Mutual inductor fault detection method and system

Abstract

The invention relates to the technical field of fault detection, in particular to a mutual inductor fault detection method and system, and the method comprises the steps: synchronously collecting a mutual inductor iron core vibration acceleration signal, a leakage magnetic flux intensity signal and a surface temperature field distribution signal, and generating a multi-physical field original data set; carrying out mechanical resonance characteristic analysis on the multi-physical field original data set, implementing adaptive noise filtering, and outputting a de-noised characteristic set; inputting into a multi-scale feature fusion module, and generating a multi-dimensional feature tensor comprising a time-frequency feature, a spatial distribution feature and an energy evolution feature; dynamically extracting a fault sensitive factor set from the multi-dimensional feature tensor, wherein the fault sensitive factor set comprises a mechanical deformation sensitive factor, an insulation degradation sensitive factor and a poor contact sensitive factor; and constructing a fault mode recognition model, and outputting corresponding fault types, fault levels and positioning information. According to the invention, through multi-physical field synchronous acquisition and fusion analysis, the accuracy and reliability of mutual inductor fault detection are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault detection, and in particular to a mutual inductor fault detection method and system. Background Art

[0002] As a key measuring device in the power system, the operational reliability of current transformers directly affects the accuracy of grid protection and control systems. Traditional transformer fault detection methods mainly rely on single physical quantity monitoring, such as vibration signal analysis or temperature threshold alarms, which have obvious technical limitations: on the one hand, typical faults such as mechanical deformation of the core and deterioration of the winding insulation only manifest as weak multi-physical field characteristic coupling changes in the early stages, which are difficult to accurately capture with a single signal; on the other hand, complex electromagnetic interference and temperature fluctuations on site will lead to an increase in the false alarm rate.

[0003] In recent years, multi-sensor fusion technology has provided new insights for fault detection, but existing solutions still suffer from three key issues: First, the timing synchronization accuracy of vibration, leakage flux, and temperature signals is insufficient, resulting in feature fusion distortion; second, traditional machine learning algorithms are inefficient in extracting multi-physics field coupling features, especially the correlation between the spatial distribution of the temperature field and the resonance characteristics in the time-frequency domain; and finally, fault location lacks a multi-dimensional evidence cross-validation mechanism. Therefore, an intelligent detection solution is urgently needed that can achieve high-precision multi-physics field synchronous analysis, has adaptive noise suppression capabilities, and can output quantitative diagnostic results. Summary of the Invention

[0004] The present invention provides a mutual inductor fault detection method and system.

[0005] A method for detecting a transformer fault comprises the following steps:

[0006] S1: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate a multi-physics field raw data set;

[0007] S2: Analyze the mechanical resonance characteristics of the multi-physics field original data set, implement adaptive noise filtering by constructing a resonance frequency offset compensation model, and output a denoised feature set;

[0008] S3: Inputting the denoising feature set into a multi-scale feature fusion module to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features, and energy evolution features;

[0009] S4: dynamically extracting a set of fault sensitive factors from the multidimensional feature tensor based on an improved random fern classifier, wherein the set of fault sensitive factors includes a mechanical deformation sensitive factor, an insulation degradation sensitive factor, and a poor contact sensitive factor;

[0010] S5: Constructing a fault pattern recognition model according to the set of fault sensitive factors, and outputting corresponding fault type, fault level and location information.

[0011] Optionally, the S1 includes:

[0012] S11: The vibration acceleration signals of the key structural positions of the core are collected by arranging a three-dimensional MEMS acceleration sensor array. The sampling frequency is set to f a =10kHz, generating the vibration acceleration signal A(t), which is expressed as:

[0013]

[0014] Among them, a xi (t),a yi (t),a zi (t) are the acceleration signals of the i-th sensor in the x / y / z direction, n is the number of acceleration sensors, t is the sampling time point, t = k / f a ,k∈N + ;

[0015] S12: High-sensitivity Hall magnetic flux sensors are placed around the core to collect transient leakage magnetic flux intensity signals and use time synchronization markers T s The time stamp is unified with the acceleration signal to generate the leakage flux intensity signal M(t), which is expressed as:

[0016] M(t)=[m1(t),m2(t),...,m k (t)];

[0017] Among them, m j (t) is the output of the jth flux sensor at time t, k is the number of flux sensors;

