Sage Journals: Discover world-class research

Abstract

The prominent vibration characteristics of amorphous alloy transformer (AMT) make it possible to apply the vibration method for real-time fault monitoring of AMT. Therefore, in order to solve the AMT vibration monitoring problem and enhance the diagnostic efficiency, this study proposes an AMT fault diagnosis model based on particle swarm optimization (PSO) to optimize the parameters of wavelet packet transform (WPT) and support vector machine (SVM).The optimal vibration signal acquisition point is determined by finite element analysis to ensure high signal quality. The PSO algorithm is used to optimize the number of WPT decomposition layers, wavelet basis functions, SVM kernel parameters g and penalty parameters c to enhance the accuracy of feature extraction and classification. Additionally, principal component analysis (PCA) reduces the dimensionality of the redundant frequency band energy after WPT feature extraction, minimizing data redundancy. Overall, the full-process optimization significantly improves AMT fault diagnosis efficiency compared with single-aspect optimization.

Keywords

Amorphous alloy transformer vibration signal fault diagnosis particle swarm algorithm support vector machine principal component analysis

Introduction

Transformer is an essential component in the process of power transmission. With the continuous advancement of technology, to accommodate the requirements of green development, the power grid has been vigorously promoting new types of transformers based on amorphous alloy transformers in recent years. Amorphous alloy transformers possess advantages such as low no-load loss and high operational efficiency. In comparison with silicon steel transformers, their no-load loss is one fourth of that of silicon steel transformers, and their no-load current is one fifth of that of silicon steel transformers. This leads to amorphous alloy transformers gradually substituting the original silicon steel transformers distributed at the power receiving end of users and approaching residential living areas. Hence, how to conduct real-time and rapid diagnoses of this new type of amorphous alloy transformers effectively holds significant importance for guaranteeing the power usage safety of the residents.

There are many existing transformer status detection methods, such as short-circuit impedance method,^1,2 dielectric response method,^3,4 frequency response analysis method,^5,6 oil chromatography analysis method,^7,8 etc. The first three methods need to be run offline, and the latter has low recognition of mechanical structure faults. Therefore, some scholars have proposed to analyze the operating status of the transformer through vibration signals. The principle of this method is to collect the vibration signal of the overall mechanical structure of the transformer through sensors to realize real-time monitoring of the mechanical structure and the electrical status of the transformer. It indicates that this method can be widely used in various types of transformers with a fault monitoring accuracy of up to 90%^9,10 and it has a good recognition of mechanical faults.

At present, fault identification based on transformer vibration signals mainly consists of three steps: signal decomposition, feature extraction, and pattern recognition. In terms of feature signal extraction, Xiang Li et al.¹¹ proposed a cross-modal feature alignment method and combined it with a deep adversarial neural network to realize fault diagnosis knowledge transfer. Kezhan et al¹² used sine cosine optimization algorithm to optimize VMD parameters and improved the recognition rate of transformer vibration signals. Xin Zhang et al¹³ proposed a multi model fusion transformer diagnosis method based on TimesNet and Informer, which combines the local features of convolution and the global correlation of attention mechanism modules for feature aggregation, enabling the model to obtain more time series information. Haikun Shang¹⁴ incorporated the approximate entropy (ApEn) of symplectic geometric mode decomposition (SGMD) into an optimized bidirectional long short-term memory (BILSTM) neural network to effectively extract the main features of partial discharge (PD), and compared it with FFT, EMD, and VMD, demonstrating the superiority of this method. For signal decomposition methods, parameter extraction often relies on empirical settings, which can reduce the signal’s representational ability and result in random decomposition effects.

Fault diagnosis is based on feature differences.¹⁵ Based on feature extraction, classification algorithms are used to classify and predict data. Wei Zhang et al¹⁶ proposed a Swin Transformer-based lithium-ion battery SOC prediction method, which uses the Swin Transformer network to process the learned features to achieve accurate prediction of the SOC of lithium-ion batteries. Huimin Zhao et al^17,18 proposed an image recognition model based on Shuffle Net-CA-SSD to improve the early defect detection capability of turbine blades. On the other hand, they developed the SGNSGA-DFBLS method using multi-objective evolutionary algorithms to enhance prediction performance and improve the quality of search and resolution through adaptive mutation. Wenlong Liao et al¹⁹ developed a GCN-based transformer fault diagnosis method that uses an adjacency matrix for sample similarity, graph convolutional layers for feature extraction, and optimized training to enhance diagnostic accuracy. In the past, fault diagnosis of transformers often only optimized one aspect of feature extraction or classification prediction, while ignoring that transformer fault diagnosis is a systematic problem.^20–23

