve abnormal conditions such as loose windings.

[0066] Furthermore, in a complex power equipment environment, vibration signals are susceptible to various noises and interferences. The setting of multiple sensors can effectively filter out environmental noise and accidental interference through signal cross-validation and fusion processing, and improve the signal-to-noise ratio. This embodiment can further enhance the anti-interference ability of the system and ensure the stability and continuity of monitoring.

[0067] In summary, this embodiment can achieve more comprehensive and accurate vibration monitoring by setting multiple sensors at one monitoring position, and provide a solid data basis for early detection of transformer winding loosening.

[0068] Preferably, the initial vibration signal collected at each monitoring point is denoised using a connection degree model, and the denoised information is fused to obtain a preprocessed signal of the target transformer, including:

[0069] A connection degree model is constructed according to the spatial distance between each acceleration sensor;

[0070] Constructing a vibration denoising function using the connection degree model;

[0071] Using the vibration denoising function to remove abnormal values ​​of vibration information at each monitoring point to generate denoised vibration information;

[0072] The fusion function is constructed using the denoised vibration information of each monitoring point;

[0073] Constructing a fusion matrix according to the fusion function;

[0074] Calculate the eigenvalues ​​corresponding to the fusion matrix;

[0075] The denoised vibration information of each monitoring point is fused using the eigenvalue to obtain a preprocessed signal of the target transformer.

[0076] Preferably, the calculation formula of the connection degree model is:

[0077]

[0078] Wherein, σ represents the average value of the vibration difference between the acceleration sensor at point m and the acceleration sensor at point n in the historical time period, dis(m,n) represents the spatial distance between the acceleration sensor at point m and the acceleration sensor at point n, and R represents the preset parameter.

[0079] In the present invention, the sensors at point m and point n are neighboring sensors. The closer the positions of the two sensors are, the greater the correlation between the two sensors. Therefore, the present invention measures the degree of connection between the two sensors based on the spatial distance between them. In practical applications, since the sensor may be affected by its own parameters or environmental factors, the measured value collected by the sensor at a certain moment may deviate greatly from the actual value. Therefore, the present invention introduces R to correct the problem of sensor correlation deviation caused by the influence of its own parameters or environmental factors. In general, the value of R is 1.

[0080] Preferably, the vibration denoising function is constructed using the connection degree model, including:

[0081] The denoising threshold is constructed using the weighted value of the vibration value of each monitoring point; wherein the denoising threshold is: Among them, Y i represents the i-th vibration value collected by the acceleration sensor, W i Represents Y i The weighted value of

[0082] A vibration denoising function is constructed based on the denoising threshold and the connection degree model; the calculation formula of the vibration denoising function is: in, represents the denoised vibration information at time t, x(m,t) represents the measurement value of the acceleration sensor at point m at time t, x(n,t) represents the measurement value of the acceleration sensor at point n at time t, ρ represents the adjustable threshold, |N m | represents the number of acceleration sensors at the same monitoring point,

[0083] Specifically, this embodiment uses a vibration denoising function to fuse the measurement values ​​of sensors at the same monitoring point, thereby avoiding a large deviation between the measurement value collected by the sensor at a certain moment and the actual value, making the measurement value of the sensor more realistic.

[0084] Preferably, the calculation formula of the fusion function is:

[0085]

[0086] Among them, q ij represents the fusion function, X i represents the vibration information of the ith monitoring point after denoising, X j Represents the vibration information of the jth monitoring point after denoising.

[0087] Preferably, the fusion matrix is:

[0088]

[0089] Preferably, the denoised vibration information of each monitoring point is fused using the characteristic value to obtain the preprocessed signal of the target transformer, including:

[0090] Using formula Get the fusion weighted coefficient; where L represents the denoised wavelet coefficient, r i represents the i-th eigenvalue of the fusion matrix;

[0091] The fusion weighted coefficient is used to perform weighted averaging on the denoised vibration values ​​of all monitoring points to obtain the preprocessed signal of the target transformer.

[0092] Specifically, this embodiment is based on the fusion matrix and dynamically allocates the corresponding fusion weighting coefficient L to each sensor according to the temperature information measured by each sensor in real time, so as to obtain the vibration information after the final fusion of all detection points, which can be closer to the actual vibration value of the entire transformer.

