A fault diagnosis method for on-load tap changer based on phase space reconstruction theory

Abstract

A method for fault diagnosis of on-load tap changers based on phase space reconstruction theory, in the technical field of fault diagnosis of on-load tap changers, is used to solve the problem that wavelet analysis and modal decomposition methods focus on time domain or frequency domain features, lack in-depth analysis of the overall spatial characteristics and complex structures of signals, and thus their recognition capabilities are limited when faced with complex fault modes. Aiming at the fault diagnosis and identification problems of KM type oil-immersed online transformer tap changers OLTC, the present invention adopts vibration signal analysis, feature extraction and machine learning technology, and proposes an effective fault classification method to improve the reliability and operation efficiency of the equipment. CatBoost is selected as the fault classification model, and the hyperparameters are systematically adjusted using Bayesian optimization to obtain the optimal hyperparameter configuration. This process significantly improves the classification accuracy of the model and enhances its robustness.

Description

Technical Field

[0001] The invention relates to the technical field of on-load tap changer fault diagnosis, in particular to an on-load tap changer fault diagnosis method based on phase space reconstruction theory. Background Art

[0002] Power transformers are indispensable core equipment in the power system and play an irreplaceable role. Their operating conditions have a great impact on the reliable operation of the entire power supply and distribution system. Among them, the on-load tap changer OLTC, as an important component of the transformer, plays an important role in stabilizing the voltage at the load center, increasing the flexibility of the power grid, and improving the quality of power. As the only frequently moving component of the transformer, OLTC works in a high voltage and high current environment for a long time, and is prone to component wear, ablation and other faults. According to the statistical analysis of fault types, OLTC faults account for 20% of the total faults of on-load tap-changing transformers, of which mechanical faults account for as high as 70%. OLTC under abnormal working conditions poses a huge safety hazard to the safe and stable operation of the transformer. Therefore, how to effectively monitor the operating status of OLTC and timely discover and prevent potential fault hazards is of great significance to ensuring the safe operation of the transformer and the reliable transmission of the power grid system.

[0003] During the gear shifting process of the mechanical on-load tap changer OLTC, the transmission mechanism transmits the power of the motor to the tap selector, the switching switch and other components and operates in sequence. The vibration signal generated during the gear shift contains rich equipment status information. At present, vibration analysis has become the research focus of mechanical OLTC fault diagnosis. OLTC vibration analysis usually includes three steps: selection of vibration burst data, feature extraction and fault diagnosis. Currently, the commonly used signal feature types include time domain features, frequency domain features and time-frequency domain features. By extracting signal features through wavelet analysis and modal decomposition, the overall status of the equipment can be better reflected, which can be used for fault diagnosis and trend prediction.

[0004] Wavelet analysis and modal decomposition can quickly extract vibration signal features, but they cannot effectively solve the impact of wavelet basis selection and modal aliasing on signal decomposition, resulting in insufficient signal feature extraction or increased errors, thus affecting the accuracy of fault identification. The above two methods usually focus on time domain or frequency domain features, lacking in-depth analysis of the overall spatial characteristics and complex structures of the signal, which limits their ability to identify complex fault modes. Vibration signals are often affected by complex mechanical structures, motion states, and external factors, and exhibit nonlinear and chaotic behaviors. Traditional linear analysis methods find it difficult to capture this complexity. Summary of the invention

[0005] The purpose of the present invention is to provide a method for fault diagnosis of on-load tap changer based on phase space reconstruction theory, which is used to solve the problem that wavelet analysis and modal decomposition method focus on time domain or frequency domain characteristics, lack in-depth analysis of the overall spatial characteristics and complex structure of the signal, and thus their recognition ability is limited when facing complex fault modes.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a method for diagnosing faults of on-load tap changers based on phase space reconstruction theory, comprising the following steps:

[0007] S1. Data collection.

[0008] S1.1 Install an acceleration sensor on the top of the OLTC to collect OLTC vibration signals.

[0009] S1.2 sets four working conditions to simulate common OLTC faults. The four working conditions include the normal state, and three mechanical fault conditions: loose arc plate and contact, loose transmission shaft screws, and stuck transmission gear.

[0010] S1.3 performs multiple tests in the same gear and multiple tests in multiple gears under the normal state of OLTC and three mechanical fault conditions, and collects 160 sets of data for each working condition.

[0011] S2. Data selection.

[0012] S2.1 is added based on the original signal sequence The modal components of the group white noise are constructed A noisy signal, .

[0013] In the formula, Indicates A noisy structural signal; is the original signal; represents signal-to-noise ratio; E represents EMD decomposition; Represents white noise.

[0014] S2.2 Use EMD to analyze noisy signals Repeat the decomposition, then subtract it from the noisy signal and calculate the mean to get the first residual signal and the corresponding modal components .

[0015] .

[0016] .

[0017] S2.3 Add multiple groups of white noise components and get And find the mean and then compare Difference modal components .

[0018] .

[0019] .

[0020] .

[0021] is the first residual signal The noisy signal after adding white noise again on the basis of represents the signal-to-noise ratio; Represents the second residual signal.

[0022] S2.4 Repeat step S2.3 until all residual signals can no longer be decomposed, and all modal components IMF are obtained; modal components IMF1-IMF4 are taken as processing objects for subsequent further feature extraction.

[0023] S3. Feature extraction.

[0024] S3.1 Phase space reconstruction.

[0025] For a given time series , N is the length of the time series, and the reconstructed phase space coordinates are expressed as: .

[0026] in, is the coordinate of the phase point in the reconstructed phase space; τ is the delay time; m is the embedding dimension.

[0027] S3.2 Phase space surface reconstruction.

[0028] The reconstructed surface is constructed using the cubic interpolation method of the griddata function in Python.