In summary, this article proposes an AMT fault diagnosis model based on PSO algorithm to optimize the entire WPT-SVM process. In the signal preprocessing stage, electromagnetic structural field coupling is adopted, and the maximum total deformation of AMT at 200 Hz is used as the basis for selecting measurement points to overcome the dependence of vibration method on feature signal selection. In the feature extraction stage, energy entropy is used as the fitness function, and PSO algorithm is used to optimize the wavelet parameters [N, Wname], thereby achieving adaptive selection of wavelet parameters for each signal. For the redundant frequency band energy after feature extraction, PCA is used to reduce the dimensionality and minimize the problem of inaccurate classification caused by complex data. Finally, in the classification prediction stage, with the highest accuracy as the optimization goal, SVM penalty parameters and kernel parameters are optimized to improve the diagnostic efficiency of the model.

AMT vibration signal acquisition and preprocessing analysis

AMT vibration test platform

The multi-channel amorphous alloy transformer vibration test platform is shown in Figure 1. The SCBH15-200/10 amorphous alloy transformer is used for vibration signal analysis and data collection. Voltage is applied to the winding of the amorphous alloy transformer through a three-phase voltage regulator, and the multimeter displays accurate readings. The piezoelectric acceleration sensor collects signals from the surface of the transformer clamp and transmits the vibration signals to the PC through the data acquisition card.

Figure 1.

AMT vibration signal testing platform.

The main equipment used in this experiment is listed in Table 1

Table 1.

Main instruments and equipment for the experiment.

Serial number	Device name	Device model
1	Vibration Sensor	AD-1000T
2	Data Acquisition Card	MPS-14801
3	Three-phase voltage regulator	TSGC three-phase voltage regulator
4	Multimeter	Fluke F18B+

This experiment uses an amorphous alloy vibration test platform to simulate the vibration signals of the transformer under three kinds of fault conditions, no-load, short circuit, and core loosening. The sampling frequency is set to 64 kHz and the number of samples is 6400. The low-voltage side is loaded with 100% rated voltage for no-load testing, and the high-voltage side is loaded with 100% short-circuit current for short-circuit testing. In the real online monitoring scenario, the looseness of the core is set by adjusting the transformer clamps and fastening bolts to change the clamping force of the cores, and the loosening signal of the cores is measured during the no-load testing.

Selection of AMT measuring points

When selecting measurement points for transformer vibration signals, it is often difficult to find the position that best represents its vibration state.

Therefore, based on the conclusion that the maximum amplitude of the AMT core occurs at 200 Hz,²⁴ this paper adopts the finite element analysis method of electromagnetic structural field coupling. Firstly, as shown in Figure 2, the volume force distribution of the amorphous alloy transformer core is obtained in the electromagnetic field analysis (when unloaded, the core serves as the main vibration source, and its volume force is approximately equal to the magnetostrictive force). Next, the volumetric force is used as the load and transmitted to the harmonic response analysis to obtain the total deformation distribution of the amorphous alloy transformer under electromagnetic vibration at 200 Hz, as shown in Figure 3(a). Finally, using the total deformation distribution at 200 Hz as the basis for selecting measurement points, the experimental measurement points shown in Figure 3(b) were chosen.

Figure 2.

Volume force of amorphous alloy core.

Figure 3.

The selection basis of measuring points.

From the total deformation distribution diagram in Figure 3(a). The vibration of AMT is transmitted from the core to the surface of the transformer fixture through the transformer connection structure from inside to outside. Therefore, according to the total deformation distribution of the transformer, the 7 suitable measuring points shown in Figure 3(b) were finally selected to collect the vibration signal of the AMT.

