Transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation
Through the method of multi-source signal fusion and fast spectrum related to the method, the problem of incomplete description of a single sensor data is solved, the accuracy of transformer fault diagnosis is improved, and the subjectivity of manual experience is reduced.
Patent Information
- Application Number
- CN202510098891.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-22
- Publication Date
- 2025-06-03
AI Technical Summary
The existing transformer fault diagnosis methods rely on single sensor data, resulting in incomplete description of fault types, and the selection of feature quantities depends on manual experience, and the diagnostic accuracy rate is low.
The method of multi-source signal fusion and fast spectrum correlation is adopted, and vibration signals at different positions on the transformer box surface are synchronized through sensor arrays, data-level fusion is used to obtain the fusion signal, and rapid spectrum correlation calculation is carried out, rapid spectrum correlation image sample set is constructed, and the mobileNetV3 network is sent for transfer learning model training to achieve fault diagnosis.
It effectively overcomes the problem of incomplete description of single sensor data, improves the accuracy of fault diagnosis, reduces the subjectivity of manual experience, and can better reflect the overall mechanical state of the transformer.
Smart Images
Figure CN120086790A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of fault diagnosis, specifically a transformer fault diagnosis method based on multi-source signal fusion and fast spectral correlation. Background Art
[0002] As one of the most important devices in the power system, faults in transformers will directly affect the safe and stable operation of the power grid. Therefore, it is very necessary to timely master the operating state of transformers and prevent accidents from occurring.
[0003] Due to the influence of external environment or its own factors, mechanical structure faults often occur in the windings and iron cores inside the transformer, directly affecting the safe and reliable operation of the transformer. In this regard, researchers have proposed many methods for diagnosing transformer mechanical faults, including frequency response method, low-voltage pulse method, vibration signal analysis method, sweep impedance method, etc. Among the above methods, the vibration signal analysis method is applicable to the fault diagnosis of both dry-type transformers and oil-immersed transformers. Due to its advantages such as no direct electrical connection with the transformer under test and easy implementation of on-line or live detection, it has become a current research hotspot.
[0004] With the development of machine learning theory, researchers use machine learning algorithms to learn the statistical features of fault original signals, and then complete fault diagnosis. Common methods include support vector machine, random forest, K-means clustering analysis, etc. However, such machine learning-based fault diagnosis methods rely on manual extraction of features based on manual experience, with great subjectivity and uncertainty, and the fault information reflected by the feature quantities is not comprehensive, resulting in low diagnostic accuracy.
[0005] Currently, due to its powerful feature self-adaptive extraction ability, convolutional neural network has shown great superiority in tasks such as image classification, image segmentation, and target detection, and has gradually been applied to the field of fault recognition and diagnosis. Common ways to construct signal images include Gram angle field, fast Fourier transform, continuous wavelet transform, recursive analysis, etc. However, these classic image coding methods generally have high parameter sensitivity and poor ability to suppress noise and interference, resulting in a decline in image quality.
[0006] In addition, transformer vibration is due to the coupling effect of multiple factors such as the internal iron core and windings. The vibration signal on the surface of its box has complex vibration characteristics of non-linearity, non-stationarity, and wide frequency band. The signals measured by sensors at different positions contain large information differences. Most current diagnostic methods are based on the signals measured by a single sensor for fault diagnosis, lacking an overall description of the mechanical state of the transformer, resulting in its robustness and generalization performance not meeting the actual engineering requirements.
[0007] In view of the above problems, the present invention proposes an on-line transformer fault diagnosis method based on multi-source signal fusion and fast spectral correlation, which overcomes the problems that the data measured by a single sensor is not comprehensive in describing the fault type and the selection of characteristic quantities depends on manual experience, and achieves good diagnostic results. Summary of the Invention
[0008] The transformer fault diagnosis method based on multi-source signal fusion and fast spectral correlation can effectively solve the problems that the data measured by a single sensor in transformer fault diagnosis is not comprehensive in describing the fault type and the selection of characteristic quantities depends on manual experience, and improve the accuracy of fault diagnosis.