[0029] S3.3 Gray-level co-occurrence matrix feature extraction.

[0030] S3.3.1Construct the gray-level co-occurrence matrix.

[0031] Given an image I, its size is , the gray-level co-occurrence matrix is ​​defined as: (8).

[0032] in, is the gray level, ranging from , is the grayscale value; is the relative distance, is the relative direction, Represents the horizontal displacement of the pixel. Represents the vertical displacement of the pixel. and Depends on the relative distance and relative direction ; The contrast Con, energy E, homogeneity H and correlation Cor of the gray-level co-occurrence matrix are selected as image texture features for feature extraction.

[0033] 1) Contrast Con: measures the intensity of the grayscale difference of the image, and the calculation formula is: (9).

[0034] 2) Energy E: reflects the uniformity of the texture. The larger the energy, the more uniform the texture. The calculation formula is: (10).

[0035] 3) Homogeneity H: measures the similarity of adjacent grayscale values. The calculation formula is: (11).

[0036] 4) Correlation Cor: reflects the linear correlation of gray values ​​between pixels. The calculation formula is: (12).

[0037] in Is grayscale The mean of Is grayscale The mean of , is the standard deviation.

[0038] S3.3.2 Texture feature extraction of surface projection image from phase space reconstruction.

[0039] S3.3.2.1 Use global minimum and maximum values ​​for normalization. The global minimum value min(x) and maximum value max(x) of the data set are calculated to ensure that the data use the same normalization standard. The formula is as follows: (13).

[0040] In the formula, is the normalized data.

[0041] S3.3.2.2 Coarse quantization of gray levels.

[0042] Reduce the grayscale to a suitable value without affecting the texture characteristics of the image, and quantize the 256 grayscale levels to 48 levels.

[0043] S3.3.2.3 Draw a grayscale image.

[0044] Map the data to grayscale between 0 and 47 and draw a grayscale image of the surface projected onto the XY plane.

[0045] S3.3.2.4 Calculate the gray-level co-occurrence matrix.

[0046] The distance is taken as 1 pixel, and the gray-level co-occurrence matrix in the four directions of 0°, 45°, 90°, and 135° is calculated.

[0047] S3.3.2.5 Calculate the four eigenvalues ​​of contrast, energy, homogeneity and correlation for each gray-level co-occurrence matrix.

[0048] S4. Sample division.

[0049] For each group of original signals, modal components IMF1, IMF2, IMF3 and IMF4 are decomposed, and 16-dimensional features are extracted for each modal component IMF. Finally, a 64-dimensional feature vector is obtained for each group of signals to form a training set and a test set, which are then input into the classifier for training and testing.

[0050] S5. OLTC fault diagnosis and identification based on Bayesian optimization CatBoost.

[0051] The steps of CatBoost are:

[0052] S5.1 Initializing the model .

[0053] S5.2 For each iteration, calculate the residual of the current model , fitting a new weak learner based on the residual And update the model: (14).

[0054] in, To represent the model after the m-1th iteration, is the corresponding learning rate.

[0055] S5.3 Based on the early stopping mechanism or the specified number of iterations, through a series of weak learners The final model is constructed cumulatively: (15).

[0056] Bayesian hyperparameter automatic optimization is used to optimize the five hyperparameters in the CatBoost model, including iterations, tree depth, learning rate, L2 regularization parameter 12 leaf reg, and bagging_temperature, to obtain the optimal parameters of the CatBoost model.

[0057] Furthermore, the simulation methods of three mechanical failure conditions are as follows:

[0058] Fault 1: The looseness of the arc plate is simulated by loosening a screw ring of the static contact on the A-phase arc plate of the OLTC; the looseness of the contact is simulated by loosening the neutral point contact and the related screws of the main on-off contact.

[0059] Fault 2: Loose screws on the transmission shaft are simulated by loosening the screws at the connection of the transmission rod.

[0060] Fault 3: The transmission gear is stuck. This is done by stuffing wood chips into the gear box of the switch transmission mechanism to increase the friction between the gears, thereby simulating the failure of the transmission shaft being stuck.

[0061] Furthermore, the mutual information method is used to determine the delay time τ: Assuming that the original vibration signal , after a time delay of τ, the signal becomes , then the mutual information value between sequences S and Q is obtained by the following information entropy and probability: (16).

[0062] Where: is the information entropy of signal S; is the information entropy of signal Q; is the joint information entropy of S and Q; The median value of the sampling point is The probability of the point appearing in the entire data segment; The median value of the sampling point is The probability of the point appearing in the entire data segment; The sampling value of the original signal is And the delayed signal sampling value is probability.

[0063] Furthermore, the pseudo nearest neighbor method is used to determine the optimal embedding dimension m.

[0064] When the embedding dimension m is small, the trajectory in the phase space is not fully unfolded, resulting in folding and entanglement of the trajectory, which in turn makes the phase points that were originally far apart become closer. These close phase points are called pseudo neighbor points. In order to separate the pseudo neighbor points, it is necessary to gradually increase the embedding dimension m. When the embedding dimension m increases to the point where the pseudo neighbor points are basically reduced to zero, the corresponding embedding dimension m is the optimal embedding dimension m of the system.

[0065] Furthermore, for the vibration signals under the four working conditions, the delay time τ and the embedding dimension m of the modal components IMF1, IMF2, IMF3 and IMF4 after ICEEMDAN decomposition are calculated respectively; to ensure the consistency of variables, all vibration signals are reconstructed in the same phase space, and the delay time τ and the embedding dimension m calculated by the modal component IMF are averaged to finally determine the reconstruction parameters τ and m.

[0066] Furthermore, the specific content of step S3.2 is:

[0067] S3.2.1 performs Delaunay triangulation on the reconstructed point set, ensuring that the circumscribed circle of each triangle does not contain other points, thereby constructing a triangular mesh covering all points.