Characteristic analysis of AMT fault signal

The main vibration signals of the transformer during operation come from the cores and windings.^25,26 The core vibration is caused by the magnetostrictive effect of the core material in the alternating magnetic field; the winding vibration is caused by the alternating Ampere force generated by the winding current in the leakage magnetic field. According to the magnetostrictive characteristics of ferromagnetic materials in the alternating magnetic field and the force characteristics of the winding current in the leakage magnetic field, it can be deduced that the fundamental frequency amplitudes of the transformer core and winding vibration are linearly related to the square of the power supply voltage and current amplitude, respectively, and the fundamental frequency of the transformer vibration signal is twice the power supply frequency.²⁷

The magnetostrictive properties of the AMT core are much greater than those of ordinary silicon steel cores,²⁸ so the core amplitude is much larger than that of ordinary transformers, which leads to a larger vibration amplitude of the entire transformer.

After the sensors were arranged on the surface of the AMT, Spectrum signals of three kinds of conditions are shown in Figure 4.

Figure 4.

Spectrum signals of three kinds of conditions.

By observing the fault signal, it can be discovered that the vibration signal of AMT is mostly concentrated within the range of 0 to 1000 Hz. As indicated in Figure 4(a) and Figure 4(c), the signal frequencies of no-load and loose-core conditions are mainly concentrated between 100 Hz and 300 Hz. It can be seen from Figure 4(b) that the frequency of the short-circuit signal is mainly concentrated at 100 Hz. However, since the signal frequency collected by this platform is 64 kHz, there exists the issue of an excessively high sampling frequency. If the sampling frequency is too large, after WPT feature extraction, the energy will be concentrated in one frequency band, making it impossible to extract effective features. Hence, the resampling frequency selected in this paper is 1600 Hz. According to the Nyquist sampling theorem, the analog signal is within the range of 0 to 800 Hz, which is consistent with results of the expectation.

Feature extraction and optimization of fault signals

Wavelet packet transform and feature extraction principle

WPT is a further extension of wavelet analysis. It divides the frequency band into multiple levels and divides each decomposed frequency band into half again, so as to achieve a more detailed analysis of the signal. A small time window is used for high-frequency signal analysis, and a large time window is used for low-frequency signal analysis.²⁹

The mathematical expression of wavelet packet decomposition is shown in formula (1)

{\begin{cases} C_{ξ, η, 2 f} = \frac{1}{\sqrt{2}} \sum_{k} h_{k - 2 η} C_{ξ + 1, k, f} \\ C_{ξ, η, 2 f + 1} = \frac{1}{\sqrt{2}} \sum_{k} g_{k - 2 η} C_{ξ + 1, k, f} \end{cases}

(1)

C_ξ,η,2f and C_ξ,η,2f+1 are the wavelet decomposition results of the next layer; C_ξ+1,k,f is the decomposition result of the upper layer; ξ is the scale coefficient; η is the position coefficient; f is the sampling frequency; h and g are orthogonal conjugate low-pass and high-pass filters.

The energy occupied by each node of the reconstructed signal can be calculated using formula (2)

E_{i, m} = \sum_{n}^{N} | x_{m n}^{2} |

(2)

Use the energy of each frequency band to construct the corresponding feature vector V

V = [E_{i, 0}, E_{i, 1}, \dots, E_{i, n}]

(3)

Then transform the feature vector so that each element in it is the ratio of the energy of each frequency band to the sum of the energies, and obtain V’ as shown in formula (4)

V^{'} = [E_{i, 0} / E, E_{i, 1} / E, \dots, E_{i, n} / E]

(4)

PSO parameter optimization for WPT

In mechanical condition monitoring, energy entropy reflects the distribution of signal energy in frequency band or time window. The vibration signal of normal equipment is stable, energy is concentrated and entropy is low. However, faults such as bearing wear, gear fracture or loose iron core will lead to changes in signal characteristics, increase frequency components and amplitude fluctuations, make energy distribution more dispersed to increase energy entropy.

In view of taking the energy entropy of the signal as the objective function and considering that the signal is not strictly continuous and differentiable in the actual acquisition process, choosing PSO to optimize WPT wavelet coefficients has inherent advantages.

