Extra-high voltage converter transformer OLTC fault diagnosis method fusing multi-feature information and RIME-CNN-BiLSTM-SAM

Abstract

The invention relates to an extra-high voltage converter transformer OLTC fault diagnosis method fusing multi-feature information and RIME-CNN-BiLSTM-SAM, and belongs to the technical field of intelligent monitoring of power systems. The method comprises the following steps: firstly, constructing a vibration signal fault data set based on an OLTC mechanical fault simulation experiment; secondly, decomposing an OLTC vibration signal into a plurality of intrinsic mode components by adopting REMD, and preferably selecting components containing rich fault feature information based on a correlation coefficient; then, the multi-scale permutation entropy of IMFs is calculated, time / frequency statistical analysis and the multi-scale permutation entropy are combined, OLTC multi-dimensional vibration characteristic parameters are integrally and locally concerned, and principal component analysis is introduced to carry out characteristic fusion and dimension reduction of OLTC vibration information; and finally, optimizing the hyper-parameters of the CNN-BiLSTM-SAM fault diagnosis model by using the RIME, and realizing the construction of the converter transformer OLTC intelligent fault diagnosis model. The method has a remarkable effect in the aspect of extracting OLTC fault feature information, the fault type recognition rate is superior to that of an existing method, and the effectiveness of the method in converter transformer OLTC fault diagnosis is verified.

Description

Technical Field

[0001] The present invention relates to a UHV converter transformer OLTC fault diagnosis method integrating multi - feature information and RIME - CNN - BiLSTM - SAM, belonging to the technical field of intelligent monitoring of power systems. Background Technique

[0002] With the vigorous development of UHV transmission technology, the importance of converter transformers in the power system has become increasingly prominent. Among them, the on - load tap changer (OLTC), as a core component in the converter transformer, mainly realizes the reactive power regulation of the power system, the stability of the load voltage, and the economic and flexible operation of the DC system. However, due to the influence of the working environment and the operation frequency, the fault frequency of the UHV converter transformer OLTC is relatively high, and its fault complexity poses a potential threat to the safe operation of the power grid. Therefore, timely and accurate diagnosis of the UHV converter transformer OLTC fault is required to take effective treatment measures to ensure the safe operation of the power grid.

[0003] Vibration signal analysis technology is the most promising method applied to OLTC fault diagnosis and on - line monitoring. Relevant researchers have carried out a large number of works on the feature extraction and fault diagnosis of OLTC vibration signals. In terms of vibration signal extraction, wavelet packet decomposition, adaptive signal decomposition algorithms, and various improved methods have been used to improve the signal - to - noise ratio of vibration signals. However, due to the complex structure of the OLTC, there are many spikes in the OLTC vibration signals, and each method has certain deficiencies, resulting in the need to improve the signal decomposition accuracy. In terms of fault diagnosis, with the application of deep learning, many fault diagnosis models have been developed. Some have limited fault recognition types, some have insufficient ability to identify compound faults, and some models are relatively complex and not conducive to practical applications. Summary of the Invention

[0004] To effectively improve the accuracy and precision of existing OLTC fault diagnosis, the present invention provides a UHV converter transformer OLTC fault diagnosis method integrating multi - feature information and RIME - CNN - BiLSTM - SAM.

[0005] The technical solution of the present invention is as follows: A UHV converter transformer OLTC fault diagnosis method integrating multi - feature information and RIME - CNN - BiLSTM - SAM includes the following steps:

[0006] Conduct vibration simulation experiments on the OLTC under different working conditions to construct a vibration signal dataset;

[0007] Use REMD (Robust Empirical Mode Decomposition) to decompose the OLTC vibration signal into multiple IMFs (Intrinsic Mode Functions), and preferably select the IMFs rich in fault feature information based on the correlation coefficient;

[0008] Calculate the multi-scale permutation entropy of the selected IMFs, combine time / frequency statistical analysis and multi-scale permutation entropy to extract multi-feature information of vibration signals, and introduce principal component analysis to carry out feature fusion and dimensionality reduction of OLTC vibration information;

[0009] Build a CNN-BiLSTM-SAM fault diagnosis model, optimize the hyperparameters of the model using RIME (rime optimization algorithm), and then use the processed data set to train the model to obtain a RIME-CNN-BiLSTM-SAM fault diagnosis model for diagnosing the OLTC of the converter transformer.