[0009] The present invention is implemented by the following technical solutions: The transformer fault diagnosis method based on multi-source signal fusion and fast spectral correlation includes the following steps:
[0010] S01 Use a sensor array to synchronously collect vibration signals at different positions on the surface of the transformer tank;
[0011] S02 Use the correlation function weighting method to perform data-level fusion on multi-sensor signals to obtain the fused signal x[n];
[0012] S03 Perform fast spectral correlation calculation on the fused signal x[n] to obtain a fast spectral correlation image, and construct an image sample set;
[0013] S04 Send the image into the MobileNetV3 network for transfer learning model training to obtain a fine-tuned neural network model;
[0014] S05 Import real-time data into the model to give the fault diagnosis result.
[0015] Further, the specific steps of step S01 are as follows:
[0016] Use a sensor array to synchronously collect vibration signals at different positions on the surface of the transformer tank.
[0017] The collected transformer sample signals include the following four types: three fault states of winding looseness, core looseness, and winding deformation and the normal state. The sensor array consists of three sensors, which are respectively set on the upper surface of the tank, the lower surface of the tank, and the front of the tank.
[0018] Further, the details of step S02 are as follows:
[0019] Use the correlation function weighting method to perform data-level fusion on multi-sensor signals to obtain the fused signal.
[0020] The vibration signals synchronously collected by the sensor array are expressed as x 1 [n], x 2 [n], x 3[n],…x k [n], the cross - correlation operation is performed on any two of these signals xi and xj to obtain the cross - correlation function:
[0021]
[0022] where n is the discrete - time variable obtained at the sampling frequency; m is the time offset for the cross - correlation calculation of the signals. In this paper, multi - source signals are synchronously collected, m = 0, and the similarity of the two signals in the state of no time delay is calculated; N is the total number of sample points.
[0023] From the energy calculation formula of discrete signals:
[0024]
[0025] where E ij is the energy of the signals obtained by performing cross - correlation operations pairwise on each signal.
[0026] Then the expression for the total correlation energy of the signal collected by the ith sensor and the signals collected by other sensors is:
[0027]
[0028] According to the fact that the weight wi is proportional to the energy of the correlation function, there is:
[0029] w 1 :w 2 :...w n =E 1 :E 2 :...E n
[0030] Also:
[0031] w 1 +w 2 +w 3 +...+w n =1
[0032] From the above two equations, the weights of each signal can be calculated, and the fusion result is:
[0033] X = w 1 x 1 +w 2 x 2 +w 3 x 3+ ...+w i x i
[0034] Furthermore, the specific steps of S03 are as follows:
[0035] Perform fast spectral correlation calculation on the fused signal to construct a fast spectral correlation image sample set;
[0036] Specifically, the STFT of the signal x[n] is:
[0037]
[0038] where N w is the window width; R is the moving step; w[n] is the window function; f k = kΔf (k = 0, 1, …, N w - 1) is the discrete frequency; Δf = Fs / N w is the frequency resolution.
[0039] The phase correction of the STFT is as follows:
[0040]
[0041] where X w (i, f k ) is the complex envelope of the signal at time iR / F S centered at f k with a bandwidth of Δf, and ∣X w (i, f k )∣ 2 represents the energy flow within the frequency band; L is the signal length.
[0042] Then the cyclic spectrum based on the STFT is:
[0043]
[0044] where the superscript * represents the conjugate complex number.
[0045] Assume that f = f k = kΔf and α = pΔf + δ, then f - α = f k - α ≈ f k-p and α ≈ pΔf. Thus, we get:
[0046]
[0047] where N 0 is the symmetric window center.
[0048] Substitute the expressions of X w (i, f k ) and X w (i, f k - α) into the cyclic spectrum formula, and we can obtain the definition of the scanning spectrum:
[0049]
[0050] Wherein, K = (L - N W + R) / R.
[0051] Use phase calibration scanning summation to calculate fast spectral correlation and obtain the fast spectral correlation spectrum:
[0052]
[0053] Furthermore, the specific steps of S04 are as follows:
[0054] Add labels to the image sample data and divide it into a training set and a test set;
[0055] Freeze the weight parameters of the Bneck layer of the pre-trained MobileNetV3 and change the output nodes to adapt to the classification task;
[0056] Use the test set data to test the model, obtain the test results and save the model.