[0068] S3.2.2 Within each triangle, the data is locally interpolated using a cubic polynomial, and within each triangular region, the values ​​between points are smoothly estimated using the cubic polynomial to generate a smooth and continuous surface, thereby reflecting the local changes in the reconstructed points.

[0069] S3.2.3 The discrete reconstruction points are connected into a smooth surface through cubic interpolation, which is the reconstruction surface. The surface reflects the information of all reconstruction points. When the mechanical state of OLTC changes, the surface generated by interpolation will also change accordingly.

[0070] The beneficial effects of the present invention are as follows: the present invention adopts vibration signal analysis, feature extraction and machine learning technology to solve the fault diagnosis and identification problems of KM type oil-immersed online transformer tap changer OLTC, and proposes an effective fault classification method, aiming to improve the reliability and operation efficiency of the equipment. (1) Through detailed experimental design of vibration signal collection of OLTC under different working conditions, common fault conditions are simulated, including loose arc plate and contact, loose transmission shaft screw and transmission gear jamming, so as to ensure a comprehensive analysis of various fault states of OLTC. (2) The collected vibration signal is processed by ICEEMDAN decomposition technology to extract the intrinsic mode function IMF containing rich fault features. Through phase space reconstruction, a high-dimensional feature space is constructed, which provides reliable support for fault classification. In addition, the gray level co-occurrence matrix GLCM is used to process the XY plane projection of the phase space reconstruction surface, and multi-dimensional image features such as contrast, energy, homogeneity and correlation are extracted, further enriching the feature description. Phase space reconstruction can expand one-dimensional time series data into multi-dimensional phase space through delayed embedding, so that the dynamic behavior of the signal, such as periodicity and chaos, can be more clearly revealed, which is helpful for analyzing the dynamic characteristics of the equipment and lays a solid theoretical foundation for nonlinear time series analysis. Due to the high sensitivity of phase space reconstruction to initial conditions, it can amplify the slight changes in the vibration signal of the equipment and realize accurate identification of faults. (3) Bayesian optimization CatBoost is selected as the fault classification model, and the hyperparameters are systematically adjusted using Bayesian optimization CatBoost to obtain the optimal hyperparameter configuration. This process significantly improves the classification accuracy of the model and enhances its robustness. After training and testing, the classification accuracy of the model on the test set reached 95%, and the classification accuracy of 93% was obtained through 5-fold cross validation. This shows that the proposed model has good performance and reliability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of the sensor layout.

[0072] Figure 2 This is a diagram of the arc plate and loose contacts.

[0073] Figure 3 This is a diagram of a loose drive shaft.

[0074] Figure 4 This is a diagram of the transmission gear getting stuck.

[0075] Figure 5 This is the original signal waveform of the whole process of vibration when the OLTC switches from the 7th gear to the 8th gear under normal conditions.

[0076] Figure 6This is the original signal waveform of the whole process of vibration when the OLTC switches from the 7th gear to the 8th gear when the arc plate and the contact are loose.

[0077] Figure 7 This is the original signal waveform of the whole process of vibration when the OLTC switches from the 7th gear to the 8th gear when the transmission shaft is loose.

[0078] Figure 8 This is the original signal waveform of the whole process of vibration when the OLTC switches from the 7th gear to the 8th gear when the transmission gear is stuck.

[0079] Fig. 9 This is the vibration signal diagram of the switch in the normal state of OLTC.

[0080] Fig.10 This is the vibration signal diagram of the switch when the OLTC arc plate and contacts are loose.

[0081] Fig.11 This is the vibration signal diagram of the switch when the OLTC transmission shaft is loose.

[0082] Fig.12 This is the vibration signal diagram of the switch when the OLTC transmission gear is stuck.

[0083] Fig.13 This is the ICEEMDAN decomposition result of the normal state switching switch vibration signal.

[0084] Fig.14 This is the grayscale image of the IMF2 phase space reconstruction surface projection under normal conditions.

[0085] Fig.15 This is the grayscale image of the surface projection of IMF2 phase space reconstruction when the arc plate and the contact are loose.

[0086] Fig.16 This is the grayscale image of the surface projection of IMF2 phase space reconstruction when the transmission shaft is loose.

[0087] Fig.17 This is the grayscale image of the surface projection reconstructed in the IMF2 phase space when the transmission gear is stuck.

[0088] Fig.18 Confusion matrix plot for the Bayesian optimization CatBoost model.

[0089] Fig.19 This is a flow chart of OLTC fault diagnosis of the present invention. DETAILED DESCRIPTION

[0090] The present invention discloses a fault diagnosis method for on-load tap changer based on phase space reconstruction theory. First, the vibration signal under different working conditions is decomposed by using improved adaptive noise complete set empirical mode decomposition ICEEMDAN. Then, the phase space reconstruction surface of the inherent mode component obtained by decomposition is obtained by using the phase space reconstruction method. The image texture features of the two-dimensional projection of the XY plane of the phase space reconstruction surface are extracted by using the gray level co-occurrence matrix GLCM. Finally, the feature database is used to train the Bayesian optimization CatBoost classification model to realize the accurate diagnosis and identification of typical fault defects of OLTC. Fig.19 As shown, the specific steps include:

[0091] S1. Data collection.

[0092] The present invention uses KM oil-immersed OLTC as the experimental object. KM oil-immersed OLTC is a typical combined on-load tap changer, which consists of a switch and a tap selector. KM oil-immersed tap changer is driven by KDJ-III electric mechanism to achieve the transformation of transformer tapping to achieve the purpose of voltage regulation.