Compared with the traditional gradient descent method or Newton method, PSO does not rely on the gradient information of the objective function, which is very suitable for dealing with the nonlinear problem in the selection of WPT parameters. The optimization objective function is shown in formula (5)

{\begin{cases} p_{i} = \frac{E_{i}}{\sum_{j = 1}^{N} E_{j}} \\ F i t n e s s = - \sum_{i = 1}^{N} p_{i} \log_{2} (p_{i}) \end{cases}

(5)

Based on the energy entropy theory in mechanical fault diagnosis mentioned above, the larger the energy entropy, the more obvious the fault characteristics.³⁰ This article proposes using the energy entropy of vibration signals as the objective function of PSO algorithm, aiming to improve the accuracy of fault signal feature extraction by optimizing the key parameters of wavelet packet decomposition [N, “Wname”], namely, the decomposition level and wavelet basis function. This method not only helps identify early faults in equipment, but also provides quantitative analysis basis for fault types,³¹ thereby achieving effective monitoring of the health status of mechanical equipment.

To show the optimization effect, taking the loose-core state as an example, the energy entropy value of the [3, “db2”] wavelet packet parameter selected by experience before improvement is compared with the fitness function value after PSO improvement, as shown in Figure 5. The larger the fitness function value, the more it can show the feature extraction effect after PSO optimization.

Figure 5.

Fitness value before and after improvement.

It can be seen from Figure 5 that, compared with before PSO optimization, the fitness values have been significantly improved and the fitness values have become larger, indicating that the optimization effect of PSO on WPT feature extraction is universal and a suitable optimal solution can be found. It also proves that PSO has a significant optimization effect on WPT feature extraction and can be used as one of the optimization methods for feature extraction, and it is feasible.

Meanwhile, according to the comparison of wavelet packet parameters before and after optimization in Table 2, the decomposition layer number of wavelet packet after optimization of transformer vibration signal is generally greater than 3 layers, and the selection of wavelet basis function is not fixed according to the difference of selected signals. It shows that the empirical wavelet parameter selection is not accurate, and the characteristics of transformer vibration signal cannot be extracted effectively.

Table 2.

Comparison of parameter optimization.

Measurement point number	Optimized parameters	Adaptation value before optimization	Optimized fitness value
1	[5,“rbio4.4”]	2.51	4.17
2	[5,“sym19”]	2.23	4.04
3	[5,“sym14”]	2.65	4.41
4	[5,“sym19”]	2.53	4.07
5	[5,“rbio2.4”]	2.20	3.87
6	[5,“rbio2.4”]	2.25	3.84
7	[5,“sym7”]	2.47	4.33

Feature extraction of fault signals

After obtaining the PSO-optimized wavelet packet parameters, the vibration signals of the transformer under three kinds of fault conditions, such as no load, short circuit, and loose iron core, are decomposed and reconstructed using wavelet packet analysis.

This article takes the loosening signal of the cores as an example, uses PSO-optimized wavelet parameters to perform wavelet packet transform and obtain the wavelet coefficients of the signal. Then, through the inverse transform of wavelet packet decomposition, the decomposed components of each node are restored to the original signal X(n), and each transformed node is reconstructed. And calculate and normalize the energy of each node after reconstruction. Obtain the energy of each node frequency band as shown in Table 3.

Table 3.

Energy proportion of frequency bands.

X1	X2	X3	X4	X5	X6	X7	X8
0.088	0.003	0.005	0.031	0.077	0.025	0.036	0.055
X9	X10	X11	X12	X13	X14	X15	X16
0.242	0.023	0.011	0.045	0.095	0.069	0.023	0.003
X17	X18	X19	X20	X21	X22	X23	X24
0.011	0.002	0.006	0.063	0.007	0.008	0.006	0.010
X25	X26	X27	X28	X29	X30	X31	X32
0.005	0.004	0.003	0.006	0.008	0.009	0.002	0.019

As shown in Table 3, due to the excessive number of wavelet packet decomposition layers after PSO optimization, the normalized feature frequency bands are too many and the energy features are too complex. This not only makes data statistics not only difficult, but also more importantly, and it makes subsequent classification more complex and inaccurate.

Feature dimensionality reduction processing based on principal component analysis

As shown in Table 2, PSO-optimized wavelet packet decomposition typically exceeds 3 layers, leading to at least 16 frequency band energy nodes. This significantly increases data redundancy. To maintain data consistency and simplify processing, PCA is used to reduce the dimensionality of the normalized frequency band energy. The steps are as follows

Step 1, use the Z-score method to standardize the data and establish a standardized matrix Z.