[0057] Furthermore, the specific steps of S05 are as follows:
[0058] Use the weight mean values of each sensor obtained in the offline stage as statistical features for data fusion of the signals collected by the on-site monitoring system;
[0059] Calculate the fast spectral correlation of the fused signal to obtain the fast spectral correlation image sample;
[0060] Import the spectral correlation image into the optimal model, and finally give the fault diagnosis result through feature extraction and classification recognition. Brief Description of the Drawings
[0061] Figure 1 It is a flowchart of the method involved in the present invention;
[0062] Figure 2 It is the sensor weight distribution under the winding loosening state in the embodiment;
[0063] Figure 3 It is the comparison of the envelope spectra of the single-sensor signal and the weighted fusion signal of the correlation function in the embodiment;
[0064] Figure 4 It is the comparison of the envelope spectra of the fused signals of different fusion methods in the embodiment;
[0065] Figure 5 It is the fast spectral correlation images of different faults in the embodiment;
[0066] Figure 6 It is the energy feature distribution of the fast spectral correlation in the embodiment;
[0067] Figure 7 It is the accuracy rate and loss curve of the model training in the embodiment;
[0068] Figure 8 Confusion matrix for fault diagnosis in the embodiment
[0069] Figure 9 Comparison of diagnostic accuracy rate curves of different sensor signals in the embodiment
[0070] Figure 10 Comparison of accuracy rate curves of different feature images in the embodiment Detailed implementation manners
[0071] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention
[0072] The first step is to collect the vibration signals of the transformer box body. The specific steps are as follows
[0073] S101: Synchronously collect the vibration signals at different positions on the surface of the transformer box body by using a sensor array
[0074] Specifically, the collected transformer sample signals include the following four types: three fault states of winding looseness, core looseness, and winding deformation, and the normal state. The sensor array consists of three sensors, which are respectively arranged on the upper surface of the box body, the lower surface of the box body, and the front surface of the box body. A total of 1600 groups of samples are collected, with 400 groups for each state
[0075] The second step is to perform data-level fusion on the signals of the three sensors by using the correlation function weighting method. The specific steps are as follows
[0076] S201: Specifically, the vibration signals synchronously collected by the sensor array are expressed as x 2 [n], x 3 [n]. For any two of the signals x i and x j , perform a cross-correlation operation to obtain the cross-correlation function: x 1 [n]
[0077]
[0078] Among them, n is the discrete time variable obtained at the sampling frequency; m is the time offset for the cross-correlation calculation of the signals. In this article, multi-source signals are synchronously collected, m = 0, and the similarity of the two signals in the state of no time delay is calculated; N is the total number of sample points
[0079] From the energy calculation formula of discrete signals
[0080]
[0081] Among them, E ijThe energy of the signal obtained by performing cross - correlation operations pairwise for each signal.
[0082] Then the expression for the total correlation energy of the signal collected by the i - th sensor and the signals collected by other sensors is:
[0083]
[0084] According to the weight w i being proportional to the energy of the correlation function, there is:
[0085] w 1 : w 2 : w 3 = E 1 : E 2 : E 3
[0086] Also:
[0087] w 1 + w 2 + w 3 = 1
[0088] From the above two equations, the weights of each signal can be calculated, and the fusion result is:
[0089] X = w 1 x 1 + w 2 x 2 + w 3 x 3
[0090] In the third step, perform fast spectral correlation calculation on the fused signal to obtain a fast spectral correlation image and construct an image sample set. The specific steps are as follows:
[0091] S301: Perform fast spectral correlation calculation on the fused signal to construct a fast spectral correlation image sample set;
[0092] Specifically, the STFT of the signal x[n] is:
[0093]
[0094] where N w is the window width; R is the moving step; w[n] is the window function; f k = kΔf (k = 0, 1,..., N w - 1) is the discrete frequency; Δf = Fs / N w is the frequency resolution.
[0095] Specifically, the phase correction of the STFT is as follows:
[0096]
[0097] Among them, X w (i, f k ) is the complex envelope of the signal at time iR / F with f as the center and Δf as the bandwidth, S at time f k as the center and Δf as the bandwidth,
[0098] |X w (i, f k )| 2 represents the energy flow within the frequency band; L is the signal length.
[0099] Then the cyclic spectrum based on STFT is:
[0100]
[0101] where the superscript * represents the conjugate complex number.