[0093] S1.1 As Figure 1 As shown, the present invention selects three YD-38D voltage output acceleration sensors with a sensitivity of 5mV / g and installs them on the top of the OLTC at an even interval of 120° along the circumferential direction to collect the vibration signal of the OLTC, and the signal sampling rate is set to 25KHz.

[0094] S1.2 The present invention sets four working conditions to simulate common faults of OLTC. The four working conditions include a normal state, and three mechanical fault conditions, namely, a loose condition of an arc plate and a contact, a loose condition of a transmission shaft screw, and a stuck condition of a transmission gear. The loose condition of an arc plate and a contact is named Fault 1, the loose condition of a transmission shaft screw is named Fault 2, and the stuck condition of a transmission gear is named Fault 3.

[0095] Fault 1: If Figure 2 As shown in the figure, the loosening of the arc plate is simulated by loosening a screw ring of the static contact on the arc plate of phase A of the OLTC. Different degrees of looseness faults can be achieved by loosening the number of turns of the fixing screws of the corresponding components; contact loosening is simulated by loosening the relevant screws of the neutral point contact and the main on-off contact.

[0096] Fault 2: If Figure 3 As shown, the loosening of the transmission shaft screws is achieved by loosening the screws at the connection of the transmission rod to realize fault simulation.

[0097] Fault 3: If Figure 4 As shown, the transmission gear jam is achieved by stuffing wood chips into the gear box of the switch transmission mechanism to increase the friction between the gears, thereby simulating the jam of the transmission shaft.

[0098] S1.3 performs multiple tests in the same gear and multiple tests in multiple gears under the normal state of OLTC and the set mechanical fault conditions (fault 1, fault 2 and fault 3), and collects 160 sets of data for each working condition. Among them, the original signal waveform of the whole process of vibration when the OLTC is switched from the 7th gear to the 8th gear under a normal state is as follows: Figure 5 As shown in Figure 1, the entire waveform records the complete OLTC shift vibration waveform of motor startup, transmission shaft rotation, switch action, and motor stop. The original vibration signal waveforms of the three mechanical fault conditions when the 7th gear is switched to the 8th gear are shown in Figure 1. Figure 6 , Figure 7 and Figure 8 shown.

[0099] Different mechanical fault conditions have different effects on the vibration waveform of the OLTC switching process, which is mainly reflected in the amplitude of the impact vibration. Since the data of the switching switch action stage has the highest amplitude and the most information content in the complete signal, the present invention selects the vibration signal of this stage as the main research object. The selected signals are all vibration signals of the same sensor under different working conditions and the same gear change. Fig. 9 , Fig.10 , Fig.11 and Fig.12 As shown in the figure, the vibration signal during OLTC switching is composed of multiple pulse components, showing nonlinear and strong time-varying characteristics, and the shape of the vibration signal varies with different working conditions. During the switching action, the transition contact a1, transition contact b1, main on-off contact b and main contact B collide in sequence, among which the action time of the main on-off contact b and the main contact B partially overlap, and the vibration waveform overlaps. Intuitively, due to the vibration or resonance of the OLTC mechanical parts caused by different fault types, the vibration signal generated by the entire equipment differs in frequency, amplitude and geometric characteristics. Although the vibration signal waveforms under different working conditions can be distinguished through manual experience, this method is often subjective and difficult to make accurate judgments. Therefore, a fault diagnosis and identification method based on signal processing and machine learning is introduced.

[0100] S2. Data selection.

[0101] ICEEMDAN is gradually evolved from the empirical mode decomposition (EMD) method. EMD completes the adaptive decomposition of the original sequence according to the time scale characteristics of the sequence itself. It has extremely high applicability and has obvious advantages in processing non-stationary and nonlinear signals. However, EMD decomposition has obvious mode aliasing. In order to suppress the mode aliasing phenomenon, the ensemble empirical mode decomposition (EEMD) adds white noise to the original signal to eliminate mode aliasing, but increases the reconstruction error. In response to the reconstruction error problem, the CEEMDAN algorithm adds positive and negative complementary white noise to each stage of mode decomposition, thereby eliminating the residual white noise interference and improving the reconstruction accuracy, but there is a problem of false components. Compared with EMD, ICEEMDAN overcomes the mode aliasing problem by adding white noise, making different modes more distinguishable. By adding different noises in each decomposition step and eliminating the noise after the decomposition is completed, the balance of noise and the stability of the decomposition results are ensured; at the same time, ICEEMDAN adjusts the noise amplitude to ensure that the extraction of each intrinsic mode function (IMF) is more robust, reducing the interference of noise on the signal.

[0102] The present invention adopts the vibration signal decomposition method of ICEEMDAN, and the specific steps are as follows:

[0103] S2.1 is added on the basis of the original signal sequence The modal components of the group white noise are constructed A noisy signal, .

[0104] In the formula, Indicates A noisy structural signal; is the original signal; represents the signal-to-noise ratio; represents EMD decomposition; Represents white noise.

[0105] S2.2 Use EMD to analyze noisy signals Repeat the decomposition, then subtract it from the noisy signal and calculate the mean to get the first residual signal and the corresponding modal components .

[0106] .

[0107] .

[0108] S2.3 Add multiple groups of white noise components and get And find the mean and then compare Make a difference .

[0109] .

[0110] .

[0111] .

[0112] is the first residual signal The noisy signal after adding white noise again on the basis of Represents the signal-to-noise ratio.

[0113] S2.4 Repeat step S2.3 until all residual signals can no longer be decomposed and all modal components are obtained.

