Multi-scale spatio-temporal data modelling and brain-like intelligent optimisation strategies in power equipment operation and inspection

Clarifying the equipment fault situation existing in the UHV converter station in the process of electric power operation and inspection helps to better achieve the identification and diagnosis of fault categories. As a result, the collection of vibration signals of the UHV converter station equipment, filtering out the noise and interference of the fault signal to ensure that the pre-processed signal is close to the real fault signal. The vibration sensor is selected to collect the dynamic signal of the equipment, and the noise part of the signal must be eliminated due to noise pollution caused by the vibration signal. [23]. Describe the equipment vibration signal containing noise, the expression is: 1A(i)=B(i)+aC(i)

Where A(i) is the noise-containing signal at sampling point i, C(i) is the noise, a is the noise level, and B(i) is the true signal. The noise in the working environment of the equipment approximately obeys a Gaussian distribution, assuming C(i) is Gaussian noise, and using the least squares method, the trend term is fitted C(i) and the polynomial function is: 2C(i)=b0+b1i+b2i2+⋯+bmim

Where b_j is the coefficient to be determined, j ∈ (0,m); m are polynomials of order. Calculate the cumulative sum of squares c of the difference between A(i) and C(i), Eq: 3 c=∑i=1n∑j=0m(bjA(i)−C(i))2

Where n is the number of sampling points. Taking the partial derivatives ∂c∂b of c with respect to b_j to be 0 and m to be 0, and substituting into the above equation, we obtain: 4∑i=1nb0i−∑i=1nC(i)i2=0

Solve the system of linear equations, get the intercept b₀ of the trend term is ∑ i=1nC(i)n make A(i) subtract C(i) trend term, decompose the vibration signal with the trend term, get the signal with the trend term D(i). multi-layer wavelet decomposition of the signal with the trend term D(i), get the low-frequency and high-frequency signals, the low-frequency signals are rich in vibration information, and the high-frequency signals are rich in noise. The threshold method is used to compare the absolute value of the threshold and wavelet coefficients, and the wavelet coefficient of the decomposed detrended signal D(i) is set to be d, and the formula calculates the wavelet coefficient: 5d={ E−feE−fF,d>f0,d<f

Where F is an arbitrary positive constant, f is a threshold, E is a variable parameter, and changing the value of E makes the wavelet coefficients d vary between soft and hard threshold functions. Set the number of layers of wavelet decomposition, using wavelet coefficients d, processing the trend term signal D(i), remove the high frequency coefficients of high frequency signals, reconstruct the low frequency coefficients of low frequency signals, retaining the vibration information, to get the vibration signal after noise cancellation. This completes the pre-processing of the vibration signal of the energy equipment.

For equipment vibration signals of a certain length, this paper introduces a learning dictionary for fault feature extraction. In the dictionary training stage, the improved K-SVD dictionary learning algorithm is firstly used to train the sample signal matrix X = [x₁,x₂,…,x_m] ∈ R^n×m in order to obtain the dictionary D. After the dictionary D is obtained, the vibration signals need to be matrixed first because the length of the collected vibration signals is generally not the same as the length of the atoms.

After matrixing the vibration signal y, the OMP algorithm is used to solve the coefficient matrix w. In the process of solving w, the total number of iterations is set to be half of the number of atoms, and each iteration for each column signal y_i(i ∈ {l,2,…,P}) searches for an atom with the largest inner product with it. After the total number of iterations is performed, the number of iterations and the approximated signal are found when the maximum value of the cliff is reached, and the corresponding sparse coefficient vectors are used as the corresponding column vectors w_i in matrix w.

After finding the coefficient matrix w, the extracted fault feature components are found by the optimisation problem. Namely: 6X=argminXη||Y−X||F2+∑i=1P12||DV(i)−X(i)||22

Where η is the penalty factor. Clearly, the above equation is a quadratic optimisation problem and its solution can be expressed as: 7X=(2ηY+DW)/(1+2η)

The obtained X˙ is in the form of a matrix, and its column vectors are connected together at the same overlap rate (the overlap area is averaged) to form a one-dimensional signal. In practice, the best value of η is found by step-by-step searching method to ensure that the craggy value of the fault characteristic component is maximised.