[0018] S13: Continuously image the surface temperature field of the transformer through the infrared thermal imaging module, spatially sample the thermal image using the region partition matrix R, and extract the surface temperature field distribution signal T(t), which is expressed as:

[0019] T(t)={T i,j (t)},i=1,…,h; j=1,…,w;

[0020] Among them, T i,j (t) represents the surface temperature value of the (i, j)th pixel area at time t, h, w are the height and width pixels of the image;

[0021] S14: Based on the unified time base T sThe vibration acceleration signal, leakage magnetic flux intensity signal, and surface temperature field distribution signal sources are time-aligned and expressed using a linear interpolation function. The asynchronously sampled signals are resampled to generate a multi-physics field raw data set consisting of three channels, which is expressed as:

[0022] X s (t) = interp(X(t i ),T s );

[0023] D(t)={A(t),M(t),T(t)};

[0024] Among them, X(t i ) represents the original signal sequence, T s The target unified time axis, interp represents the linear interpolation function, X s (t) is the output signal after alignment, and D(t) is the original data set of the multi-physics field.

[0025] Optionally, the S2 includes:

[0026] S21: Extract the frequency domain features of the vibration signal through short-time Fourier transform and combine it with temperature field modeling to compensate for the temperature drift of the resonant frequency;

[0027] S22: Adaptive filtering is used to suppress leakage flux noise interference, and vibration, leakage flux intensity and temperature signals are integrated to generate a multi-dimensional denoising feature set.

[0028] Optionally, the S21 includes:

[0029] S211, Mechanical Resonance Frequency Domain Analysis: Perform short-time Fourier transform (STFT) on each sensor channel in the vibration acceleration signal A(t) to extract the main resonant frequency f r and its harmonic component set {f r1 ,f r2 ,...,f rm}, satisfying P(f r ,τ)≥α.max(P(f,τ)), α is the energy threshold coefficient, and α∈(0,1), calculate the time-frequency energy spectrum matrix P(f,τ), expressed as:

[0030]

[0031] Where w(t-τ) is the Hanning window function, τ is the center position of the time window, f is the frequency component, and i∈{1,2,...,n} represents the sensor number;

[0032] S212, Resonant frequency temperature drift modeling and compensation: Based on the surface temperature field distribution signal T(t), establish the resonant frequency temperature drift model Δfr (t), expressed as:

[0033]

[0034] in, is the average temperature rise, T0 is the reference temperature, β is the static temperature coefficient, and γ is the dynamic temperature coefficient;

[0035] Calculate the resonant frequency after compensation, expressed as:

[0036] f r ′(t)=f r (t)-Δf r (t);

[0037] Among them, f r ′(t) is the resonant frequency after compensation, f r (t) is f r (t) Original main resonant frequency.

[0038] Optionally, the S22 includes:

[0039] S221, Adaptive noise suppression based on LMS algorithm: Construct an adaptive filter based on the LMS algorithm, take the leakage flux intensity signal M(t) as the reference noise input, the vibration signal as the desired signal, and output the filtered image, which is expressed as:

[0040]

[0041] Among them, w p (t+1)=w p (t)+μ·e(t)·m j (tp) is the weight update formula, L is the filter order, μ is the convergence factor, e(t)=a xi (t)-y(t) is the error signal;

[0042] S222, multi-physics field feature fusion and denoising set generation: fuse the denoising results of each physical field and output a multi-dimensional denoising feature set F(t), expressed as:

[0043]

[0044] Among them, M′(t)=M(t)-y(t) is the flux signal after decorrelation, T′(t)=T(t)·H(f r ′(t)) is the frequency-domain weighted temperature field distribution signal, and H(f) is the resonant frequency matching window function.

[0045] Optionally, the S3 includes:

[0046] S31, time-frequency feature extraction: vibration acceleration signal after denoising Perform continuous wavelet transform to extract the peak sequence in the time-frequency energy distribution and generate the time-frequency feature vector W(a,b) that represents the dynamic change of mechanical resonance, which is expressed as:

[0047]

[0048] Among them, ψ(t) is the mother wavelet function (Morlet wavelet is selected), a is the scale parameter, corresponding to the frequency a0 is the scale reference corresponding to the wavelet center frequency, b is the translation parameter, which represents the time offset, and i∈{1,2,...,n} is the sensor photo number;

[0049] Extract the time-frequency energy peak sequence E TF (t)=max a |W(a,b)| 2 , forming the time-frequency feature vector F TF (t);