[0010] Furthermore, the different working conditions of the OLTC include normal conditions, drive shaft jamming faults, contact ablation faults, and contact wear faults.

[0011] Furthermore, the process of REMD decomposing the OLTC is as follows:

[0012] Let E i (r) be the i-th qualified r component obtained by EMD decomposition, x(n) be the target signal to be decomposed, and k be the number of implementation times. The specific steps are as follows:

[0013] (1) Import the original data for initialization, h k [n] = G i (n);

[0014] (2) Calculate the mean envelope line m k [n] from h k [n]. The calculation formula is as follows:

[0015]

[0016] In the formula, g max [n] represents the upper envelope line, and g min [n] represents the lower envelope line;

[0017] (3) Screen to obtain the signal h1[n]:

[0018] h1[n] = h0[n] - m0[n]

[0019] (4) Judge whether the extracted modal component r meets the conditions; if it does not meet the conditions, use h1[n] as the new initial signal, and repeat steps (2) to (4) until the first r component is obtained;

[0020] Before reaching the maximum iteration number D max When:

[0021] r1 = h k-2 [n]

[0022] Reach the maximum number of iterations D max When:

[0023] r1 = h k [n]

[0024] Where h k [n] represents the signal after screening and iteration;

[0025] (5) Obtain a new initial signal G i (n), and determine whether G i (n) is monotonic or a constant. If not, repeat steps (1) to (5) until the initial signal cannot be decomposed; at this time, k components are obtained, and the final residual component is expressed as R(n), then the initial signal x(n) is expressed as:

[0026]

[0027] Where I is the total number of modal components.

[0028] Furthermore, the correlation coefficient for the preferred IMFs process is as follows:

[0029] (1) The correlation coefficient ρ between the original vibration signal x(t) and its IMFs, h1(t), h2(t), …, h n (t) after REMD decomposition is calculated as follows: i The formula is:

[0030]

[0031] Where cov() is the covariance; σ is the standard deviation; h i (t) represents the signal after screening and iteration;

[0032] (2) Set the threshold Th as the standard deviation of the correlation coefficient, that is:

[0033]

[0034] If ρ i > Th, then retain the i-th IMF, otherwise reject the i-th IMF.

[0035] Furthermore, the steps for data dimensionality reduction by the principal component analysis are as follows:

[0036] 1) Assume the number of samples is m, and the eigenvector corresponding to each sample is n. The eigenvector matrix H can be expressed as an n×m order matrix;

[0037] First, normalize the data to generate a standard matrix H * :

[0038]

[0039] In the formula, represents the mean value of each dimension, and S j represents the variance of each dimension;

[0040] 2) Construct the feature covariance matrix Σ:

[0041]

[0042] 3) Obtain the eigenvalues of Σ and sort them in descending order. Select the eigenvectors corresponding to the first l eigenvalues to form a new vector matrix U, and calculate the principal component Z:

[0043] Z = U T X * .

[0044] Furthermore, the CNN-BiLSTM-SAM fault diagnosis model includes a convolutional layer, a pooling layer, a BiLSTM layer, a self-attention layer, and a fully connected layer; first, the convolutional layer captures the local features of the input data, and the pooling layer reduces the spatial dimension of the feature map and retains the key features; then, the BiLSTM structure is used to extract the temporal features hidden in the spatial feature information, and the extracted spatial and temporal feature information is input to the next layer to establish the logical relationship between the input parameters and the output parameters; then, the self-attention layer is used to define the importance of the features, so as to identify the long-term dependence relationship between the input elements. Finally, it is input to the fully connected layer for weight allocation.

[0045] Furthermore, the RIME optimization model parameters include a regularization coefficient, a learning rate, and the number of hidden neurons. The soft rime search strategy is used to conduct extensive parameter searches in the initial stage, and the hard rime puncture mechanism is combined to make precise adjustments in the later stage to determine a high-performance hyperparameter combination.