After the time-domain fault signal feature extraction of the UHV converter equipment is completed, it is necessary to extract the frequency-domain signals for the equipment faults, and perform the signal modal decomposition in its frequency domain. In order to obtain a suitable instantaneous amplitude, the modal component needs to satisfy the following conditions, i.e., the number of signal extreme points should be ensured to be the same as the number of over-zero points over the entire length of the data, and the calculation formula is: 8f(t)=∑i=1nki(t)+εn(t)

Where f(t) is the t nd component obtained after modal decomposition of the signal, k_i(t) is the i th modal function, and ε_n(t) is the n th residual.

The modes are independent of each other, indicating that each vibration mode is actually effective, which provides a quantitative index for the fault type identification of UHV converter equipment through the independent vibration mode signal characteristics.

After obtaining the fault frequency domain signal after modal decomposition, the coefficient matrix composed of IMF components in each frequency band is decomposed by singular value to obtain the singular value that can reflect the basic characteristics of the original coefficient matrix. The complexity of this singular value is evaluated using the statistics of information entropy, so that there is a definite measure to reach the purpose of evaluating the complexity of the original signal [24].

Let EWT transform the original signal to obtain n layer of IMF components from low frequency to high frequency. The IMF components of each layer are formed into a matrix A of order m×n, which can be expressed according to the singular value decomposition theory as: 9Am×n=Um×lΛl×lVl×nT

Matrix U is of order m×l, matrix A is an orthogonal matrix of order l×l, and matrix V is of order l×n.

The diagonal matrix A = diag(λ₁,λ₂,…,λ_p). is in descending order and the main diagonal element of Λ is the singular value of matrix A. After normalising these p singular values, the empirical wavelet singular entropy is defined as: 10WSE=∑i=1kΔpi

Δp_i is the j nd order empirical wavelet singular entropy. Then: 11Δpi=−(λi/∑j=1lλj)ln(λi/∑j=1lλj)

The empirical wavelet singular entropy evaluates the complexity of the analysed signal. The simpler the analysed signal is, the smaller the empirical wavelet singular entropy is. The more complex the signal, the larger the empirical wavelet singular entropy is.

Combined with the boundary characteristics of the UHV multi-terminal hybrid DC line, it can be seen that when the in-area fault occurs, the voltage transient quantity signal contains rich high-frequency components, the signal is more complex, and the empirical wavelet singular entropy is larger. When an out-of-area fault occurs, the high-frequency components contained in the voltage transient signal after boundary attenuation are far less than those in the in-area fault, and the signal is simpler, and the empirical wavelet singular entropy is smaller. Therefore, the empirical wavelet singular entropy can be used to evaluate the complexity of the voltage transient signal, so as to realize the discrimination of in-zone and out-of-zone faults.

Due to the existence of two different categories of data in the time domain and frequency domain of the fault signal of UHV converter equipment, the signal data and the extracted features are more and thus cause redundancy of information, which makes some useful feature information can not be recognised, resulting in the fault identification of UHV converter equipment is not accurate enough. Therefore, this paper introduces kernel principal component analysis (KPCA) to carry out multi-scale time-frequency feature fusion and then complete the feature dimensionality reduction [25]. The specific steps are as follows: 1)

The vibration signal of the UHV converter equipment is collected, and the vibration signal is recorded as q = {s_x,s_y,s_z}, and the subscript x,y,z indicates the x-axis, y-axis and z-axis of the vibration signal, respectively. The sample length is set to m. Obviously, sample q is the data of m×3. The number of samples for each state is N_p(p = 1,2,…,P).

The UHV converter equipment fault dataset used in this paper comes from the operation and inspection data of the GZ section of the NF grid from 2020 to 2023. The UHV converter equipment fault types mainly include normal (Nor), inner ring fault (IR), outer ring fault (OR), cage fault (Cage), inner and outer ring composite fault (IR-OR), and inner and outer ring + rolling body composite fault (TCF). 400 samples are taken for each mode, 200 data points for each sample, and 80,000 data points for each fault type.

In order to verify the effectiveness of the multiscale time-frequency feature fusion algorithm proposed in this paper, time-frequency feature extraction is performed based on the above dataset and multiscale feature fusion is performed by KPCA. Firstly, the parallel method is used for fusion to get the fused time-frequency feature data, which avoids the high dimensionality of features in the fusion process and reduces the computational difficulty. Then KPCA method is used to map the fused time-frequency features to the high-dimensional space for PCA operation to extract the principal components and to get the final fusion features with obvious fault information and low dimensionality. After many experiments, the classification effect is best when 6 principal components are extracted. Therefore, in this paper, the 6 principal components of the features are extracted and used for the next analysis, and the feature values of some samples after fusion are shown in Table 1.