[0050] S32, spatial distribution feature modeling: perform spatial convolution coding on the temperature field distribution signal T′(t), using the convolution kernel K s Extract local temperature gradient features C ij (t), expressed as:

[0051]

[0052] Among them, K s The Sobel operator of (2d+1)×(2d+1) is used to detect the spatial gradient of the temperature field, d=1, the convolution kernel size is defined as 3×3, h, w are the height and width pixels of the temperature field image;

[0053] The convolution output C(t) is divided into regions to calculate the mean and variance, and generate the spatial distribution feature matrix F SP (t)∈R h ×w×2 ;

[0054] S33, energy evolution characteristics analysis: Through sliding window integration and entropy calculation, the energy accumulation and complexity of the leakage flux intensity signal M′(t) are quantified and expressed as:

[0055]

[0056] Among them, E M (t) is the sliding energy window integral, N = f a ΔT is the window length, ΔT = 0.1s is the integration time window, and k is the number of magnetic flux sensors;

[0057] The energy evolution complexity is quantified by the entropy value, and the energy evolution feature vector is generated, which is expressed as:

[0058]

[0059] F EN (t)=[E M (t),H E (t)];

[0060] Among them, H E (t) is the energy evolution complexity, F EN (t) is the energy evolution characteristic;

[0061] S34, multi-dimensional feature tensor fusion: the time-frequency feature F TF (t), spatial distribution characteristics F SP (t) and energy evolution characteristics F EN (t) Splicing by channel dimension to form a multi-dimensional feature tensor Expressed as:

[0062]

[0063] Flatten(·) means flattening the matrix into a vector, D = dim(F TF )+2hw+2 is the total feature dimension.

[0064] Optionally, the S4 includes:

[0065] S41, Improved Random Fern Classifier Construction: Define dynamic weight random fern classifier, composed of N f fern structure Composition, each fern From the multidimensional feature tensor Randomly select L feature subsets S from m ={f m1 ,f m2 ,...,f mL}, and calculate its dynamic classification weight ω m (t), expressed as:

[0066]

[0067] Among them, λ is the weight adjustment factor (λ>0), Var(S m ) represents the subset S m The characteristic variance, L is the number of features selected for each fern (L≤D, D is The total dimension of ), m is the fern structure index number;

[0068] S42, dynamic extraction of fault-sensitive factors: For three types of fault factors (mechanical deformation, insulation degradation, and poor contact), a fern classifier is used to output probability mapping;

[0069] S43, sensitive factor set generation: Fusion of three types of sensitive factors, output of dynamically updated fault sensitive factor set Q(t), expressed as:

[0070] Q(t)={Q mech (t),Q insul (t),Q contac (t)};

[0071] Each type of sensitivity factor is normalized by the threshold and expressed as:

[0072]

[0073] Among them, μ type ,σ type are the mean and standard deviation of historical data respectively.

[0074] Optionally, the S42 includes:

[0075] S421, Mechanical deformation sensitivity factor: based on time-frequency characteristics F TF (t) and energy evolution characteristics F EN (t), calculate the weighted voting result of the fern classifier, expressed as:

[0076]

[0077] Among them, θ mech is the mechanical deformation sensitivity factor, δ(·) is the indicator function (output 1 when the condition is met, otherwise 0), N f The number of random ferns is set to 50 by default;

[0078] S422, Insulation degradation sensitivity factor: Based on spatial distribution characteristics F SP (t), calculated by convolution feature response, is expressed as:

[0079]

[0080] Among them, Q insul (t) is the insulation degradation sensitivity factor, σ(·) is the Sigmoid function, C ij (t) is the temperature field convolution output;

[0081] S423, poor contact sensitivity factor: combined with the energy entropy H of the leakage flux signal M′(t) E (t), dynamically adjust the fern classifier weight, expressed as:

[0082]

[0083] Among them, Q contac (t) is the poor contact sensitivity factor, and k is the number of leakage flux sensors.