[0046] Advantages of the present invention:

[0047] (1) The present invention uses the robust empirical mode decomposition (REMD) algorithm to decompose the OLTC fault vibration signal, which can effectively alleviate the mode mixing effect of the EMD method, thereby improving the decomposition accuracy.

[0048] (2) The composite feature analysis method that fuses time-domain, frequency-domain, and multi-scale permutation entropy feature information realizes a more comprehensive extraction of on-load tap-changer fault information. The feature fusion method based on PCA effectively improves the fault recognition accuracy and reduces the influence of redundant information on the diagnosis efficiency.

[0049] (3) Utilize the efficient search mechanism and forward greedy mechanism of the RIME algorithm, with the minimum envelope entropy as the fitness function, to obtain the global optimal parameter combination under different operating conditions of the on-load tap-changer, and improve the fault diagnosis accuracy of the model.

[0050] (4) The fault diagnosis model of the UHV converter transformer OLTC that integrates multi-feature information and RIME-CNN-BiLSTM-SAM has significantly optimized the diagnosis effect, and the model proposed by the present invention has higher diagnostic accuracy when diagnosing on-load tap-changer faults. Description of the Drawings

[0051] Figure 1 is the structure diagram of the CNN-BiLSTM-SAM model in this embodiment;

[0052] Figure 2 is the construction diagram of the RIME-CNN-BiLSTM-SAM model in this embodiment;

[0053] Figure 3 is the REMD decomposition flow chart. Detailed Implementation Manner

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention. It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be arbitrarily combined with each other.

[0055] Embodiment 1: A UHV converter transformer OLTC fault diagnosis method that integrates multi-feature information and RIME-CNN-BiLSTM-SAM specifically includes the following steps:

[0056] (1) OLTC vibration simulation experiment

[0057] In this embodiment, a UCG type on-load tap-changer produced by ABB is used to carry out a prototype simulation test on the OLTC of the spare converter transformer in the West Converter Station. The OLTC body is respectively simulated in normal and spring fatigue states. The ULT2001 piezoelectric acceleration sensor produced by Beijing Kangtai Technology Co., Ltd. is selected and installed on the edge of the OLTC prototype cover.

[0058] By analyzing the mechanical faults that may occur in the on-load tap-changer during operation, vibration simulations and comparative studies are respectively carried out under 4 working conditions: normal condition (N), transmission gear jamming fault (G1), contact ablation fault (G2), and contact wear fault (G3). Among them, the jamming fault is simulated by adding wood chips to the gear box; for the contact system in the switching switch, the contact ablation fault is simulated by using a drill to grind the contact surface rough; the contact wear fault is simulated by using a grinding machine to thin the contact.

[0059] Collect and organize the data collected under each working condition into a data set.

[0060] (2) Use the Ensemble Empirical Mode Decomposition (EEMD) to decompose the on-load tap changer (OLTC) vibration signal into multiple Intrinsic Mode Functions (IMFs), and select the IMFs with rich fault feature information based on the correlation coefficient.

[0061] The steps of the EEMD signal decomposition are as follows:

[0062] The EEMD defines the screening stop determination condition as F(k), and F(k) should be as small as possible.

[0063]

[0064] F(k) = e RMS + |E k |

[0065] In the formula, g k [n] represents the component obtained after k iterations of the original signal, e RMS represents the root mean square of g k [n], and E k represents the excess kurtosis of g k [n], and n is the signal length.

[0066] The EEMD screening stop process:

[0067] (1) The number of zero points (N zp ) and extreme points (N ep ) is the same, or the difference between the two is less than 1.

[0068] (2) If F k-2 < F k-1 and F k-1 < F k , then stop the screening process and return the (k - 2)th decomposition result; if the condition is not met, continue the screening process until the maximum number of iterations is reached.

[0069] The decomposition flow chart of the EEMD is as Figure 3 shown.

[0070] Let E i (r) be the rth component that meets the conditions obtained by the EMD decomposition for the i-th (i = 0, 1, 2,...), x(n) be the target signal to be decomposed, and k be the number of realizations (k = 0, 1, 2,..., D max ).