[0102] Assume that f = f k = kΔf and α = pΔf + δ, then f - α = f k - α ≈ f k-p and α ≈ pΔf. Thus, we get:
[0103]
[0104] where N 0 is the center of the symmetric window.
[0105] Substitute the expression of X w (i, f k ) and the expression of X w (i, f k - α) into the cyclic spectrum formula, and we can obtain the definition of the scanning spectrum:
[0106]
[0107] where K = (L - N W + R) / R.
[0108] Use phase calibration to scan and sum, calculate the fast spectral correlation, and obtain the fast spectral correlation spectrum:
[0109]
[0110] Step 4: Send the image into the MobileNetV3 network for transfer learning model training to obtain a fine-tuned neural network model. The specific steps are as follows:
[0111] S401: Add labels to the image sample data and divide it into a training set and a test set;
[0112] S402: Freeze the weight parameters of the Bneck layer of the pre-trained MobileNetV3 and change the output nodes to adapt to the classification task;
[0113] S403: Use the test set data to test the model, obtain the test results and save the model.
[0114] The fifth step is to import the real-time data into the model and give the fault diagnosis result. The specific steps for online transformer fault diagnosis are as follows:
[0115] S501: Use the weight means of each sensor obtained in the offline stage as statistical features for data fusion of the signals collected by the on-site monitoring system;
[0116] S502: Calculate the fast spectral coherence of the fused signal to obtain fast spectral coherence image samples;
[0117] S503: Import the spectral coherence image into the optimal model, and finally give the fault diagnosis result through feature extraction and classification recognition.
[0118] Example:
[0119] First, collect transformer fault vibration signal samples. The sensor array consists of three sensors, which are respectively set on the upper surface of the box body, the lower surface of the box body and the front of the box body. The transformer states include three fault states of winding looseness, core looseness, winding deformation and the normal state, with 400 groups of samples for each state.
[0120] Secondly, use the correlation function weighting method to perform data-level fusion on the multi-sensor signals to obtain the fused signal. Taking the winding looseness fault as an example, 400 sample data under 3 sensors are collected and preprocessed, and the correlation function weighting method is used for real-time fusion according to the correlation analysis of the data. As Figure 2 shown, it is the weight distribution of the sensor data at each measuring point. The weight range of the top measuring point A is 0.350 - 0.399, and the mean value is 0.373; the weight range of the bottom measuring point B is 0.320 - 0.379, and the mean value is 0.348; the weight range of the front measuring point C is 0.202 - 0.347, and the mean value is 0.273. The weights of the top and bottom measuring points A and B are generally higher than those of the front measuring point, because there are close mechanical connections between the top and bottom of the transformer and the windings and the iron core, following the principle of the shortest vibration transmission path. In order to receive the vibration information transmitted by the insulating oil, the setting of sensor C is also necessary. After calculation, the same rule is shown in other state samples.
[0121] To verify the effectiveness and superiority of the multi-sensor fusion, perform envelope spectrum analysis on the single-sensor signal and the fused signal. The results are as Figure 3As shown. It can be seen from the figure that the frequency peaks shown in the envelope spectrum of a single sensor signal are not complete, while the fused signal under the correlation function weighting method shows more complete and prominent frequency characteristic components in a wider frequency band. As Figure 3 (d) shows, there are obvious energy peaks in the envelope spectrum of the fused signal under the correlation function weighting method at the fundamental frequency f = 100 Hz and its harmonics 200 Hz, 300 Hz, 400 Hz, and 500 Hz. In contrast, Figure 3 (a) only has amplitude peaks of energy at 100 Hz, 200 Hz, and 300 Hz, and no frequency characteristic components are detected in the frequency band above 300 Hz; Figure 3 (b) only has prominent energy peaks at 300 Hz and 400 Hz; Figure 3 (c) only has an energy peak at 500 Hz, and the information loss is relatively serious. This shows that the description of the transformer operating state by a single sensor signal is not accurate enough, while the fused signal can not only inherit the original frequency characteristics, but also has more comprehensive frequency components, and can better reflect the overall vibration characteristics of the faulty transformer, providing a strong basis for improving the accuracy of fault diagnosis.