[0084] Optionally, the S5 includes:

[0085] S51, Fault mode feature matrix construction: Fuse the fault sensitive factor set Q(t) with the multi-scale feature tensor F(t) to construct the fault mode feature matrix G(t)∈R N×M , expressed as:

[0086]

[0087] Where W∈R M×D is a trainable weight matrix, D is the total feature dimension, b is the bias vector, N = 3 is the number of fault types (mechanical deformation, insulation degradation, poor contact), M is the feature mapping dimension (M = 64 by default), and || represents the vector concatenation operation;

[0088] S52, hierarchical identification based on attention mechanism: multi-head attention mechanism is used to calculate the probability distribution P of each fault type type and fault level score S level , expressed as:

[0089]

[0090] S level =softmax(W l ·P type +b l );

[0091] Among them, W q ,W k ,W v are the weight matrices corresponding to query, key, and value, respectively. d is the dimension of each attention head. The softmax operation is used to normalize the attention weights and output the probability distribution of each fault type. l is the level weight matrix (level 5 corresponds to 3 types of faults), b l is the bias vector;

[0092] S53, Fault Location Analysis: Analyze spatial positioning and vibration source positioning to obtain positioning information, including:

[0093] (1) Spatial positioning L heat :For temperature field characteristics F SP (t) performs deconvolution operation, which is expressed as:

[0094] L heat=Deconv(F SP (t),K d );

[0095] Among them, K d is a 3×3 deconvolution kernel;

[0096] (2) Vibration source positioning L vib :Based on time-frequency features F TF (t) Calculate the energy center of gravity, expressed as:

[0097]

[0098] Where n is the number of acceleration sensors, W(a i ,b j ) is the wavelet coefficient;

[0099] S54, state identification and early warning: Based on the fused fault type probability, fault level score and location information, a comprehensive decision function Y(t) is constructed, which is expressed as:

[0100] Y(t)=αP type +βS level +γ(L heat +L vib );

[0101] Among them, α, β, γ are weight coefficients, satisfying α+β+γ=1;

[0102] When Y(t)>η (η is the preset threshold, set according to historical data statistics), a graded warning (such as mild, moderate, severe) is triggered, and the fault type, level and location information (temperature abnormality area and vibration source location) are simultaneously output;

[0103] The preset threshold η is set based on historical data statistics, specifically including:

[0104] (1) Initial threshold calculation: Based on historical normal data, calculate the mean μ and standard deviation σ of the comprehensive decision function Y(t), and set the initial threshold η0 = μ + 3σ (covering 99.7% normal operating conditions);

[0105] (2) Dynamic adjustment mechanism: Set a sliding window (1 week of data). Specific adjustment rules include:

[0106] When the false alarm rate exceeds 5%, the threshold is increased by a step size of Δη = 0.1σ;

[0107] When the false negative rate exceeds 5%, the threshold is lowered according to the same step length.

[0108] A transformer fault detection system, used to implement the above-mentioned transformer fault detection method, includes the following modules:

[0109] Data acquisition module: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate multi-physics field raw data sets;

[0110] Signal processing module: performs mechanical resonance characteristic analysis on the multi-physics field raw data set, implements adaptive noise filtering by constructing a resonance frequency offset compensation model, and outputs a denoised feature set;

[0111] Feature fusion module: performs multi-scale feature fusion on the denoising feature set to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features and energy evolution features;

[0112] Feature extraction module: using an improved random fern classifier to dynamically extract a set of fault sensitive factors from the multidimensional feature tensor, wherein the set of fault sensitive factors includes a mechanical deformation sensitive factor, an insulation degradation sensitive factor, and a poor contact sensitive factor;

[0113] Fault diagnosis module: constructs a fault pattern recognition model based on the fault sensitive factor set, and outputs the corresponding fault type, fault level and location information.

[0114] Beneficial effects of the present invention:

[0115] The present invention significantly improves the accuracy and reliability of transformer fault detection through synchronous acquisition and fusion analysis of multiple physical fields. First, time synchronization marking technology is used to achieve microsecond-level synchronous acquisition of vibration, leakage flux and temperature signals, solving the problem of multi-source data timing mismatch in traditional methods; second, by constructing a resonant frequency temperature drift compensation model and an adaptive noise filtering algorithm, signal distortion caused by on-site electromagnetic interference and temperature fluctuations is effectively suppressed.