The distribution plots of the samples are made for the fused multi-scale time-frequency eigenvalues of the UHV converter equipment, and the results are shown in Fig. 1, in which Figs. 1(a)~(f) is the distribution results of PCA1~PCA6, respectively, and the vertical axis of the plots is the eigenvalues, and the horizontal axis is the ordinal number of all the categories of samples.

Figure 1.

The sample distribution of the fusion eigenvalue

Taking 400 samples as an interval, the 1st-1600 samples are normal, inner ring faults, outer ring faults, and cage faults in order, the 1601st-2000th samples are inner and outer ring composite faults, and the 2001th-2400th samples are inner and outer ring + rolling body composite faults. From the figure, it can be seen that the differences in eigenvalues of different failure modes are relatively large, and a single eigenvalue can not completely distinguish the different types of failures.PCA1 has relatively good differentiation for almost all failure modes, and the differences in PCA1 eigenvalues between normal samples and other failure types are so large that normal samples and fault samples can be well distinguished, and the inner ring and outer ring and other faults also have a relatively good differentiation. The PCA2 eigenvalues of the cage can be clearly distinguished from other fault types, and there is also a relatively obvious difference between the inner and outer ring composite faults as well as the inner and outer ring + rolling element composite faults, but the eigenvalues of the normal, inner and outer ring fault types overlap slightly. The outer ring faults PCA3 and PCA4 eigenvalues differ from normal and inner ring faults, but there is a relatively large overlap with the other types. The inner and outer ring composite faults of PCA5 and PCA6 are significantly different from the other fault types. However, the fault types cannot be accurately determined by these plots alone, and further classification algorithms are needed to differentiate between them. The PCA1-PCA6 obtained in this section are inputted into the classifier, which visualises the classification results and accurately classifies the fault patterns.

Generative Adversarial Network (GAN) is a deep learning model that is unique in that it is trained through a game between the generator and the discriminator, enabling the generator to gradually improve its ability to generate realistic samples while the discriminator continuously improves its ability to discriminate between true and false samples. This dynamic equilibrium forms a zero-sum adversarial game between very small and very large entities, similar to the process of two intelligences improving each other in competition. [26].

The generator’s objective is to learn the data’s underlying distribution from random noise and generate samples that are similar to the actual data. During training, the generator tries to trick the discriminator into not being able to distinguish the generated samples from the real training data by receiving a random vector as input. During backpropagation, the generator adjusts its parameters based on feedback from the discriminator, gradually increasing the fidelity of the generated samples. The discriminator is trained to be a professional discriminator whose goal is to effectively distinguish between fake samples generated by the generator and real training data. The discriminator’s ability to classify fakes and forgeries is continuously improved by receiving samples from both sides, which increases the difficulty of adversarial training. In this process, the discriminator continuously improves its ability to discriminate more accurately between the generator’s output and the actual data. Throughout the training process, the generator and the discriminator fight against each other to form a dynamic equilibrium. The optimisation function of GAN can be expressed as: 14minGmaxDV(D,G)=Ex~pdata(x)[logD(x)]+Ez~pz(z)[log(1−D(G(z)))]

The generator endeavours to improve the realism of the generated samples in order to deceive the discriminator, while the discriminator endeavours to improve its ability to recognise authentic samples in order to better discriminate between the generator’s outputs and the real data. The GAN’s zero-sum adversarial gaming process allows the generator to gradually understand the latent distribution of the data and generate more realistic samples.

Nearest Neighbour Interpolation (NNI) is a KD-tree based interpolation algorithm, where the main idea is to simply determine the “nearest” neighbour sample point and assume its value with a given “y” value, instead of calculating the mean value by some weighting criteria or generating an intermediate value according to complex Instead of calculating an average value by some weighting criteria or generating an intermediate value based on complex rules [27].

Suppose {(x₁,y₁),(x₂,y₂)…,(x_n,y_n)} is a sample point in a given two-dimensional sample space, and their labels {z₁,z₂,…,z_n} are already known. If (x_m,y_m) is a sample point interpolated into the sample space, the whole computation process can be described as follows: 1)

Use the known labels to create data points C as: 15C={(xm,ym,z1),(xm,ym,z2)…,(xm,ym,zn)}

Iterate steps (1) and (2) with different sample points in the interpolation grid.