[0122] In addition, other fusion methods are compared and analyzed with the correlation function weighting method proposed in this paper, including the weighted average method, the entropy weight method, the least squares method, and the Kalman filtering method. The envelope spectra of different methods are as Figure 4 shown. It can be seen from the figure that for the fused signals obtained by using the weighted average method and the least squares method, their frequency components are almost submerged by noise during the fusion, lacking anti-interference performance. Although the fused signals obtained by the entropy weight method and the Kalman filtering method play a certain role in suppressing noise, the frequency information they reflect is not comprehensive and there is a certain amount of information loss. In contrast, Figure 3 (d) shows that the information contained in the fused signal under the method proposed in this paper is still the most complete, further verifying the information integrity and noise suppression of the fused signal obtained by the correlation function weighting method.
[0123] Again, the fast spectral correlation calculations are respectively carried out on the fused signals under different faults, the maximum cyclic frequency is set to 1000 Hz, and an image sample set is constructed, where winding looseness, core looseness, winding deformation and some images are as Figure 5 shown. As Figure 5 (a) and (c) show, there is different degrees of energy concentration at the fundamental frequency and multi-harmonic frequencies for winding looseness and winding deformation, containing less noise; Figure 5 (b) shows that there are two energy concentrations at the fundamental frequency and the third harmonic frequency; as Figure 5As shown in (d), there is no prominent energy concentration phenomenon under normal operating conditions. Thus, Fast-SC can effectively reveal transformer fault information and suppress the background noise of the non-periodic cyclic characteristics.
[0124] Based on the average energy at the vibration fundamental frequency and its 2 - 5 times frequency in the fast spectral correlation spectrum, the sample energy characteristic distribution diagram is drawn as the basis for the accuracy of the image classification task. The results are as Figure 6 shown. It can be seen from the figure that different from the fault state samples, the characteristic energy of the normal state samples is at a relatively low level; the characteristic energy distribution intervals of different fault state samples are different and can all be distinguished to a certain extent. Thus, the fast spectral correlation method proposed in this paper can effectively express transformer fault characteristics.
[0125] Then, the image is sent into the MobileNetV3 network for transfer learning model training. The training set and the test set are randomly divided at a ratio of 4∶1, and the number of classifications is 4. To avoid overfitting, a regularization method is introduced, Dropout is set to 0.15, the loss function uses the cross-entropy function, the activation function uses h-swish, the classifier uses the Softmax classifier, the maximum number of training epochs is set to 150, the batch size batchsize is 32, and the learning rate is 0.001. The accuracy and loss curves are obtained as Figure 7 shown. When the number of iterations reaches about 25 times, the training set loss value converges from the initial 1.395 to 0.04 and tends to be stable until it converges completely; when the number of iterations reaches about 40 times, the test set accuracy rapidly increases to more than 95% and steadily grows to more than 98% as the model iterates.
[0126] The confusion matrix is used to measure the classification effect of the model in this paper, as Figure 8 shown. The labels of the four types of winding looseness, core looseness, winding deformation and no-fault state are 1, 2, 3, and 4 respectively. It can be seen from the figure that the model accuracy reaches 98.75%. Especially, the recognition accuracy of the winding deformation and normal state samples reaches 100%. It can effectively identify and diagnose common mechanical faults of different degrees. Only two samples of winding looseness are misclassified as winding deformation because there is a slight similarity in the energy characteristics in the frequency band of 50Hz - 400Hz. Only two samples of slight core looseness are misclassified as the normal state because there is a little similarity in the energy characteristics of these two states above the 300Hz frequency band.
[0127] To verify the improvement of the diagnostic accuracy by multi-source signal fusion, the diagnostic effects of the models of single-sensor signals and signals weighted and fused using the correlation function are compared. The obtained precision curves are as Figure 9As shown in the figure, it can be seen that the accuracy curve of the fused signal converges the fastest and reaches the highest level, with an average increase of 10.52% compared to the other three single-sensor signals. Moreover, when the number of iterations reaches 100 or even as high as 150, there is still a certain degree of volatility in the accuracy of the single-sensor model, and the convergence level is relatively low. In summary, the signal fused using the correlation function weighting contains more fault feature information and achieves the best diagnostic effect.