[0116] The innovative design of the present invention is also reflected in the comprehensiveness and practicality of fault diagnosis. The time-frequency-space-energy multidimensional feature tensor generated by the multi-scale feature fusion module realizes the three-dimensional characterization of the transformer fault. The fault pattern recognition model based on the attention mechanism can not only accurately distinguish the fault type, but also output a quantitative severity assessment of levels 1-5 and locate the fault location. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0118] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0119] Figure 2 2 is a system module diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0120] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0121] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0122] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0123] like Figure 1 As shown, a method for detecting a transformer fault includes the following steps:

[0124] S1: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate a multi-physics field raw data set;

[0125] S2: Analyze the mechanical resonance characteristics of the original multi-physics field data set, implement adaptive noise filtering by building a resonant frequency offset compensation model, and output a denoised feature set;

[0126] S3: Input the denoised feature set into the multi-scale feature fusion module to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features, and energy evolution features;

[0127] S4: Dynamically extract a set of fault-sensitive factors from the multidimensional feature tensor based on an improved random fern classifier. The set of fault-sensitive factors includes mechanical deformation sensitivity factor, insulation degradation sensitivity factor, and poor contact sensitivity factor.

[0128] S5: Construct a fault pattern recognition model based on the set of fault sensitive factors and output the corresponding fault type, fault level and location information.

[0129] S1 includes:

[0130] S11: The vibration acceleration signals of the key structural positions of the core are collected by arranging a three-dimensional MEMS acceleration sensor array. The sampling frequency is set to f a =10kHz, generating the vibration acceleration signal A(t), which is expressed as:

[0131]

[0132] Among them, a xi (t),a yi (t),a zi (t) are the acceleration signals of the i-th sensor in the x / y / z direction, n is the number of acceleration sensors, t is the sampling time point, t = k / f a ,k∈N + ;

[0133] S12: High-sensitivity Hall magnetic flux sensors are placed around the core to collect transient leakage magnetic flux intensity signals and use time synchronization markers T s The time stamp is unified with the acceleration signal to generate the leakage flux intensity signal M(t), which is expressed as:

[0134] M(t)=[m1(t),m2(t),...,m k (t)];

[0135] Among them, m j (t) is the output of the jth flux sensor at time t, k is the number of flux sensors;

[0136] S13: Continuously image the surface temperature field of the transformer through the infrared thermal imaging module, spatially sample the thermal image using the region partition matrix R, and extract the surface temperature field distribution signal T(t), which is expressed as:

[0137] T(t)={T i,j (t)},i=1,…,h; j=1,…,w;

[0138] Among them, T i,j (t) represents the surface temperature value of the (i, j)th pixel area at time t, h, w are the height and width pixels of the image;

[0139] S14: Based on the unified time base T sThe vibration acceleration signal, leakage magnetic flux intensity signal, and surface temperature field distribution signal sources are time-aligned and expressed using a linear interpolation function. The asynchronously sampled signals are resampled to generate a multi-physics field raw data set consisting of three channels, which is expressed as:

[0140] X s (t) = interp(X(t i ),T s );

[0141] D(t)={A(t),M(t),T(t)};

[0142] Among them, X(t i ) represents the original signal sequence, T s The target unified time axis, interp represents the linear interpolation function, X s (t) is the output signal after alignment, and D(t) is the original data set of the multi-physics field.

[0143] S2 includes:

[0144] S21: Extract the frequency domain features of the vibration signal through short-time Fourier transform and combine it with temperature field modeling to compensate for the temperature drift of the resonant frequency;

[0145] S22: Adaptive filtering is used to suppress leakage flux noise interference, and vibration, leakage flux intensity and temperature signals are integrated to generate a multi-dimensional denoising feature set.

[0146] S21 includes:

[0147] S211, Mechanical Resonance Frequency Domain Analysis: Perform short-time Fourier transform (STFT) on each sensor channel in the vibration acceleration signal A(t) to extract the main resonant frequency f r and its harmonic component set {f r1 ,f r2 ,...,f rm}, satisfying P(f r ,τ)≥α.max(P(f,τ)), α is the energy threshold coefficient, and α∈(0,1), calculate the time-frequency energy spectrum matrix P(f,τ), expressed as:

[0148]

[0149] Where w(t-τ) is the Hanning window function, τ is the center position of the time window, f is the frequency component, and i∈{1,2,...,n} represents the sensor number;

[0150] S212, Resonant frequency temperature drift modeling and compensation: Based on the surface temperature field distribution signal T(t), establish the resonant frequency temperature drift model Δf r(t), expressed as:

[0151]

[0152] in, is the average temperature rise, T0 is the reference temperature, β is the static temperature coefficient, and γ is the dynamic temperature coefficient;

[0153] Calculate the resonant frequency after compensation, expressed as:

[0154] f r ′(t)=f r (t)-Δf r (t);

[0155] Among them, f r ′(t) is the resonant frequency after compensation, f r (t) is f r (t) Original main resonant frequency.