[0114] Take a set of normal switch vibration signals as an example to perform ICEEMDAN decomposition. The decomposition results are as follows: Fig.13 As shown in the figure, each order of intrinsic mode component is arranged according to frequency after the vibration signal is decomposed, and contains components with different vibration information and time characteristic scales, which can contain the chaotic system information in the mechanical system. Fig.13 It can be seen that the high-frequency noise component of the intrinsic mode component increases from IMF5, the useful information contained gradually decreases, and the contribution to the signal is small. In order to effectively reduce the complexity of subsequent calculations while retaining the main signal features, the present invention takes IMF1-IMF4 as the processing object for subsequent further feature extraction.

[0115] S3 feature extraction.

[0116] S3.1 Phase space reconstruction.

[0117] OLTC is a complex nonlinear dynamic system. Phase space reconstruction can maintain the geometric invariance of the original nonlinear dynamic system. The vibration signal generated during the operation of OLTC is a complex nonlinear time series and contains rich equipment information. This sequence is an important information source for classifying OLTC mechanical fault characteristics. The intrinsic mode function sequence obtained by decomposing the original signal ICEEMDAN is a sequence in the time domain. By performing phase space reconstruction on it, more OLTC state information can be obtained.

[0118] For a given time series , N is the length of the time series, and the reconstructed phase space coordinates are expressed as: .

[0119] in, is the coordinate of the phase point in the reconstructed phase space, τ is the delay time, and m is the embedding dimension. The most critical parameters for reconstructing a one-dimensional time series into a high-dimensional phase space are the delay time τ and the embedding dimension m. The present invention uses the mutual information method to determine the delay time τ.

[0120] Assuming the original vibration signal , after a time delay of τ, the signal becomes , then the mutual information value between sequences S and Q can be obtained by the following information entropy and probability: .

[0121] Where: is the information entropy of signal S, is the information entropy of signal Q, is the joint information entropy of S and Q. The median value of the sampling point is The probability of a point appearing in the entire data segment is The median value of the sampling point is The probability of a point appearing in the entire data segment is The sampling value of the original signal is And the delayed signal sampling value is probability.

[0122] The present invention adopts the pseudo-nearest neighbor method to determine the optimal embedding dimension m. When the embedding dimension m is small, the trajectory in the phase space is not fully unfolded, resulting in folding and winding of the trajectory, which in turn makes the phase points that were originally far apart become closer. These close phase points are called pseudo-nearest neighbor points. In order to separate the pseudo-nearest neighbor points, it is necessary to gradually increase the embedding dimension m. When the embedding dimension m increases to the point where the pseudo-nearest neighbor points are basically reduced to zero, the corresponding embedding dimension m is the optimal embedding dimension m of the system.

[0123] For the vibration signals in the four states, the first four IMFs (i.e., IMF1, IMF2, IMF3, and IMF4) after ICEEMDAN decomposition are used to calculate their delay time τ and embedding dimension m. To ensure the consistency of variables, the present invention reconstructs all vibration signals in the same phase space, and takes the average of the delay time τ and embedding dimension m calculated by the IMF. The final reconstruction parameters are shown in Table 1.

[0124] Table 1

[0125] .

[0126] S3.2 Phase space surface reconstruction.

[0127] After completing the phase space reconstruction, the signal will be presented in the high-dimensional space in the form of reconstructed points. Through appropriate interpolation operations, these reconstructed points form a reconstructed surface. Since the surface fully reflects the information of all reconstructed points and changes accordingly when the OLTC state changes, the reconstructed surface can be used as a feature of the OLTC mechanical state.

[0128] The present invention uses the cubic interpolation method of the griddata function in Python to construct a reconstructed surface, and the steps are as follows:

[0129] S3.2.1 performs Delaunay triangulation on the reconstructed point set, ensuring that the circumscribed circle of each triangle does not contain other points, thereby constructing a triangular mesh covering all points.

[0130] S3.2.2 Within each triangle, the data is locally interpolated using a cubic polynomial, and within each triangular region, the values ​​between points are smoothly estimated using the cubic polynomial to generate a smooth and continuous surface, thereby reflecting the local changes in the reconstructed points.

[0131] S3.2.3 The discrete reconstruction points are connected into a smooth surface through cubic interpolation, which is the reconstruction surface. The surface reflects the information of all reconstruction points. When the mechanical state of OLTC changes, the surface generated by interpolation will also change accordingly.

[0132] This method can generate smoother and more continuous surfaces than linear interpolation, which is particularly important for phase space reconstruction results in high-dimensional space. Compared with linear interpolation, cubic interpolation can better capture the changing trend of data and ensure the smoothness of the interpolation results, which helps to accurately reflect the subtle changes in the OLTC state. By comparing the phase space trajectories of the phase space surface reconstruction results of IMF1-IMF4 after the original signal was decomposed by ICEEMDAN under various working conditions, it was found that there were significant differences in the phase trajectories of each inherent modal component under different working conditions. Due to the limitations of the time domain analysis method, the OLTC mechanical defects cannot be accurately identified by the time domain waveform of the vibration signal, but the phase trajectory characteristics of the vibration signal are equivalent to the characteristics of the original power system. The reconstructed surface can more clearly reflect the differences in vibration characteristics between various working conditions and can well reflect the characteristics of the vibration signal.

[0133] S3.3 Gray-level co-occurrence matrix feature extraction.

[0134] S3.3.1Construct the gray-level co-occurrence matrix.

[0135] The phase space reconstructed surface accurately contains the information of all reconstructed points. For a three-dimensional surface, its Gaussian curvature, surface normal vector and other surface inherent geometric indicators can be calculated as signal features, but the amount of calculation is large and the process is cumbersome. To simplify the calculation, the present invention uses the gray level co-occurrence matrix to process the XY plane projection of the phase space reconstructed surface to extract image features.