[0071] The specific steps are as follows:

[0072] (1) Import the original data for initialization, h k [n] = G i(n);

[0073] (2) Obtain the mean envelope m k [n] from h k [n]. The calculation formula is as follows:

[0074]

[0075] In the formula, g max [n] represents the upper envelope, and g min [n] represents the lower envelope.

[0076] (3) Screen to obtain the signal h1[n], and obtain F(k) from m0[n]

[0077] h1[n] = h0[n] - m0[n]

[0078] (4) Judge whether the extracted modal component r meets the conditions; if it does not meet the conditions, use h1[n] as the new initial signal and repeat steps (2) to (4) until the first r component is obtained.

[0079] Before reaching the maximum number of iterations D max :

[0080] r1 = h k-2 [n]

[0081] When reaching the maximum number of iterations D max :

[0082] r1 = h k [n]

[0083] In the formula, h k [n] represents the signal after screening and iteration.

[0084] (5) Obtain the new initial signal G i (n) (i = 1, 2,...), judge whether G i (n) is monotonic or constant. If not, repeat steps (1) to (5) until the initial signal cannot be decomposed. At this time, k components are obtained, and the final residual component is expressed as R(n). Then the initial signal x(n) can be expressed as:

[0085]

[0086] In the formula, I is the total number of modal components.

[0087] Among the several IMFs obtained by decomposing the vibration signal of the on-load tap-changer of the converter transformer using REMD, some IMFs contain the fault information of the original vibration signal, while the remaining IMFs are interference components caused by iterative errors during decomposition and on-site noise. To effectively extract the fault information of the vibration signal, it is necessary to remove the IMFs irrelevant to the fault and interference to improve the accuracy of fault recognition. In this embodiment, the correlation coefficient method is used to select the effective IMFs.

[0088] The correlation coefficient reflects the degree of correlation between the IMFs and the original vibration signal. The IMFs containing the main fault information have a high correlation with the original vibration signal, while the correlation of the remaining IMFs is low.

[0089] (1) The correlation coefficient ρ between the original vibration signal x(t) and its IMFs, h1(t), h2(t), …, h n (t) obtained by REMD decomposition is i Calculated by the formula:

[0090]

[0091] where cov() is the covariance; σ is the standard deviation; h i (t) represents the signal after screening and iteration.

[0092] (2) Set the threshold Th as the standard deviation of the correlation coefficient, that is:

[0093]

[0094] If ρ i > Th, then retain the i-th IMF, otherwise remove the i-th IMF.

[0095] In this embodiment, the correlation coefficients of the IMFs are calculated and the IMFs with correlation coefficients greater than the threshold are selected as sensitive IMFs.

[0096] The correlation coefficients in the four states of the OLTC are shown in Table 1.

[0097] Table 1 Correlation coefficients of IMF components under four OLTC working conditions

[0098] IMF component Normal Jamming Contact ablation Contact wear Threshold 0.2548 0.2580 0.2579 0.2617 <![CDATA[IMF1]]> 0.7636 0.7650 0.6491 0.8193 <![CDATA[IMF2]]> 0.4634 0.5097 0.6591 0.3747 <![CDATA[IMF3]]> 0.0852 0.2004 0.1770 0.3285 <![CDATA[IMF4]]> 0.0710 0.1270 0.2087 0.1096 <![CDATA[IMF5]]> 0.0747 0.1083 0.0793 0.0654 <![CDATA[IMF6]]> 0.0513 0.0496 0.0566 0.0532 <![CDATA[IMF7]]> 0.0205 0.0117 0.0046 0.0365 <![CDATA[IMF8]]> 0.0052 0.0074 0.0140 0.0147 <![CDATA[IMF9]]> -0.0015 0.0004 0.0012 -0.0021 <![CDATA[IMF 10 > -0.0018 0.0001 -0.0005 0.0007

[0099] As can be seen from Table 1, the correlation coefficient thresholds for the normal, jammed, contact ablation, and contact wear conditions of the OLTC are 0.2548, 0.2580, 0.2579, and 0.2617, respectively. The correlation coefficients between IMF1 and IMF2 in the normal condition, IMF1 and IMF2 in the jammed fault, IMF1 and IMF2 in the contact ablation fault, and IMF1 - IMF3 in the contact wear fault and the original signal are relatively high. Therefore, these IMF components with relatively high correlation coefficients can be selected for experimental analysis.