[0156] S22 includes:

[0157] S221, Adaptive noise suppression based on LMS algorithm: Construct an adaptive filter based on the LMS algorithm, take the leakage flux intensity signal M(t) as the reference noise input, the vibration signal as the desired signal, and output the filtered image, which is expressed as:

[0158]

[0159] Among them, w p (t+1)=w p (t)+μ·e(t)·m j (tp) is the weight update formula, L is the filter order, μ is the convergence factor, e(t)=a xi (t)-y(t) is the error signal;

[0160] S222, multi-physics field feature fusion and denoising set generation: fusion of the denoising results of each physical field, output of the multi-dimensional denoising feature set F(t), expressed as:

[0161]

[0162] Among them, M′(t)=M(t)-y(t) is the magnetic flux signal after decorrelation, T′(t)=T(t)·H(f r ′(t)) is the frequency-domain weighted temperature field distribution signal, and H(f) is the resonant frequency matching window function.

[0163] S3 includes:

[0164] S31, time-frequency feature extraction: vibration acceleration signal after denoising Perform continuous wavelet transform to extract the peak sequence in the time-frequency energy distribution and generate the time-frequency feature vector W(a,b) that represents the dynamic change of mechanical resonance, which is expressed as:

[0165]

[0166] Among them, ψ(t) is the mother wavelet function (Morlet wavelet is selected), a is the scale parameter, corresponding to the frequency a0 is the scale reference corresponding to the wavelet center frequency, b is the translation parameter, which represents the time offset, and i∈{1,2,...,n} is the sensor photo number;

[0167] Extract the time-frequency energy peak sequence E TF (t)=max a |W(a,b)| 2 , forming the time-frequency feature vector F TF (t);

[0168] S32, spatial distribution feature modeling: perform spatial convolution coding on the temperature field distribution signal T′(t), using the convolution kernel K s Extract local temperature gradient features C ij (t), expressed as:

[0169]

[0170] Among them, K s The Sobel operator of (2d+1)×(2d+1) is used to detect the spatial gradient of the temperature field, d=1, the convolution kernel size is defined as 3×3, h, w are the height and width pixels of the temperature field image;

[0171] The convolution output C(t) is divided into regions to calculate the mean and variance, and generate the spatial distribution feature matrix F SP (t)∈R h ×w×2 ;

[0172] S33, energy evolution characteristics analysis: Through sliding window integration and entropy calculation, the energy accumulation and complexity of the leakage flux intensity signal M′(t) are quantified and expressed as:

[0173]

[0174] Among them, E M (t) is the sliding energy window integral, N = f a ΔT is the window length, ΔT = 0.1s is the integration time window, and k is the number of magnetic flux sensors;

[0175] The energy evolution complexity is quantified by the entropy value, and the energy evolution feature vector is generated, which is expressed as:

[0176]

[0177] F EN (t)=[E M (t),H E (t)];

[0178] Among them, H E (t) is the energy evolution complexity, F EN (t) is the energy evolution characteristic;

[0179] S34, multi-dimensional feature tensor fusion: the time-frequency feature F TF (t), spatial distribution characteristics F SP (t) and energy evolution characteristics F EN (t) Splicing by channel dimension to form a multi-dimensional feature tensor Expressed as:

[0180]

[0181] Flatten(·) means flattening the matrix into a vector, D = dim(F TF )+2hw+2 is the total feature dimension.

[0182] S4 includes:

[0183] S41, Improved Random Fern Classifier Construction: Define dynamic weight random fern classifier, composed of N f fern structure Composition, each fern From the multidimensional feature tensor Randomly select L feature subsets S from m ={f m1 ,f m2 ,...,f mL}, and calculate its dynamic classification weight ω m (t), expressed as:

[0184]

[0185] Among them, λ is the weight adjustment factor (λ>0), Var(S m ) represents the subset S m The characteristic variance, L is the number of features selected for each fern (L≤D, D is The total dimension of ), m is the fern structure index number;

[0186] S42, dynamic extraction of fault-sensitive factors: For three types of fault factors (mechanical deformation, insulation degradation, and poor contact), a fern classifier is used to output probability mapping;

[0187] S43, sensitive factor set generation: Fusion of three types of sensitive factors, output of dynamically updated fault sensitive factor set Q(t), expressed as:

[0188] Q(t)={Q mech (t),Q insul (t),Q contac (t)};

[0189] Each type of sensitivity factor is normalized by the threshold and expressed as:

[0190]

[0191] Among them, μ type ,σ type are the mean and standard deviation of historical data respectively.