[0136] Gray Level Co-occurrence Matrix (GLCM) is a tool for image texture analysis that can describe the spatial relationship between grayscale values ​​in an image. The basic idea is to count the grayscale relationship between adjacent pixels in an image and construct a matrix that reflects the co-occurrence of different grayscale levels. The grayscale co-occurrence matrix counts the frequency of occurrence of the grayscale value combination of each pixel in the image and its neighboring pixels by setting a fixed relative distance and direction. Different distances and directions can reflect the texture features of images at different scales and directions. Common directions are 0°, 45°, 90°, and 135°, and the common distance is 1 pixel. Given an image I, its size is , the gray-level co-occurrence matrix is ​​defined as: .

[0137] in, is the gray level, ranging from , is the grayscale value; is the relative distance, is the relative direction, Represents the horizontal displacement of the pixel. Represents the vertical displacement of the pixel. and Depends on the relative distance and relative direction The gray-level co-occurrence matrix cannot be used directly to describe the texture features of an image. The present invention selects the contrast, energy, homogeneity and correlation of the gray-level co-occurrence matrix as image texture features for feature extraction.

[0138] (1) Contrast Con: measures the intensity of image grayscale differences. The calculation formula is: .

[0139] (2) Energy E: reflects the uniformity of the texture. The larger the energy, the more uniform the texture. The calculation formula is: .

[0140] (3) Homogeneity H: measures the similarity of adjacent grayscale values. The calculation formula is: .

[0141] (4) Correlation Cor: reflects the linear correlation of gray values ​​between pixels. The calculation formula is: .

[0142] in Is grayscale The mean of Is grayscale The mean of , is the standard deviation.

[0143] S3.3.2 Texture feature extraction of surface projection image from phase space reconstruction.

[0144] Through color mapping (such as grayscale or color), the projection image can show the change of Z value (height) in the three-dimensional surface. In the XY plane, the color difference of different areas reflects the height difference of the corresponding position on the surface. At the same time, the projection image can reflect the shape, boundary and distribution of the three-dimensional surface in the XY plane, and show the overall layout of the data. In addition, the local change trend on the surface, such as the peak, valley or smooth area in certain areas, will also be reflected in the projection image through color changes. The specific steps for extracting texture features of the phase space reconstruction surface projection image are as follows:

[0145] S3.3.2.1 In order to ensure that the same data is mapped to the same grayscale value, the present invention uses the global minimum and maximum values ​​for normalization. The overall minimum and maximum values ​​of the data set are calculated to ensure that the data uses the same normalization standard. The formula is as follows:

[0146] .

[0147] In the formula, is the normalized data.

[0148] S3.3.2.2 Coarse quantization of gray levels.

[0149] The gray level is reduced to a suitable value without affecting the texture characteristics of the image. In the present invention, the gray level of 256 is quantized into 48 levels.

[0150] S3.3.2.3 Draw a grayscale image.

[0151] Map the data to grayscale between 0 and 47 and draw a grayscale image of the surface projected onto the XY plane.

[0152] S3.3.2.4 Calculate the gray-level co-occurrence matrix.

[0153] The distance is taken as 1 pixel, and the gray-level co-occurrence matrix in the four directions of 0°, 45°, 90°, and 135° is calculated.

[0154] S3.3.2.5 Calculate the four eigenvalues ​​of contrast, energy, homogeneity and correlation for each gray-level co-occurrence matrix.

[0155] S4 sample division.

[0156] After completing the above steps, for each group of original signals, 4 IMFs (IMF1, IMF2, IMF3 and IMF4) are decomposed, and 16-dimensional features are extracted for each IMF. Finally, a 64-dimensional feature vector is obtained for each group of signals to form a training set and a test set, which are then input into the classifier for training and testing.

[0157] Fig.14 , Fig.15 , Fig.16 and Fig.17 The grayscale images of the reconstructed surface projection of a group of IMF2 phase space under four working conditions are shown in Table 2.

[0158] Table 2

[0159] .

[0160] S5. OLTC fault diagnosis and identification based on Bayesian optimization CatBoost.

[0161] CatBoost (Categorical Boosting) is an algorithm based on gradient boosting decision trees GBDT (GradientBoosting Decision Trees), which is mainly used to solve data problems with categorical features. It optimizes on the additive model through gradient descent to solve the target leakage and gradient deviation problems existing in the traditional gradient boosting decision tree. CatBoost is a complex model with multiple hyperparameters, and manual adjustment of these hyperparameters often requires a lot of experiments and time, and the results cannot guarantee the best accuracy. In order to solve this problem, the present invention introduces Bayesian optimization CatBoost, which quickly finds the optimal combination in the hyperparameter space through an efficient search algorithm.

[0162] The main steps of the Bayesian optimization CatBoost algorithm are:

[0163] (1) Initialize the model .

[0164] (2) For each iteration, calculate the residual of the current model , fitting a new weak learner based on the residual And update the model: .

[0165] (3) Based on the early stopping mechanism or the specified number of iterations, a series of weak learners The final model is constructed cumulatively: .

[0166] Using Bayesian optimization CatBoost can significantly reduce the time required for parameter adjustment, which not only improves optimization efficiency, but also improves model performance and accuracy. Bayesian hyperparameter automatic optimization is used to optimize the five hyperparameters in the CatBoost model, namely, iterations, tree depth, learning rate, L2 regularization parameter 12 leaf reg, and bagging_temperature. The search range of hyperparameters in the Bayesian hyperparameter automatic optimization framework is shown in Table 3. Based on the hyperparameter search range shown in Table 3, the optimal parameter results obtained by the Bayesian hyperparameter automatic optimization framework are shown in Table 4.

[0167] Table 3

[0168] .

[0169] Table 4

[0170] .

[0171] The optimal hyperparameters obtained by Bayesian hyperparameter automatic optimization are used to train the training set using the CatBoost model and tested on the test set. The confusion matrix is ​​as follows: Fig.18 shown.