Z_{i j} = \frac{x_{i j} - μ}{σ}

(6)

Among them, x_ij represents the element in row i and column j in the original data matrix, µ represents the mean of each feature, σ is the standard deviation of each feature.

Step 2, calculate the covariance matrix covZ of Z.

c o v Z = [\begin{array}{l} c o v (z_{1}, z_{1}) & \dots & c o v (z_{1}, z_{n}) \\ ⋮ & ⋮ \\ c o v (z_{n}, z_{1}) & \dots & c o v (z_{n}, z_{n}) \end{array}]

(7)

cov (z_{a}, z_{b}) = \frac{1}{m - 1} \sum_{i = 1}^{m} (z_{i a} - z_{a}) (z_{i b} - z_{b})

(8)

Step 3, calculate the characteristic roots of the matrix and arrange them according to their size, $λ_{1} \geq λ_{2} \geq \dots λ_{n}$

Step 4, calculate the contribution rate and cumulative contribution rate of each principal component

ρ_{i} = \frac{λ_{i}}{\sum_{j = 1}^{n} λ_{j}}

(9)

ρ = ρ_{1} + ρ_{2} + \dots + ρ_{k}

(10)

This article suppose that the principal components selected when the cumulative contribution rate exceeds 95% can represent all the original information. According to formula (10), the contribution rate reaches 98.27% when the first four dimensions are taken, which has reached the set standard. Therefore, the first four dimensions are taken as the new principal components, and their individual/cumulative contribution rates are shown in Table 4.

Table 4.

Covariance matrix eigenvalues and contribution rates.

Dimensions	Eigenvalue	Individual contribution	Cumulative contribution
1st	0.171	0.87	0.87
2st	0.012	0.06	0.93
3st	0.006	0.03	0.96
4st	0.003	0.02	0.98

The eigenvector corresponding to the eigenvalue is used to reduce the dimension of the frequency band energy data set. Finally, dimension 1, dimension 2, dimension 3, and dimension 4 as shown in Table 5 are used as data integration scores, with four dimensional features as feature groups and an additional column as the state recognition group. It is labeled as a feature data set for SVM classification.

Table 5.

Energy data after PCA dimensionality reduction.

States	1st	2st	3st	4st
No load	−0.12	−0.03	0.14	0.04
Short Circuit	0.16	−0.18	0.22	−0.01
Loose core	−0.28	−0.10	−0.02	−0.02

Classification and diagnosis of AMT vibration signals

Support vector machine theory

SVM classification is a machine learning method based on statistical learning theory and structural risk minimization. It can handle binary and multi-classification problems, and is applicable to both linearly separable and nonlinearly separable samples. Transformer fault diagnosis is a nonlinear multi-classification problem, and the nonlinear multi-classification transformation model of support vector machine is shown in formula (11) and formula (12)

\min (ω, ξ) = \frac{1}{2} ω^{2} + c \sum_{i = 1}^{l} ξ_{i}

(11)

s . t . {\begin{cases} y_{i} [ω^{T} φ (x_{i}) + b] ⩾ 1 - ξ_{i} \\ ξ_{i} ⩾ 0, i = 1, 2, \dots, l \end{cases}

(12)

The sample set is mapped from the input space to the high-dimensional feature space, and the Lagrangian function and decision function are constructed as follows

L (ω, b, ξ, α, β) = \frac{1}{2} ω^{2} + c \sum_{i = 1}^{l} ξ_{i} -

\sum_{i = 1}^{l} α_{i} {y_{i} [ω^{T} φ (x_{i}) + b] - 1 + ξ_{i}} - \sum_{i = 1}^{l} β_{i} ξ_{i}

(13)

f (x) = sign [\sum_{i = 1}^{l} α_{i} y_{i} K (x, x_{i}) + b]

(14)

where α_i, β_i are Lagrange multipliers; K(x, xi) is the kernel function for implementing SVM nonlinear transformation; x is the eigenvector. The selection of K(x, xi) has a great influence on the classification effect of the SVM model.

The Gaussian kernel function (radial basis function, RBF) has good generalization performance and only needs to determine one parameter. The selected RBF kernel function is shown in formula (15)

K (x, x_{i}) = \exp (- g ∥ x, x_{i} ∥^{2}), g > 0

(15)

In the equation, g is the free parameter of the RBF kernel function.