We construct an interpolation grid in this paper and use NNI to extend the original dataset. It aims to fully capture a small range of information for each sample point. Due to the large difference in values between different descriptors, normalisation would make them lose their original physical meaning. Therefore, we assume that all interpolated points obey a normal distribution when interpolating N(N, σ²). The mean and variance are described as: 16μ=xij 17σ2=λm∑i=1mxij

Where x_ij is the value of the jrd feature of the ind sample, m is the number of samples, and λ is the variable that controls the size of the interpolated region. Choosing a larger λ means that the larger the area covered by the interpolation, the more information can be included.

In order to achieve effective identification of UHV faults, a nearest-neighbour generative fragment model (GPNN) combining production adversarial network and nearest-neighbour difference algorithm is introduced based on multi-scale time-frequency features of faults. The fusion of multi-scale time-frequency features and proximity sample similarity is used to generate UHV converter station equipment fault samples, which are applied to UHV converter station equipment fault identification, helping to enhance the diagnostic accuracy of brain-like intelligence.

Where u denotes the membrane voltage, τ_m is the membrane time constant, and τ_m = RC. The voltage values of the IAF model are not directly represented as pulses, but are accumulated until they exceed a threshold value of V_mh, at which point the model releases a spike and enters an undesired state.

The IAF model equation is: 23ui(t)=∑ti′∈Fiηi(t−tif)+∑j∈Γi∑t/∈Fiwij∈ij(t−tif)

Where η_i and ò_ij are kernel functions of the specific form: 24{ ηi(s)=−(ϑ−urezt)exp(−sτm)H(s)∈(s)=∫0∞exp(−s′τm)1τsexp(−s−s′τm)H(s

The learning rule of SNNs is to change the weights by the relative time difference of the impulses emitted by the neurons before and after the synapse. The learning rule of SNNs can be classified into three modes, namely supervised learning, semi-supervised learning and unsupervised learning, according to the supervision mode. In this paper, we mainly introduce the time-dependent plasticity (STDP) based algorithm, which is applied to the learning of SNNs.

The update of weights under STDP learning relies on the pre and post-synaptic pulse sequences, which can be expressed as: 25si(t)=∑fδ(t−tif),sj(t)=∑fδ(t−tjf)

The expression for the STDP learning rule is: 26dwij(t)dt=a0+si(t)[ a1pre+∫0∞a2pre,pozt(s)sj(t−s)ds ]+sj(t)[ a1pozt+∫0∞a2pozt,pre(s)si(t−s)ds ] s_i(t) & s_j(t) are sequences of presynaptic neurons i and postsynaptic neurons j. Parameter a₀ is a decay term, where the synaptic weights decay over time when no signal transmission occurs. aipre and aipost are non-Hebbian terms, indicating that synchronous hyperpolarisation is not required before and after the synapse. The kernel functions a2pre.poct and a2pozt.pre in the integrals indicate the correlation between the change in weights and the pulse duration, s is the time difference between pre- and post-synaptic neurons s=tjf−tif , and the kernel functions a2pre.poct and a2poct.pwe express the characteristics of the learning window for STDP learning. The different parameters determine different forms of STDP learning rules, which are often taken as a0<0,alpre>0,alpozt<0 .

The response of the postsynaptic neuron occurs when the presynaptic pulse arrives, when s>0 exhibits an incremental weight change. When the postsynaptic impulse is generated first, s<0 then this information is likely to be ignored and therefore the weights show a decreasing change.

Where, according to the characteristics of the learning window, the STDP learning window W(s)^STDP can be expressed as: 27W(s)STDP={ a2post.pro(s)=+A+exp(−sτ+),s≥0a2pro.post(−s)=−A−exp(sτ−)s<0

Where τ₊ and τ_– are time constants, and A₊>0 and A_–>0 denote the maximum values of synaptic changes.

With the support of the time-frequency characteristics of UHV converter equipment faults, GPNN is used to generate equipment fault time sequence samples, and the brain-like computing SNNs model is introduced to carry out the intelligent diagnosis of UHV converter equipment faults, which provides technical support for improving the efficiency of grid operation and inspection and intelligent management.