[0093] Furthermore, when the preprocessed signal is normalized and time-divided in this embodiment, the preprocessed signal must first be amplitude normalized. The purpose of normalization is to adjust the amplitude of the signal to a uniform range, which is usually achieved by subtracting the mean and dividing by the standard deviation. This process helps to eliminate amplitude differences between different sensors or under different measurement conditions, making subsequent analysis more consistent and reliable. Next, the normalized signal is divided into several time windows of fixed length by time, each time window representing a sub-vibration time domain waveform signal. This time division can capture the characteristic changes of the signal in different time periods, providing a basis for subsequent feature extraction.

[0094] Furthermore, the present embodiment obtains the maximum value of each sub-vibration time domain waveform signal in order to extract the most significant vibration feature in each time window. For each sub-waveform signal, its maximum value in the time window is calculated, and these maximum values ​​represent the peak vibration amplitude in the time period. The maximum values ​​of all time windows are combined together to form a peak sequence signal. This sequence signal reflects the changing trend of the vibration signal during the entire monitoring period and is an important characteristic parameter for identifying winding looseness. By analyzing this peak sequence, abnormal vibration patterns can be identified, thereby determining whether there are problems such as winding looseness.

[0095] Optionally, the steps of normalization processing in this embodiment are as follows:

[0096] The preprocessed signal is normalized by the following formula:

[0097]

[0098] Wherein, amplify_vib_arr is the preprocessed signal, i=0, 1, 2, 3, ... n-1, amplify_vib_arri is the i-th preprocessed signal, n is the number of preprocessed signals; amplify_vib_arr is the normalized preprocessed signal;

[0099] The normalized preprocessed signal is rounded to one decimal place to obtain normalized data.

[0100] Specifically, this embodiment divides the normalized data according to preset time intervals. The determination of the preset time intervals needs to comprehensively consider factors such as the operating characteristics of the transformer, the frequency characteristics of the vibration signal, and the resolution of the monitoring system. Usually, the time interval selected should be small enough to capture the dynamic changes and detailed characteristics of the vibration signal, but it should not be too small to avoid generating too much data processing burden. Specifically, this embodiment can determine the main frequency components and change periods of the signal by performing spectral analysis on the vibration signal of the transformer under normal and abnormal conditions. Based on these frequency characteristics, a time interval that can cover the main frequencies and does not distort the signal characteristics is selected. In addition, the sampling rate of the monitoring system must also be considered to ensure that the selected time interval can match the sampling capacity of the system, thereby achieving efficient and accurate signal division.

[0101] Preferably, the characteristic parameters include frequency and amplitude.

[0102] Specifically, this embodiment uses a time-frequency analysis method to extract characteristic parameters of a peak sequence signal. First, it is necessary to select a suitable time-frequency analysis tool. Commonly used time-frequency analysis methods include short-time Fourier transform (STFT), wavelet transform, and Hilbert-Huang transform. These methods can provide both time and frequency information of the signal and are suitable for analyzing non-stationary signals. When selecting a specific method, it is necessary to consider the characteristics of the signal and the target of the analysis. For example, the wavelet transform is suitable for processing signals with mutation characteristics, while the STFT is suitable for analyzing signals with better stability.

[0103] When performing time-frequency analysis, the peak sequence signal is first segmented so that spectrum analysis can be performed in each time period. Through time-frequency transformation, the one-dimensional time series signal is converted into a two-dimensional time-frequency diagram to show the frequency distribution of the signal at different time points. Next, characteristic parameters such as instantaneous frequency, energy distribution, bandwidth, etc. are extracted from the time-frequency diagram. These parameters can reflect the frequency change and energy concentration of the signal, and are an important basis for identifying winding looseness.

[0104] Finally, the extracted characteristic parameters are input into the trained recognition model for pattern recognition and classification. By comparing the characteristic parameters under normal and abnormal conditions, the characteristic pattern of winding looseness can be identified. In order to improve the accuracy of recognition, the present embodiment fully trains and verifies the model to ensure that it can effectively distinguish different vibration states. Through the combination of this time-frequency analysis and feature extraction, accurate monitoring and diagnosis of the transformer winding state can be achieved.