[0100] (3) Calculate the multi-scale permutation entropy of the selected IMFs, combine time / frequency statistical analysis and multi-scale permutation entropy to extract multi-feature information of the vibration signal, and introduce principal component analysis to carry out feature fusion and dimensionality reduction of the OLTC vibration information.

[0101] Based on the correlation coefficient method, the IMF components of the four working conditions of the OLTC are optimized, and their multi-scale permutation entropy values are calculated to form a new feature vector. In this embodiment, the length N of a single sample signal is selected as 1280, the embedding dimension m is controlled to be 2 - 8, and the delay time τ is taken as 1 - 10.

[0102] Calculate the multi-scale permutation entropy values of the IMF components with relatively high correlation coefficients screened by the REMD algorithm and the correlation coefficient method for the 4 working conditions of the on-load tap-changer respectively. Select the PE means with scale factors s of 8 and 10 to form the composite vibration fault features. Therefore, 360 groups of fault feature samples corresponding to the four working conditions of the OLTC can be obtained, and the fault sample set is dimensionally reduced by the principal component analysis method for secondary feature extraction. And divide the training set and the test set according to the ratio of 8:2.

[0103] (1) Time-domain features

[0104] Time-domain analysis is the most commonly used method for extracting the features of equipment fault signals, and the extraction of time-domain feature parameters is the most crucial. The time-domain features of the vibration signals of the OLTC under different faults are also different. Therefore, it is necessary to calculate the signal features in the time domain scale.

[0105] The commonly used time-domain feature parameters and their mathematical calculation formulas are shown in Table 2.

[0106] Table 2 Time-domain feature parameters and their calculation formulas

[0107]

[0108] (2) Frequency-domain features

[0109] When diagnosing faults in equipment, relying solely on time-domain feature analysis cannot accurately determine the fault type. Therefore, it is necessary to combine frequency-domain feature analysis. Different fault states of the OLTC will have significant changes in its features on the frequency-domain scale compared to the normal state. The collected OLTC fault signals are transformed into spectrograms through the Fast Fourier Transform (FFT) to achieve the conversion of OLTC fault time-domain signals to frequency-domain signals. The commonly used frequency-domain feature parameters and their mathematical calculation formulas are shown in Table 3.

[0110] Table 3 Frequency-domain Feature Parameters and Their Calculation Formulas

[0111]

[0112]

[0113] Among them, f(k) is the spectrum of the signal x(n), k = 1, 2,..., K (K is the spectral line value), and f k is the frequency value of the k-th spectral line.

[0114] In the above table:

[0115]

[0116] p i is used to calculate the probability of each frequency point.

[0117] (3) Multiscale Permutation Entropy

[0118] Permutation Entropy (PE) is an average entropy function that measures the complexity of time series and detects dynamic mutations, and is often used to extract mechanical fault features. However, the fault feature information of the OLTC during operation is distributed in multiple scales. Analyzing only with the permutation entropy of a single scale will miss the fault feature information in other scales. Multiscale permutation entropy is to coarsen the time series at multiple scales on the basis of permutation entropy, and then calculate the permutation entropy of the coarsened series at different scales. The specific calculation process is as follows:

[0119] 1) Assume there is a time series X = {x1, x2,..., x N} of length N, and after coarsening, the sequence obtained is:

[0120]

[0121] In the formula, s is the scale factor, and [N / s] represents taking the integer.

[0122] 2) Reconstruct y j (s) to obtain

[0123]

[0124] In the formula, l represents the l-th reconstructed component; τ represents the delay time; m represents the embedding dimension.

[0125] 3) Arrange the reconstructed sequences in ascending order:

[0126]

[0127] After reconstruction, a set of sequences S(g) = {j1, j2,..., j m} is obtained, where g = 1, 2,..., r and r ≤ m, and calculate the probability {P1, P2,..., P r} of each sequence appearing.