[0128] To verify the superiority of the proposed Fast-SC, it is compared with different image construction methods, including Fast Fourier Transform (FFT), Discrete Wavelet Transform (DWT), and Gramian Angular Field (GAF). The accuracy curves are as Figure 10 shown. The training of DWT and FFT experiences significant fluctuations throughout the process because their parameter sensitivity leads to poor image quality. The training of GAF only reaches a better convergence effect after the 80th epoch, which is slower than Fast-SC, but its average accuracy in the first 80 epochs is still 7.34% lower than that of Fast-SC. Compared with the other three image construction methods, the average accuracy of the proposed Fast-SC in this paper has increased by 10.86%, showing significant superiority in the field of transformer fault diagnosis.
[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features. Therefore, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.
Claims
1. A transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation, characterized in that: include: S01 uses a sensor array to synchronously collect vibration signals at different locations on the transformer box surface; S02 uses a correlation function weighting method to perform data-level fusion on multi-sensor signals to obtain a fusion signal x[n]; S03 performs fast spectral correlation calculation on the fused signal x[n] to obtain a fast spectral correlation image and construct an image sample set; S04 sends the image to the MobileNetV3 network for transfer learning model training to obtain a fine-tuned neural network model; S05 imports real-time data into the model, gives fault diagnosis results, and realizes online fault diagnosis.
2. The transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation according to claim 1 is characterized in that: In step S01, the vibration signals at different positions on the surface of the transformer box include three fault states: loose winding, loose core, and deformed winding, and a normal state; the sensor array consists of three sensors, which are respectively arranged on the upper surface of the box, the lower surface of the box, and the front surface of the box.
3. The transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation according to claim 1 is characterized in that: In step S02, the vibration signals synchronously collected by the sensor array are represented as x1[n], x2[n], x3[n], ... x k [n], for any two signals x i and x j Perform cross-correlation operation to obtain the cross-correlation function: Among them, n is the discrete time variable obtained by the sampling frequency; m is the time offset of the signal for cross-correlation calculation, and N is the total number of sample points. The energy calculation formula of the discrete signal is: Among them, E ij The energy of the signal obtained by performing cross-correlation operation on each signal is expressed as follows: According to the weight w i Is proportional to the energy of the correlation function, then: <h2 style=";text-align:left;direction:ltr">w1:w2:...w<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> (E1:E2:...E)<h2 style=";text-align:left;direction:ltr"> n again: <h2 style=";text-align:left;direction:ltr">w1+w2+w3+...+w<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> =1 The weights of each signal can be calculated by the above two formulas, and the fusion result is: X=w1x1+w2x2+w3x 3+ ...+w i x i 。 4. The transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation according to claim 1 is characterized in that: In step S03, the STFT of the signal x[n] is: Among them, N w is the window width; R is the moving step length; w[n] is the window function; f k =kΔf(k=0,1,…,N w -1) is the discrete frequency; Δf=Fs / N w For the frequency resolution, the phase correction of STFT is as follows: Among them, X w (i,f k ) is the signal at iR / F S Time f k The complex envelope with Δf as the bandwidth and ∣X w (i,f k )∣ 2 represents the energy flow within the frequency band; L is the signal length, then the cyclic spectrum based on STFT is: The superscript * indicates a conjugate complex number. Assume that f = f k =kΔf and α=pΔf+δ, then f-α=f k -α≈f k-p And α≈pΔf, thus we get: Among them, N0 is the center of the symmetric window, and X w (i,f k ) expression and X w (i,f k -α) expression into the cyclic spectrum, we can get the definition of the scanning spectrum: Where K = (LN W +R) / R, use phase calibration scan summation to calculate the fast spectral correlation and obtain the fast spectral correlation spectrum:
5. The transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation according to claim 1 is characterized in that: Step S4 is specifically as follows: label the image sample data and divide it into a training set and a test set, freeze the weight parameters of the Bneck layer of the pre-trained MobileNetV3, change the output node to adapt to the classification task, test the model using the test set data, obtain the test results and save the model.
6. The transformer fault diagnosis method based on multi-source signal fusion and fast spectrum correlation according to claim 1 is characterized in that: Step S5 is specifically as follows: using the weighted mean of each sensor obtained in the offline stage as a statistical feature for data fusion of the signals collected by the on-site monitoring system, calculating the fast spectral correlation of the fused signal, obtaining a fast spectral correlation image sample, importing the spectral correlation image into the optimal model, and finally giving the fault diagnosis result through feature extraction and classification recognition.