[0105] Preferably, the method for constructing the winding looseness identification model includes:

[0106] Obtaining a preset frequency sample data set and an amplitude sample data set;

[0107] Construct a convolutional neural network for multi-object segmentation;

[0108] The convolutional neural network is trained according to the frequency sample data set, the amplitude sample data set and the knee joint ligament data set to obtain a trained first classifier and a second classifier;

[0109] Using the first classifier to identify loose results in the frequency sample data set to obtain a frequency confidence set, and using the second classifier to identify loose results in the amplitude sample data set to obtain an amplitude confidence set;

[0110] Summing the confidence levels of the loosening results in the bone confidence set and the cartilage confidence set, respectively obtaining the loosening confidence sums of the frequency confidence set and the amplitude confidence set;

[0111] Select the data with the largest sum of loose confidence as the training threshold data, and retrain the classifiers corresponding to the other sums of loose confidence so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network;

[0112] The trained LSTM neural network is connected after the classification network to obtain the winding looseness identification model.

[0113] Specifically, the present embodiment constructs a winding loosening identification model by obtaining two types of sample data sets: a frequency sample data set and an amplitude sample data set. These data sets contain the vibration signal characteristics of the transformer under different operating conditions, specifically reflecting the frequency and amplitude changes under the influence of winding loosening. The accuracy and comprehensiveness of the data set directly affect the effect of subsequent model training, so it is necessary to ensure the quality of the data set acquisition and annotation process. Next, the present embodiment constructs a convolutional neural network (CNN) for multi-target segmentation. This network is designed to analyze data of different feature dimensions, so it includes multiple parallel sub-networks (sub-classifiers), which are trained for frequency and amplitude data respectively. When the frequency and amplitude sample data sets are trained independently, it is necessary to consider the distribution characteristics of different features, adjust the network architecture and hyperparameters, so as to ensure that the training process can effectively identify the change pattern under the loose state. During the training process, the two classifiers of the CNN are trained using the preset sample data sets respectively to obtain the first classifier (for frequency signals) and the second classifier (for amplitude signals). Each classifier will output the loosening identification result of the corresponding sample data set and generate its confidence set. The confidence set reflects the judgment strength of the classifier and is an important basis for subsequent decision fusion. This strategy ensures that the classifier objectively evaluates the contribution of various signal features. In the evaluation stage, the sum of the confidences of each loosening result in the frequency and amplitude confidence set is calculated, and the result with the largest sum of confidences is selected as the training threshold data. This threshold data is used to reversely adjust the classifier so that it can better produce high-confidence judgments in data with similar features. Through this optimization step, the recognition ability and stability of each sub-classifier are improved, and the outputs of different classifiers are more coordinated. Finally, a comprehensive classification network is constructed by cascading the trained first and second classifiers, and a trained LSTM neural network is connected to it. The LSTM network can capture the time dependency in the sequence signal and help identify the dynamically changing characteristic parameters. Combining the above steps, the construction of the entire winding loosening identification model is completed. This model can accurately and efficiently analyze the characteristics of transformer vibration signals and provide an intelligent solution for winding loosening detection.

[0114] Furthermore, in the winding loosening identification model, LSTM (Long Short-Term Memory Network) is used to process and analyze time series data to capture the dynamic changes and time dependencies of the signal. LSTM network is particularly suitable for processing sequence data because it can remember information over a long time span and effectively solve the gradient vanishing problem common in traditional RNN (recurrent neural network). The specific steps of applying LSTM network in this embodiment are as follows:

[0115] 1. Input data preparation: After being processed by the convolutional neural network (CNN), the feature vector sequence obtained is used as the input of the LSTM network. These feature vectors contain the time-frequency characteristics of the frequency and amplitude signals, which are high-level representations of the original vibration signals. The input data needs to be arranged in chronological order so that the LSTM can learn the temporal dependency of the signal.

[0116] 2. LSTM network structure design. When designing the structure of the LSTM network, it is necessary to determine the number of layers of the network, the number of units in each layer (i.e. the dimension of the hidden layer), and other hyperparameters (such as learning rate, regularization parameters, etc.). Usually, the complexity of the network matches the scale and complexity of the dataset to avoid overfitting or underfitting. The number of layers and units of the LSTM can be adjusted through experiments to find the best configuration.

[0117] 3. During the training process, the LSTM network updates its weights through the back-propagation algorithm to minimize the prediction error. The training data set includes vibration signals under various operating conditions to ensure the generalization ability of the model. Through its memory unit and gating mechanism, LSTM can effectively capture the long-term dependencies of the signal, thereby improving the accuracy of identifying winding looseness.

[0118] 4. Output results. The output of the LSTM network is usually a probability distribution, which indicates the possibility that the input signal belongs to different states (such as normal or loose). By setting a threshold, this embodiment converts the probability distribution into a specific classification result. The output results of the LSTM are combined with the feature extraction results of the CNN to form a complete recognition model.