[0128] The permutation entropy of the time series vibration signal is:

[0129]

[0130] 4) When P r = 1 / m!, the probabilities of each sequence are the same. At this time, the complexity of the time series is the highest, and the permutation entropy is the largest, which is ln(m!). Perform normalization processing on the multi-scale permutation entropy to obtain:

[0131]

[0132] The value range of H p is [0, 1], and the magnitude of its value reflects the complexity and randomness of the time series. The larger H p , the more random the time series is. On the contrary, it means that the time series is more regular.

[0133] (4) Feature-level fusion method of principal component analysis

[0134] To reduce the complexity of the feature vectors extracted by the composite statistical features, this embodiment introduces the principal component analysis method (Principal Component Analysis, PCA) for data dimensionality reduction. The specific calculation steps are as follows:

[0135] 1) Assume that the number of samples is m, the feature vector corresponding to each sample is n, and the feature vector matrix H can be expressed as an n×m order matrix.

[0136] First, perform normalization processing on the data to generate the standard matrix H * :

[0137]

[0138] In the formula, represents the mean value of each dimension, and S j represents the variance of each dimension.

[0139] 2) Construct the feature covariance matrix ∑:

[0140]

[0141] 3) Obtain the eigenvalues of ∑ and sort them in descending order. Select the eigenvectors corresponding to the first l eigenvalues to form a new vector matrix U, and calculate the principal component Z.

[0142] Z = U T X * .

[0143] (4) Build a CNN - BiLSTM - SAM fault diagnosis model, optimize the hyperparameters of the model using RIME, then use the processed dataset to train the model to obtain a RIME - CNN - BiLSTM - SAM fault diagnosis model, and diagnose the on - load tap changer (OLTC) of the converter transformer.

[0144] Since the on - load tap changer of the UHV converter transformer is in a harsh in - service environment and has a complex structure, its vibration signal has strong nonlinear and multi - time - scale characteristics. Therefore, the present invention proposes a UHV converter transformer OLTC fault diagnosis model of RIME - CNN - BiLSTM - SAM that fuses multi - feature information. The basic structure of the model is as Figure 1 shown. First, input the fused and dimension - reduced feature vectors into the model. The CNN captures the local features of the input data through the convolutional layer, and reduces the spatial dimension of the feature map through the pooling layer while retaining the key features. Second, use the BiLSTM structure to extract the temporal features hidden in the spatial feature information. Based on its bidirectionality, the BiLSTM layer can capture the long - term dependencies of the CNN feature map through the forward and backward propagation layers, and input the extracted spatial and temporal feature information into the next layer to establish the logical relationship between the input parameters and the output parameters. Then, to enhance the fault feature recognition ability, introduce the self - attention mechanism SAM to define the importance of features, thereby identifying the long - term dependencies between input elements. Finally, input it into the fully - connected layer for weight allocation, and improve the feature extraction effect by automatically adjusting the weights, so that the fault diagnosis model can pay more attention to the key features and output the fault diagnosis result.

[0145] As Figure 2 shown, to improve the diagnostic performance of the model, use the RIME algorithm to optimize the three hyperparameters of the model: the regularization coefficient, the learning rate, and the number of hidden neurons. Conduct extensive parameter search in the initial stage through the soft rime search strategy, and combine the hard rime piercing mechanism for precise adjustment in the later stage to determine a high - performance hyperparameter combination. Use the training set to train the model, and use the cross - entropy loss function to measure the difference between the model prediction result and the true label until the training is completed.

[0146] All kinds of parameter settings are shown in Table 4.