[0172] In order to verify the rationality of the model selection, the present invention uses 5-fold cross validation for testing. The classification model accuracy and cross-validation accuracy are shown in Table 5. The results show that the correct classification ratio of the model on the test set is 95%. The classification accuracy of 95% and the cross-validation accuracy of 93% indicate that the model performs well on both the training and test sets, and has good classification ability and certain robustness.

[0173] Table 5

[0174] .

[0175] The present invention constructs a systematic and efficient diagnostic process by analyzing and processing vibration signals, from signal decomposition to feature extraction, and then to the application of classification models. The present invention can not only effectively extract key features in the signal, but also further enhance the representativeness of the features through gray-level co-occurrence matrix analysis, and realize high-precision fault identification and diagnosis in combination with Bayesian optimization CatBoost classifier.

[0176] Aiming at the fault diagnosis and identification problem of KM type oil-immersed online transformer tap changer OLTC, this paper adopts vibration signal analysis, feature extraction and machine learning technology, and proposes an effective fault classification method, aiming to improve the reliability and operation efficiency of the equipment. (1) Through the detailed experimental design of vibration signal collection of OLTC under different working conditions, common fault conditions are simulated, including loose arc plate and contact, loose transmission shaft screw and transmission gear jamming, to ensure a comprehensive analysis of various fault states of OLTC. (2) The collected vibration signal is processed by ICEEMDAN decomposition technology to extract the intrinsic mode function IMF containing rich fault features. Through phase space reconstruction, a high-dimensional feature space is constructed, which provides reliable support for fault classification. In addition, the gray level co-occurrence matrix GLCM is used to process the XY plane projection of the phase space reconstruction surface, and multi-dimensional image features such as contrast, energy, homogeneity and correlation are extracted, further enriching the feature description. Phase space reconstruction can expand one-dimensional time series data into multi-dimensional phase space through delayed embedding, so that the dynamic behavior of the signal, such as periodicity and chaos, can be more clearly revealed, which is helpful for analyzing the dynamic characteristics of the equipment and lays a solid theoretical foundation for nonlinear time series analysis. Due to the high sensitivity of phase space reconstruction to initial conditions, it can amplify the slight changes in the vibration signal of the equipment and realize accurate identification of faults. (3) CatBoost was selected as the fault classification model, and the hyperparameters were systematically adjusted using Bayesian optimization to obtain the optimal hyperparameter configuration. This process significantly improved the classification accuracy of the model and enhanced its robustness. After training and testing, the classification accuracy of the model on the test set reached 95%, and the classification accuracy of 93% was obtained through 5-fold cross validation. This shows that the proposed model has good performance and reliability in practical applications.

Claims

1. A method for fault diagnosis of on-load tap changer based on phase space reconstruction theory, characterized in that: The following steps are involved: S1. Data Collection S1.1 Install an acceleration sensor on the top of the OLTC to collect OLTC vibration signals; S1.2 sets four working conditions to simulate common OLTC faults. The four working conditions include the normal state, and three mechanical fault conditions: loose arc plate and contactor, loose transmission shaft screws, and stuck transmission gears. S1.3 performs multiple tests in the same gear and multiple tests in multiple gears under normal OLTC conditions and three mechanical fault conditions, and collects 160 sets of data for each condition; S2. Data selection S2.1 Add i groups of white noise modal components to the original signal sequence to construct N noisy signals, that is, In the formula, represents the i-th noisy constructed signal; x(t) is the original signal; β1 represents the signal-to-noise ratio; E represents EMD decomposition; ω (i )(t) represents white noise; S2.2 uses EMD to repeatedly decompose the noisy signal N times, then subtracts it from the noisy signal and calculates the mean to obtain the first residual signal R1(t) and the corresponding modal component IMF1; IMF1=x(t)-R1(t) (3) S2.3 Add multiple groups of white noise components to R1(t), and get And find the mean value and then make a difference with R1(t) to get the modal component IMF2; IMF2=R1(t)-R2(t) (6) It is the noise-added signal after adding white noise again on the basis of the first residual signal R1(t), β2 represents the signal-to-noise ratio; R2(t) represents the second residual signal; S2.4 Repeat step S2.3 until all residual signals cannot be further decomposed, and all modal components IMF are obtained; modal components IMF1-IMF4 are taken as processing objects for subsequent further feature extraction; S3. Feature extraction S3.1 Phase space reconstruction For a given time series {x(i), i = 1, 2, ..., N}, where N is the length of the time series, the reconstructed phase space coordinates are expressed as: Where X(i) is the coordinate of the phase point in the reconstructed phase space; τ is the delay time; m is the embedding dimension; S3.2 Phase space surface reconstruction Use the cubic interpolation method of the griddata function in Python to construct the reconstructed surface; S3.3 Gray-level co-occurrence matrix feature extraction S3.3.1 Constructing the gray-level co-occurrence matrix Given an image I with size M×N, the gray-level co-occurrence matrix is ​​defined as: Among them, i, j are grayscale levels, ranging from [0, L-1], L is the grayscale value; d is the relative distance, θ is the relative direction, Δx represents the horizontal displacement of the pixel, Δy represents the vertical displacement of the pixel, Δx and Δy depend on the relative distance d and the relative direction θ; the contrast Con, energy E, homogeneity H and correlation Cor of the grayscale co-occurrence matrix are selected as image texture features for feature extraction; 1) Contrast Con: measures the intensity of image grayscale difference, and the calculation formula is: 2) Energy E: reflects the uniformity of the texture. The larger the energy, the more uniform the texture. The calculation formula is: 3) Homogeneity H: measures the similarity of adjacent grayscale values, and the calculation formula is: 4) Correlation Cor: reflects the linear correlation of gray values ​​between pixels, and the calculation formula is: where μ i is the mean value of gray level i, μ j is the mean value of gray level j; σ i , σ j is the standard deviation; S3.3.2 Phase space reconstruction and surface projection image texture feature extraction S3.3.2.1 Use global minimum and maximum values ​​for normalization. Calculate the overall minimum min(x) and maximum max(x) of the data set to ensure that the data uses the same normalization standard. The formula is as follows: In the formula, x′ is the normalized data; S3.3.2.2 Grayscale coarse quantization Reduce the gray level to a suitable value without affecting the texture characteristics of the image, and quantize the gray level of 256 to 48 levels; S3.3.2.3 Drawing grayscale images Map the data to grayscale levels between 0 and 47, and draw a grayscale image of the surface projected onto the XY plane; S3.3.2.4 Calculation of gray-level co-occurrence matrix Taking the distance as 1 pixel, calculate the gray level co-occurrence matrix in the four directions of 0°, 45°, 90°, and 135°; S3.3.2.5 Determine the four eigenvalues ​​of contrast, energy, homogeneity, and correlation for each gray-level co-occurrence matrix; S4. Sample Division For each group of original signals, modal components IMF1, IMF2, IMF3 and IMF4 are decomposed, and 16-dimensional features are extracted for each modal component IMF. Finally, a 64-dimensional feature vector is obtained for each group of signals to form a training set and a test set, which are then input into the classifier for training and testing. S5. OLTC Fault Diagnosis and Identification Based on Bayesian Optimization CatBoost S5.1 Initialize model F0(x); S5.2 For each iteration, calculate the residual of the current model Fitting a new weak learner h based on the residual m (x) and update the model: F(x)=F m-1 (x)+γ m h m (x) (14) Among them, F m-1 (x) represents the model after the m-1th iteration, γ m is the corresponding learning rate; S5.3 Based on the early stopping mechanism or the specified number of iterations, through a series of weak learners h m (x) Accumulate and build the final model: Bayesian hyperparameter automatic optimization is used to optimize the five hyperparameters in the CatBoost model, including iterations, tree depth, learning rate, L2 regularization parameter 12 leaf reg, and bagging_temperature, to obtain the optimal parameters of the CatBoost model.