PSO optimizes SVM parameters

Since the transformer classification diagnosis problem is a nonlinear multi-classification problem, SVM with kernel function is usually used. When adjusting the parameters of the SVM, the c parameter controls the degree of penalty for the error term, and the gamma parameter determines the influence range of the support vector of the RBF kernel.

The performance of SVM largely depends on the selection of its parameters. Although traditional grid search and random search methods are simple, their efficiency is relatively low, especially in high-dimensional spaces. As an intelligent optimization algorithm, PSO can adaptively adjust and find better parameter combinations. And for the multi class nonlinear problem of transformer fault diagnosis, it does not rely on gradient effects and gradually approaches the optimal solution by adjusting the velocity and position of particles.

Therefore, the objective function for optimizing the SVM model in this paper is as follows

{\begin{cases} θ = [c, g] \\ J (θ) = - A c c (y t e s t, \hat{y} t e s t (θ)) \end{cases}

(16)

The PSO algorithm aims to minimize the objective function. Therefore, it necessitates taking the negative value to ensure accuracy.

According to the optimized objective function, the steps for PSO to optimize SVM are shown in Figure 6

Figure 6.

PSO optimization strategy for SVM.

AMT fault diagnosis examples

According to the algorithm flow in Figure 6, the feature samples are divided into 192 training samples and 48 test samples. The training samples include 72 no-load states, 72 short-circuit states, and 48 loose states, covering three kinds of fault conditions. This paper takes PSO-SVM as an example to analyze the optimization effect of particle swarm algorithm on support vector machine parameters. The training sample set is used to select the best SVM parameters.

Firstly, initialize the PSO parameters, with a maximum initialization population of 50 and a maximum population evolution of 50. Randomly assign each particle position (SVM parameter) and define the penalty parameter c $\in$ [0.1100] and kernel function g $\in$ [0.01,1000] range. Next, calculate the fitness value and use the maximum model accuracy as the objective function. Train the SVM with the SVM parameters represented by each particle position, compare the accuracy, and determine the position quality. Finally, find the optimal fitness value among the particles and update it to the global best position. The optimal SVM parameters obtained are penalty parameter c = 2.580991426 and kernel function parameter g = 10. To show the effect of PSO on transformer process optimization, its confusion matrix and fitness curve are shown in Figure 7. In the confusion matrix, the labels 1, 2, and 3, respectively represent “no-load,” “short circuit,” and “loose core.”

Figure 7.

Optimization effect of PSO-WPT-PSO-SVM model.

Ablation experiment

In order to verify the practicability of the whole-process optimization model, a WPT-SVM model without PSO optimization was defined, respectively, and the PSO-WPT-SVM model only optimized for feature extraction and the WPT-PSO-SVM model only optimized for classification prediction were compared with the whole-process optimization model in this paper. Among them, the SVM without optimization uses fixed parameters c = 1 and g = 0.1, while the WPT without optimization uses empirical wavelet parameters [3,“db2”]. Other parameters are set, and the model structure is consistent with the model in this paper. The results of ablation experiments are shown in Table 6.

Table 6.

Model evaluation before and after PSO optimization.

Feature extraction	Classification method	Precision	F1-Sore	Recall
WPT	SVM	0.78	0.77	0.77
WPT	PSO-SVM	0.85	0.83	0.83
PSO-WPT	SVM	0.82	0.82	0.85
PSO-WPT	PSO-SVM	0.92	0.91	0.92

Further comparing the impact of the main parameters on the prediction model, the confusion matrix shown in Figure 8 can be used to compare the actual measured values with the classification prediction results.

Figure 8.

Confusion matrices before and after parameter optimization.

According to Table 6 and Figure 8, after PSO optimization, WPT and SVM are better able to extract effective features of vibration signals, improving recognition accuracy, F1-Sore, and Recall compared to before optimization. Whether optimizing the wavelet parameters [N, Wname] or the SVM parameters [c, g], each method has shown significant optimization effects. Therefore, combining both approaches results in an effect that surpasses their individual contributions, achieving a synergistic effect greater than the sum of its parts.

The above explanation shows that using particle swarm optimization algorithm to optimize the WPT-SVM transformer fault recognition model has high accuracy and good stability.It can accurately and reliably identify three types of fault conditions in transformers: no-load, short circuit, and loose core.