For the SNNs-based fault diagnosis model of UHV converter equipment, the mapping relationship between fault characteristics and fault modes is to be established. Namely: 28f:X→Y

In the mapping relationship f, the original data set X = {(x_i, y_i)|1≤i≤N}, where N represents the total sample data, Y represents the label space. The core of SNNs model is to continuously update the parameters, so that the error between the actual output and the expected value is gradually reduced, and the loss function is continuously reduced until the model is optimised at the end of training. In the learning process of all kinds of network models, it is essential to update the weights, given ω represents the weights, α represents the learning rate, E represents the loss function, and we get: 29ω=ω−α∂E∂ω

Therefore, let Njpoct be the jnd posterior neuron in the output layer, let Nipre be the ith anterior neuron in the input layer, let ω_ij be the value of the connection weights of Nipre , and Njpoer and Uij(t) be the value of the membrane voltage of neuron Njpact at moment t for the nth sample.

Theorem 1 Let U_j(t),U_j(t+1) denote the membrane voltage value of the postsynaptic neuron Njpozt at moments t and t+1, then the membrane voltage update formula is: 30Uj(t+1)=exp(−1/τ)Uj(t)+(1/n)∑i=1nωij−(θexp(−1/τ))/n

Where η denotes the phase of ineligibility after the pulsed neuron has been activated, during which the pulsed neuron gradually returns to its resting state. When the pulsed neuron is during the period of ineligibility, no response is generated during that time. The computational expression for the period of ineligibility is: 31η=−θexp(Δt/τ)

Where θ denotes the threshold, Δt denotes the time interval, and τ denotes the time constant. The introduction of the pulse train expression indicates the change in the relationship between the output data and the membrane voltage.

In general, when performing model tuning, it is necessary to select the appropriate loss function for the current network model. In the SNNs in this paper, the loss function is selected using the cross-entropy loss function. When the training model produces a predicted output during the training process, the predicted output is compared with the true output, the resulting loss is calculated, and a penalty value in logarithmic form is set based on this loss. It follows that the cross-entropy loss function usually reflects the degree of similarity between the actual output data of the network model and that in the training data, let E denote the loss function, then: 32E(y,y^)=−[ylogy^+(1−y)log(1−y^)]

In the formula of cross-entropy function, y represents the desired output and y^ is the actual label.

Based on the loss function formula, the sensitivity change function between each connected neuron in the multilayer network is given ξil(t) , l denotes the number of layers where the ith neuron is located, and t denotes the current moment, then the sensitivity change function is: 33ξil(t)=∂E(y,y^)∂Uil(t)

At this point, the weight update formula evolves into the following equation, such that K is the total number of samples, T is the total time step, and the amount of weight change Δω_i between the jth and ith neuron in layer r neurons is: 34Δωij=α1KT∑k=1K∑t=1Tξk,jr+1(t)∂Uk,jr+1(t)∂ωij

Combining the optimal learning of SNNs with the STDP learning algorithm given in the previous section, and then applying the model to the field of fault diagnosis of UHV converter equipment, the multi-scale spatio-temporal sample dataset is inputted into the SNNs model, so as to realise the intelligent diagnosis of faults of UHV converter equipment.

Based on the multi-scale time-frequency characteristic data of UHV converter station equipment faults obtained in the previous paper, combined with the GPNN model constructed in this paper, it is applied to the multi-scale fault sample generation of UHV converter station equipment. The dataset is divided into a training set and test set according to 8:2, and FID is selected as the evaluation index of the quality of fault sample generation and the process data of FID for different categories of faults are shown in Table 2. The last column represents the minimum FID value in the entire training process, which saves the parameters of the GPNN model at this time and generates fault time-frequency samples based on it. The smaller the FID is, the smaller the distribution distance between the two datasets is, which can be regarded as higher quality fault sample generation. FID has better robustness to noise, and the computational complexity is not high. Only the first-order and second-order moments of the samples are considered. As can be seen from the table, as the number of iterations increases from 50 to 500, the generated FID values of the six different categories of fault samples will gradually decrease, in which the FID value of the inner and outer ring + rolling body composite fault (TCF) is the smallest 21.16. The above results show that the GPNN model can obtain generated samples closer to the real samples of faults, and the distribution of the generated data is becoming more and more similar to the real data distribution.

After the training of the GPNN model is completed, the network continues to perform 10,000 network iterations using the model and the multiscale time-frequency fault sample data for fault sample generation at the sampling time step, and its specific generation sequence is shown in Fig. 2 for the comparison with the real sample segments. The horizontal axis in the figure represents the time step, while the vertical axis indicates the normalized values of the corresponding multi-scale time-frequency fault types.