2. The on-load tap changer fault diagnosis method based on phase space reconstruction theory according to claim 1 is characterized in that: Three mechanical failure conditions are simulated as follows: Fault 1: The loosening of the arc plate is simulated by loosening a screw ring of the static contact on the A-phase arc plate of the OLTC; the loosening of the contact is simulated by loosening the screws related to the neutral point contact and the main on-off contact; Fault 2: Loose screws on the transmission shaft are used to simulate the fault by loosening the screws at the connection of the transmission rod; Fault 3: The transmission gear is stuck. This is done by stuffing wood chips into the gear box of the switch transmission mechanism to increase the friction between the gears, thereby simulating the failure of the transmission shaft being stuck.

3. The on-load tap changer fault diagnosis method based on phase space reconstruction theory according to claim 1 is characterized in that: The mutual information method is used to determine the delay time τ: Assuming that the original vibration signal S = {x(i), i = 1, 2, ..., N}, after a time delay τ, the signal becomes Q = {x(i+τ), i = 1, 2, ..., N}, then the mutual information value between the sequences S and Q is obtained by the following information entropy and probability: Where: H(S) is the information entropy of signal S; H(Q) is the information entropy of signal Q; H(S,Q) is the joint information entropy of S and Q; P s (s i ) is the median value of the sampling point s i The probability of a point appearing in the entire data segment; p q (q j ) is the median value of the sampling point q j The probability of a point appearing in the entire data segment; P sq (s i ,q j ) is the sampling value of the original signal s i And the delayed signal sampling value is q j probability.

4. The on-load tap changer fault diagnosis method based on phase space reconstruction theory according to claim 1 is characterized in that: The pseudo nearest neighbor method is used to determine the optimal embedding dimension m: When the embedding dimension m is small, the trajectory in the phase space is not fully unfolded, resulting in folding and entanglement of the trajectory, which in turn makes the phase points that were originally far apart become closer. These close phase points are called pseudo neighbor points. In order to separate the pseudo neighbor points, it is necessary to gradually increase the embedding dimension m. When the embedding dimension m increases to the point where the pseudo neighbor points are basically reduced to zero, the corresponding embedding dimension m is the optimal embedding dimension m of the system.

5. The on-load tap changer fault diagnosis method based on phase space reconstruction theory according to claim 1 is characterized in that: For the vibration signals under four working conditions, the delay time τ and the embedding dimension m of the modal components IMF1, IMF2, IMF3 and IMF4 after ICEEMDAN decomposition are calculated respectively; To ensure the consistency of variables, all vibration signals are reconstructed in the same phase space, and the delay time τ and the embedding dimension m calculated from the modal component IMF are averaged to finally determine the reconstruction parameters τ and m.

6. The on-load tap changer fault diagnosis method based on phase space reconstruction theory according to claim 1 is characterized in that: The specific content of step S3.2 is: S3.2.1 Perform Delaunay triangulation on the reconstructed point set to ensure that the circumscribed circle of each triangle does not contain other points, thereby constructing a triangular mesh covering all points; S3.2.2 In each triangle, the data is interpolated locally using a cubic polynomial, and the values ​​between points are smoothly estimated by the cubic polynomial within each triangular region to generate a smooth and continuous surface, thereby reflecting the local changes of the reconstructed points; S3.2.3 The discrete reconstruction points are connected into a smooth surface through cubic interpolation, which is the reconstruction surface. The surface reflects the information of all reconstruction points. When the mechanical state of OLTC changes, the surface generated by interpolation will also change accordingly.

Images