Conclusions

In this study, a PSO-optimized WPT-SVM model for fault diagnosis of AMT is proposed based on the vibration signal of AMT clamp. Based on this model, not only can significantly improve the diagnostic efficiency of AMT, but also provide a new optimization idea for the subsequent research of transformer fault diagnosis. The conclusions are as follows:

(1) The PSO algorithm is employed to optimize the parameters [N, Wname] in wavelet packet decomposition, enhancing feature extraction efficiency and mitigating the uncertainty associated with manual parameter selection. Furthermore, PCA reduces data redundancy from the WPT feature extraction process, improve the efficiency of data processing.

(2) The model combines PSO to enhance the WPT-SVM transformer fault diagnosis process, improve fault identification. With a precision of 0.92, an F1-Score of 0.91, and a recall of 0.92, these metrics are notably higher than pre-optimization levels. Consequently, this method possesses a significant reference value.

(3) Transformer fault diagnosis is a systemic problem. Optimizing the entire fault diagnosis process, including signal acquisition, feature extraction, and classification prediction, rather than just the classification algorithm, can more effectively enhance the model’s diagnostic efficiency.

However, the adaptability of this model to different types of AMT is still lack of research, and how to establish a universal diagnosis model of AMT is still a problem to be solved. In the future, the research team will solve the above problems by expanding the scope of sample collection and summarizing the vibration characteristics of different types of AMT.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is supported by National Natural Science Foundation of China(52167017);Natural Science Foundation of Jiang xi Province (20224BAB204054);Ganpo Talent Support Program(20232BCJ22006).

ORCID iDs

Daosheng Liu

Wentao Yang

Longsheng Liu

Zhe Zhao

References

Jia

Cai

, et al. A new method of transformer short-circuit impedance regulation based on magnetic shunts. Energies 2024; 17(15): 3714–3735.

Tian

Fang

Zhao

. Short-circuit impedance analysis of HTS traction transformer with flux diverters using magnetic energy method. IEEE Trans Appl Supercond 2024; 34(8): 1–5.

Liu

Cheng

Zhang

, et al. Dielectric response evaluation method for multi-factor aging of converter transformer oil-paper insulation. High Volt Eng 2022; 48(5): 1684–1694.

Zhang

LiLi

Liu

, et al.fnm Method for quantitative assessment of transformer oil-paper insulation non-uniform ageing parameters based on frequency domain dielectric response. IET Sci Meas Technol 2022; 16(2): 118-129.

Zhang

. Characterizing transformer HV–LV winding FRA curves through derivation of transfer functions from FRA data. Electr Power Syst Res 2024; 228(5): 110055.

Wang

Qiu

Xie

, et al. Diagnosing fault types and degrees of transformer winding combining FRA method with SOA-KELM. IEEE Access 2024; 12(4): 50287–50299.

Aciu

A-M

Enache

Nițu

M-C

. A Reviewed turn at of methods for determining the type of fault in power transformers based on dissolved gas analysis. Energies 2024; 17(10): 2331–2356.

Taha

IBM

Dessouky

Ghoneim

SSM

. Transformer fault types and severity class prediction based on neural pattern-recognition techniques. Electr Power Syst Res 2021; 191(10): 106899–106909.

Liu

Longwu

, et al. Power transformer fault detection based on multi-eigenvalues of vibration signal. Trans Chin Electrotech Soc 2023; 38(1): 83–94.

10.

Yuan

Yan

Zhang

, et al. Research on vibration characteristic analysis and fault diagnosis method of oil-immersed transformer based on multi-physics coupling. IEEJ Trans Electr Electron Eng 2024; 19(10): 1621–1630.

11.

Lei

, et al. Dynamic vision-based machinery fault diagnosis with cross-modality feature alignment. IEEE/CAA J Autom Sin 2024; 11(10): 2068–2081.

12.

Sun

Yinjun

, et al. Transformer fault diagnosis method based on SCA-VMD and improved googlenet. Appl Sci Basel 2024; 14(2): 861–881.

13.

Zhang

Yang

Zheng

. Transformer fault diagnosis method based on timesnet and informer. Actuators 2024; 13(2): 74–92.

14.