[0147] Table 4 Experimental parameters

[0148] Parameter Value RIME population size 10 RIME optimization dimension 3 RIME iteration times 20 CNN convolutional layer convolution kernel 32 CNN activation layer convolution kernel 64 CNN activation function ReLU BiLSTM activation function Sigmod Attention mechanism activation function Softmax Batch size 32 Dropout 0.3

[0149] The RIME algorithm simulates the growth processes of soft rime and hard rime of rime ice, and realizes position update through mathematical modeling of the soft rime search strategy, hard rime piercing mechanism, and greedy selection mechanism to solve the problem to be optimized. This algorithm has the advantages of fast optimization speed and high solution accuracy. By actively changing the search position during the update process, the global optimization ability of the algorithm and the ability to jump out of the local optimum are improved. The rime optimization algorithm is divided into the following four stages:

[0150] (1) Initialization. In the d-dimensional search space, randomly generate the initial positions of a population of rime particles with a population size of R, described as:

[0151] x ij = Lb ij + r(Ub ij - Lb ij )

[0152] In the formula, x ij represents the position of the i-th rime particle in the j-th dimensional space; Ub ij , Lb ij represent the upper and lower limits of the space respectively; r is a random number between (0, 1); i = 1, 2,..., R; j = 1, 2,..., d.

[0153] (2) Soft rime search strategy. This strategy can quickly complete the global search during the early iteration process without falling into the local optimum. The position update of the soft rime particle is as follows:

[0154]

[0155] In the formula, represents the updated position of the soft rime particle; R best,j is the best position of the rime particle in the j-th dimensional space; r1 is a random number between (-1, 1); θ is the growth angle of the soft rime; β represents the environmental factor; h represents the adhesion degree, and its value is a random number between (0, 1); r2 is a random number between (0, 1); E is the adhesion coefficient. The expressions of θ, β, and E are as follows:

[0156]

[0157] In the formula, t is the current iteration number; T is the maximum iteration number; w is the control coefficient, and in the present invention, it is taken as 5.

[0158] (3) Hard rime piercing mechanism. Realize the position update of the piercing rime particles, jump out of the local extreme value optimization, and strengthen the algorithm convergence. The extreme value is described as follows:

[0159]

[0160] In the formula, is the updated position of the hard rime particle; F normr (S i ) is used to normalize the current particle fitness value; r3 is a random number between (-1, 1); S i is the rime crystal.

[0161] (4) Greedy selection mechanism. Compare the fitness of the updated rime particle with that of the rime particle before update. If the result is better than before, replace it with the updated position of the rime particle; otherwise, retain the original position of the rime particle.

[0162] To further illustrate the superiority of the RIME-CNN-BiLSTM-SAM model, the test feature dataset is respectively input into the trained CNN-LSTM, CNN-LSTM-SAM, CNN-GRU, CNN-GRU-SAM, CNN-BiLSTM and the model of this embodiment for diagnosis. The comparison results of different models based on the same OLTC vibration signal test feature dataset are shown in Table 5.

[0163] As can be seen from Table 5, all 6 fault diagnosis models can identify various states of the OLTC, but the average fault diagnosis accuracy is different, which are 94.79%, 95.49%, 95.83%, 96.53%, 97.57% and 99.65% respectively. The RIME-CNN-BiLSTM-SAM model constructed by the present invention has only 1 sample misjudgment in fault diagnosis, and the fault recognition accuracy for the OLTC is significantly higher than that of the CNN-LSTM, CNN-LSTM-SAM, CNN-GRU, CNN-GRU-SAM and CNN-BiLSTM models, showing good fault diagnosis effect. Therefore, the fault diagnosis method integrating multi-feature information and the RIME-CNN-BiLSTM-SAM model proposed by the present invention can complete the identification of typical faults of the UHV converter transformer OLTC.

[0164] Table 5 Comparison results of OLTC fault diagnosis methods

[0165]

[0166] The specific embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the purpose of the present invention.

Claims

1. A UHV converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM, characterized in that It includes the following steps: Conduct vibration simulation experiments on OLTC under different working conditions to construct a vibration signal dataset; Use REMD to decompose the OLTC vibration signal into multiple IMFs, and preferentially select the IMFs containing rich fault feature information based on the correlation coefficient; Calculate the multi-scale permutation entropy of the selected IMFs, combine time / frequency statistical analysis and multi-scale permutation entropy to extract multi-feature information of the vibration signal, and introduce principal component analysis to carry out feature fusion and dimensionality reduction of the OLTC vibration information; Build a CNN-BiLSTM-SAM fault diagnosis model, use RIME to optimize the hyperparameters of the model, then use the processed dataset to train the model to obtain a RIME-CNN-BiLSTM-SAM fault diagnosis model, and conduct fault diagnosis on the OLTC of the converter transformer.