[0119] 5. Model evaluation and optimization. After the model is built, it needs to be evaluated and optimized. Use independent validation and test sets to evaluate the performance of the model, and adjust the network structure and hyperparameters to improve accuracy and robustness. Through continuous iteration and optimization, ensure that the LSTM network can effectively identify the loose state of transformer windings in practical applications.

[0120] Through the above steps, the LSTM network in this embodiment plays a key role in the winding looseness identification model, helping to capture and analyze the temporal dynamic characteristics of the vibration signal, thereby achieving more accurate fault detection and diagnosis.

[0121] Corresponding to the above method, such as Figure 2 As shown, this embodiment also provides a transformer winding looseness detection system, including:

[0122] A sensor installation module is used to install a plurality of acceleration sensors at various monitoring points of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time;

[0123] A preprocessing module, used to denoise the initial vibration signal collected at each monitoring point using a connection degree model, and fuse the denoised information to obtain a preprocessing signal of the target transformer;

[0124] A signal acquisition module, used for normalizing and time-dividing the pre-processed signal to obtain a sub-vibration time-domain waveform signal;

[0125] A sequence acquisition module is used to obtain the maximum value of each sub-vibration time domain waveform signal and combine them to obtain a peak sequence signal;

[0126] The identification module is used to extract the characteristic parameters of the peak sequence signal by using a time-frequency analysis method, and input the characteristic parameters into a trained winding loose identification model to obtain an identification result; the structure of the winding loose identification model is a CNN-LSTM network.

[0127] The beneficial effects of the present invention are as follows:

[0128] (1) The present invention installs acceleration sensors at multiple monitoring points to monitor the vibration signals of the transformer in real time, and can more comprehensively capture early signals of winding loosening. The multi-point monitoring of the present invention greatly improves the accuracy of detection and helps to discover potential problems in a timely manner.

[0129] (2) The present invention uses the connection degree model to denoise the initial vibration signal, which can effectively remove the noise generated by the environment and the operation process and improve the signal quality. The denoised information is integrated to help form a more reliable preprocessing signal and enhance the effectiveness of subsequent analysis.

[0130] (3) The present invention obtains the sub-vibration time domain waveform signal through normalization processing and time division, so that the vibration characteristics of each time period are fully analyzed. The present invention can capture subtle changes and provide more possibilities for subsequent feature extraction.

[0131] (4) The time-frequency analysis method of the present invention extracts characteristic parameters of the peak sequence signal, which can reflect the frequency characteristics and amplitude changes of the vibration signal. These parameters of the present invention are important bases for identifying winding looseness and can improve the accuracy of the identification model.

[0132] (5) The present invention utilizes the combination of CNN-LSTM network, making full use of the advantages of convolutional neural network (CNN) in feature extraction and the ability of long short-term memory network (LSTM) in time series data analysis to realize intelligent recognition of winding looseness. This intelligent detection scheme can adapt to complex vibration patterns and improve recognition efficiency and accuracy.

[0133] (6) The present invention realizes automated detection, reduces reliance on human intervention, helps improve work efficiency and reduces the probability of human error.

[0134] (7) The present invention combines real-time monitoring and intelligent identification technology to achieve dynamic monitoring of the transformer winding status and issue a warning in time when signs of looseness are found, thereby ensuring the safe operation of the transformer.

[0135] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0136] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for detecting loose transformer windings, characterized in that: include: A plurality of acceleration sensors are installed at each monitoring point of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time; The initial vibration signal collected at each monitoring point is denoised using the connection degree model, and the denoised information is fused to obtain the preprocessed signal of the target transformer; Performing normalization processing and time division on the preprocessed signal to obtain a sub-vibration time domain waveform signal; Obtain the maximum value of each sub-vibration time domain waveform signal and combine them to obtain a peak sequence signal; The characteristic parameters of the peak sequence signal are extracted by using a time-frequency analysis method, and the characteristic parameters are input into a trained winding looseness identification model to obtain an identification result; the structure of the winding looseness identification model is a CNN-LSTM network.