Shang

Zhao

, et al. Partial discharge fault diagnosis in power transformers based on SGMD approximate entropy and optimized bilstm. Entropy 2024; 26(7): 551–559.

15.

Wang

Dong

Cheng

, et al. Transformer-based intelligent fault diagnosis methods of mechanical equipment: a survey. Open Phys 2024; 22(1): 107815–107823.

16.

Zhang

Hongshen

Zhang

. State of charge prediction of lithium-ion batteries for electric aircraft with swin transformer. IEEE/CAA J Autom Sin 2024; 11: 1–3. DOI: 10.1109/JAS.2023.124020.

17.

Zhao

Gao

Deng

. Defect detection using Shuffle net-CA-SSD lightweight network for turbine blades in IoT. IEEE Internet Things J 2024; 11(20): 32804–32812.

18.

Zhao

Deng

. Fuzzy broad neuroevolution networks via multi-objective evolutionary algorithms: balancing structural simplification and performance. IEEE Trans Instrum Meas. 2025; 74: 1–10. DOI: 10.1109/TIM.2024.3485438.

19.

Liao

Yang

Wang

, et al. Fault diagnosis of power transformers using graph convolutional network. CSEE J Power Energy Syst 2021; 7(2): 241–249.

20.

Feng

Zhang

Ouyang

, et al. Fault diagnosis of oil-immersed transformer based on TSNE and IBASA-SVM. Recent Pat Mech Eng 2022; 15(5): 504–514.

21.

Zhao

Wang

Peng

, et al. Transformer fault diagnosis based on a multi-strategy improved dung beetle optimizer. Power Syst Prot Control 2024; 52(6): 120–130.

22.

Han

Shi

, et al. A novel power transformer fault diagnosis model based on harris-hawks-optimization algorithm optimized kernel extreme learning machine. Electr Eng Technol 2022; 17(3): 1993–2001.

23.

Chen

Yang

, et al. Convolutional neural network-based transformer fault diagnosis using vibration signals. Sensors 2023; 23(10): 4781–4795.

24.

Song

Han

Yang

, et al. Magnetostrictive vibration characteristics of amorphous alloy transformer with three-dimensional wound core. IEEE Access 2024; 12(2): 43958–43967.

25.

Vandevelde

Melkebeek

JAA

. Magnetic forces and magnetostriction in electrical machines and transformer cores. IEEE Trans Magn 2003; 39(3): 1618–1621.

26.

Jaroslaw

Magdalena

. Investigation of distribution transformers vibrations in terms of core and winding condition assessment. Energies 2022; 15(1): 13–30.

27.

Kammeugue

Liu

. Investigation on vibration characteristics of amorphous metal alloy core dry-type distribution transformer. CES Trans Electr Mach Syst 2022; 6(3): 324–331.

28.

Liu

Zhang

, et al. Measurement analysis and improvement of vibroacoustic characteristics of amorphous alloy transformer. Energies 2022; 15(3): 949–964.

29.

Saleh

Aktaibi

. The development of a d–q Axis WPT-based digital protection for power transformers. IEEE Trans Power Deliv 2012; 27(4): 2255–2269.

30.

Min

Zhihui

. Fault location in hybrid transmission lines on the basis of VMD energy entropy and SVM. J Comput Methods Sci Eng 2023; 23(4): 1–17.

31.

Tang

Zhao

. Motor rolling bearing fault diagnosis based on MVMD energy entropy and GWO-SVM. J Vibroeng 2023; 25(6): 1096–1107.

Research on fault diagnosis of amorphous alloy transformers by using vibration signals and a PSO-optimized full-process WPT-SVM model

Abstract

Keywords

Introduction

AMT vibration signal acquisition and preprocessing analysis

AMT vibration test platform

Selection of AMT measuring points

Characteristic analysis of AMT fault signal

Feature extraction and optimization of fault signals

Wavelet packet transform and feature extraction principle

PSO parameter optimization for WPT

Feature extraction of fault signals

Feature dimensionality reduction processing based on principal component analysis

Classification and diagnosis of AMT vibration signals

Support vector machine theory

PSO optimizes SVM parameters

AMT fault diagnosis examples

Ablation experiment

Conclusions

Footnotes

Declaration of Conflicting Interests

Funding

ORCID iDs

References