2. The UHV converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM according to claim 1, characterized in that, The different working conditions of the OLTC include normal conditions, drive shaft jamming faults, contact ablation faults, and contact wear faults.

3. The UHVDC converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM according to claim 1, characterized in that The process of REMD decomposing OLTC is as follows: Let E i (r) be the i-th r-component that meets the conditions obtained by EMD decomposition, x(n) be the target signal to be decomposed, and k be the number of implementation times. The specific steps are as follows: (1) Initialize by importing the original data, h k [n] = G i (n); (2) From h k Find the mean envelope line m k [n], the calculation formula is as follows: where g max [n] represents the upper envelope, and g min [n] represents the lower envelope; (3) Screen to obtain the signal h1[n]: h1[n] = h0[n] - m0[n] (4) Judge whether the extracted modal component r meets the conditions; if it does not meet the conditions, use h1[n] as the new initial signal, and repeat steps (2) to (4) until the first r component is obtained; The maximum number of iterations D has not been reached max When: r1 = h k-2 [n] Reach the maximum number of iterations D max At this time: r1 = h k [n] where h k [n] represents the signal after screening and iteration; (5) Obtain a new initial signal G i (n), and determine whether G i (n) is monotonic or constant. If not, repeat steps (1) to (5) until the initial signal cannot be decomposed. At this time, k components are obtained, and the final residual component is expressed as R(n). Then the initial signal x(n) is expressed as: In the formula, I is the total number of modal components.

4. The UHV converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM according to claim 1, wherein, The process of preferentially selecting IMFs by the correlation coefficient is as follows: (1) The correlation coefficient ρ between the original vibration signal x(t) and its IMFs, h1(t), h2(t), …, h n (t) after being decomposed by REMD is calculated as follows: i The calculation formula is: where cov() is the covariance; σ is the standard deviation; h i (t) represents the signal after screening and iteration; (2) Set the threshold Th as the standard deviation of the correlation coefficient, that is: If ρ i > Th, then retain the i-th IMF; otherwise, reject the i-th IMF.

5. The fault diagnosis method for UHV converter transformer OLTC integrating multi - feature information and RIME - CNN - BiLSTM - SAM according to claim 1, characterized in that, The steps of data dimensionality reduction by the principal component analysis are as follows: 1) Assume that the number of samples is m, the feature vector corresponding to each sample is n, and the feature vector matrix H can be expressed as an n×m-order matrix; Normalize the data to generate the standard matrix H * : In the formula, represents the mean of each dimension, and S j represents the variance of each dimension; 2) Construct the feature covariance matrix ∑: 3) Obtain the eigenvalues of ∑ and arrange them in descending order, select the feature vectors corresponding to the first l eigenvalues to form a new vector matrix U, and calculate the principal component Z: Z = U T X * 。 6. The UHVDC converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM according to claim 1, characterized in that The CNN-BiLSTM-SAM fault diagnosis model includes a convolutional layer, a pooling layer, a BiLSTM layer, a self-attention layer, and a fully connected layer; first, the convolutional layer captures the local features of the input data, and the pooling layer reduces the spatial dimension of the feature map and retains the key features; then, the BiLSTM structure is used to extract the temporal features hidden in the spatial feature information, and the extracted spatial feature and temporal feature information are input to the next layer to establish the logical relationship between the input parameters and the output parameters; then, the self-attention layer is used to define the importance of the features, so as to identify the long-term dependence relationship between the input elements. Finally, it is input to the fully connected layer for weight allocation.

7. The UHV converter transformer OLTC fault diagnosis method integrating multi-feature information and RIME-CNN-BiLSTM-SAM according to claim 1, characterized in that The RIME optimizes the model parameters including the regularization coefficient, learning rate, and the number of hidden neurons, conducts extensive parameter searches at the initial stage through the soft rime search strategy, and combines the hard rime piercing mechanism for precise adjustment in the later stage to determine a high-performance hyperparameter combination.

Images