Figure 2.

The GPNN generates the fault sample sequence

From the comparison of the real fault samples with the generated fault sample sequence data in terms of numerical size and changes, it can be seen that the two types of data do not fit numerically, but there is a certain temporal correlation within the time step, which indicates that the generated data retains the static and temporal characteristics of the real data to a certain extent, and on the whole, the fault sample sequences generated by using the GPNN model are in good agreement with the real fault sample data can be more than 90.57% consistent. In addition, the difference in the numerical size between the generated data and the real data also indicates that the generated data are not numerically close to the real data, and the use of such data for fault sample generation is conducive to increasing the diversity of fault samples.

Six different fault types generated by the GPNN model are used as data labels to validate the validity of the SNN model using the training and test sets, respectively. Each fault type contains 150 training samples and 50 test samples, and each sample consists of three types of vibration feature maps, corresponding to time-domain feature maps, frequency-domain feature maps and time-frequency-domain feature maps, respectively. In this paper, the cross-entropy function is mainly used as the loss function for model training, which aims to improve the generalization ability of the brain-like fault diagnosis model. The trend of model accuracy and loss value during training is shown in Fig. 3.

Figure 3.

Change of accuracy rate and loss value during training

In order to reduce the volatility of the curve, the training process of the model is shown, and the horizontal coordinate is the iteration period, i.e., every traversal of the training set for 1 time, the value of the loss function, and the accuracy rate is recorded. It can be seen that the model converges quickly after the 6th iteration cycle, the loss function value drops to about 0.034, and the accuracy of both the training set and the test set reaches more than 94%. In addition, because this paper adopts the small batch stochastic gradient descent algorithm, the small batch test set taken out each time has a certain degree of randomness, so the test set accuracy rate fluctuates slightly, but the overall trend remains up. It can be seen that the brain-like intelligent fault diagnosis model can better fit the multi-scale time-frequency fault vibration data set of the UHV converter equipment, with high fault diagnosis speed and accuracy.

In order to verify the comprehensive fault diagnosis performance of the brain-like intelligent fault diagnosis model designed in this paper, XGBoost, SVM, CNN, LSTM, GRU, and CNN-GRU fault diagnosis models were constructed, respectively, and fault diagnosis was carried out on the test set to compare their test accuracies. In order to avoid chance errors and objectively evaluate the diagnostic ability of the models, multiple fault diagnosis experiments are required, and in this paper, a total of 50 times of fault diagnosis were conducted for each model, and the average of the 50 times of fault diagnosis accuracy was taken as the final diagnostic accuracy. Figure 4 displays a fault diagnosis accuracy rate of fifty times.

Figure 4.

Comparison results of fault diagnosis model

In terms of multi-scale time-frequency fault diagnosis of UHV converter station equipment, the model in this paper has both multi-scale deep features and time-series feature extraction capability, and the accuracy rate is improved by 24.35%, 21.19%, 12.16%, 11.93% and 8.41% compared to SVM, XGBoost, LSTM, GRU and CNN models that have only single-feature extraction capability, respectively, and the accuracy rate is improved by 6.22% compared to the combined CNN-GRU model with a single branch. As for the fastness, the structurally simple machine learning models SVM, XGBoost and the single deep learning models LSTM, GRU and CNN perform better, but the models built as a whole satisfy the fastness requirement. Although the single deep learning model CNN, GRU, and LSTM have shorter diagnostic times, it can be seen from several experiments that its accuracy and stability are insufficient. From the diagnostic process of the fault diagnosis model, the diagnostic duration of this paper’s model is less than 2.42ms, which meets the requirement for rapid fault diagnosis.

The accuracy of this paper’s model in 50 times fault diagnosis has a significant improvement over other models, reaching 98.06%, and its training time is only 37.14 seconds, which is much lower than other models. Since the model in this paper is based on brain-like computing and extracts the multi-scale features of the faults of the UHV converter equipment through the GPNN model, it is more capable of extracting the features at different scales and has a faster computational speed. Due to the optimization of the key parameters in the model, the diagnostic model with the best parameter ratios is obtained, which further improves the computing performance and diagnostic accuracy of the model. In summary, the model in this paper has higher accuracy while still being fast.