2. The transformer winding loosening detection method according to claim 1, characterized in that: The initial vibration signal collected at each monitoring point is denoised using the connection degree model, and the denoised information is fused to obtain the preprocessed signal of the target transformer, including: A connection degree model is constructed according to the spatial distance between each acceleration sensor; Constructing a vibration denoising function using the connection degree model; Using the vibration denoising function to remove abnormal values ​​of vibration information at each monitoring point to generate denoised vibration information; The fusion function is constructed using the denoised vibration information of each monitoring point; Constructing a fusion matrix according to the fusion function; Calculate the eigenvalues ​​corresponding to the fusion matrix; The denoised vibration information of each monitoring point is fused using the eigenvalue to obtain a preprocessed signal of the target transformer.

3. The transformer winding loosening detection method according to claim 2, characterized in that: The calculation formula of the connection degree model is: Wherein, σ represents the average value of the vibration difference between the acceleration sensor at point m and the acceleration sensor at point n in the historical time period, dis(m,n) represents the spatial distance between the acceleration sensor at point m and the acceleration sensor at point n, and R represents the preset parameter.

4. The transformer winding loosening detection method according to claim 3, characterized in that: The vibration denoising function is constructed using the connection degree model, including: The denoising threshold is constructed using the weighted value of the vibration value of each monitoring point; wherein the denoising threshold is: Among them, Y i represents the i-th vibration value collected by the acceleration sensor, W i Represents Y i The weighted value of A vibration denoising function is constructed based on the denoising threshold and the connection degree model; the calculation formula of the vibration denoising function is: in, represents the denoised vibration information at time t, x(m,t) represents the measurement value of the acceleration sensor at point m at time t, x(n,t) represents the measurement value of the acceleration sensor at point n at time t, ρ represents the adjustable threshold, N m Indicates the number of acceleration sensors at the same monitoring point.

5. The transformer winding loosening detection method according to claim 4, characterized in that: The calculation formula of the fusion function is: Among them, q ij represents the fusion function, X i represents the vibration information of the ith monitoring point after denoising, X j Represents the vibration information of the jth monitoring point after denoising.

6. The transformer winding loosening detection method according to claim 5, characterized in that: The fusion matrix is:

7. The transformer winding loosening detection method according to claim 6, characterized in that: The denoised vibration information of each monitoring point is fused using the eigenvalue to obtain the preprocessed signal of the target transformer, including: Using formula Get the fusion weighted coefficient; where L represents the denoised wavelet coefficient, r i represents the i-th eigenvalue of the fusion matrix; The fusion weighted coefficient is used to perform weighted averaging on the denoised vibration values ​​of all monitoring points to obtain the preprocessed signal of the target transformer.

8. The transformer winding loosening detection method according to claim 1, characterized in that: The characteristic parameters include frequency and amplitude.

9. The transformer winding loosening detection method according to claim 2, characterized in that: The method for constructing the winding looseness identification model includes: Obtaining a preset frequency sample data set and an amplitude sample data set; Construct a convolutional neural network for multi-object segmentation; The convolutional neural network is trained according to the frequency sample data set, the amplitude sample data set and the knee joint ligament data set to obtain a trained first classifier and a second classifier; Using the first classifier to identify loose results in the frequency sample data set to obtain a frequency confidence set, and using the second classifier to identify loose results in the amplitude sample data set to obtain an amplitude confidence set; Summing the confidence levels of the loosening results in the bone confidence set and the cartilage confidence set, respectively obtaining the loosening confidence sums of the frequency confidence set and the amplitude confidence set; Select the data with the largest sum of loose confidence as the training threshold data, and retrain the classifiers corresponding to the other sums of loose confidence so that the classification confidence of each classifier is greater than or equal to the training threshold data, and finally cascade the trained classifiers to obtain a classification network; The trained LSTM neural network is connected after the classification network to obtain the winding looseness identification model.

10. A transformer winding loosening detection system, characterized in that: include: A sensor installation module is used to install a plurality of acceleration sensors at various monitoring points of the target transformer; the acceleration sensors are used to monitor the initial vibration signal of the target transformer in real time; A preprocessing module, used to denoise the initial vibration signal collected at each monitoring point using a connection degree model, and fuse the denoised information to obtain a preprocessing signal of the target transformer; A signal acquisition module, used for normalizing and time-dividing the pre-processed signal to obtain a sub-vibration time-domain waveform signal; A sequence acquisition module is used to obtain the maximum value of each sub-vibration time domain waveform signal and combine them to obtain a peak sequence signal; The identification module is used to extract the characteristic parameters of the peak sequence signal by using a time-frequency analysis method, and input the characteristic parameters into a trained winding loose identification model to obtain an identification result; the structure of the winding loose identification model is a CNN-LSTM network.

Images