[0192] S42 includes:

[0193] S421, Mechanical deformation sensitivity factor: based on time-frequency characteristics F TF (t) and energy evolution characteristics F EN (t), calculate the weighted voting result of the fern classifier, expressed as:

[0194]

[0195] Among them, θ mech is the mechanical deformation sensitivity factor, δ(·) is the indicator function (output 1 when the condition is met, otherwise 0), N f The number of random ferns is set to 50 by default;

[0196] S422, Insulation degradation sensitivity factor: Based on spatial distribution characteristics F SP (t), calculated by convolution feature response, is expressed as:

[0197]

[0198] Among them, Q insul (t) is the insulation degradation sensitivity factor, σ(·) is the Sigmoid function, C ij (t) is the temperature field convolution output;

[0199] S423, poor contact sensitivity factor: combined with the energy entropy H of the leakage flux signal M′(t) E (t), dynamically adjust the fern classifier weight, expressed as:

[0200]

[0201] Among them, Q contac (t) is the poor contact sensitivity factor, and k is the number of leakage flux sensors.

[0202] S5 includes:

[0203] S51, Fault mode feature matrix construction: Fuse the fault sensitive factor set Q(t) with the multi-scale feature tensor F(t) to construct the fault mode feature matrix G(t)∈R N×M , expressed as:

[0204]

[0205] Where W∈R M×D is a trainable weight matrix, D is the total feature dimension, b is the bias vector, N = 3 is the number of fault types (mechanical deformation, insulation degradation, poor contact), M is the feature mapping dimension (M = 64 by default), and || represents the vector concatenation operation;

[0206] S52, hierarchical identification based on attention mechanism: multi-head attention mechanism is used to calculate the probability distribution P of each fault type type and fault level score S level , expressed as:

[0207]

[0208] S level =softmax(W l ·P type +b l );

[0209] Among them, W q ,W k ,W v are the weight matrices corresponding to query, key, and value, respectively. d is the dimension of each attention head. The softmax operation is used to normalize the attention weights and output the probability distribution of each fault type. l is the level weight matrix (level 5 corresponds to 3 types of faults), b l is the bias vector;

[0210] S53, Fault Location Analysis: Analyze spatial positioning and vibration source positioning to obtain positioning information, including:

[0211] (1) Spatial positioning L heat :For temperature field characteristics F SP (t) performs deconvolution operation, which is expressed as:

[0212] L heat =Deconv(F SP (t),K d );

[0213] Among them, K dis a 3×3 deconvolution kernel;

[0214] (2) Vibration source positioning L vib :Based on time-frequency features F TF (t) Calculate the energy center of gravity, expressed as:

[0215]

[0216] Where n is the number of acceleration sensors, W(a i ,b j ) is the wavelet coefficient;

[0217] S54, state identification and early warning: Based on the fused fault type probability, fault level score and location information, a comprehensive decision function Y(t) is constructed, which is expressed as:

[0218] Y(t)=αP type +βS level +γ(L heat +L vib );

[0219] Among them, α, β, γ are weight coefficients, satisfying α+β+γ=1;

[0220] When Y(t)>η (η is the preset threshold, set according to historical data statistics), a graded warning (such as mild, moderate, severe) is triggered, and the fault type, level and location information (temperature abnormality area and vibration source location) are simultaneously output;

[0221] The preset threshold η is set based on historical data statistics, specifically including:

[0222] (1) Initial threshold calculation: Based on historical normal data, calculate the mean μ and standard deviation σ of the comprehensive decision function Y(t), and set the initial threshold η0 = μ + 3σ (covering 99.7% normal operating conditions);

[0223] (2) Dynamic adjustment mechanism: Set a sliding window (1 week of data). Specific adjustment rules include:

[0224] When the false alarm rate exceeds 5%, the threshold is increased by a step size of Δη = 0.1σ;

[0225] When the false negative rate exceeds 5%, the threshold is lowered according to the same step length.

[0226] like Figure 2 As shown, a transformer fault detection system is used to implement the above-mentioned transformer fault detection method, including the following modules:

[0227] Data acquisition module: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate multi-physics field raw data sets;

[0228] Signal processing module: Analyzes the mechanical resonance characteristics of the original multi-physics field data set, implements adaptive noise filtering by building a resonant frequency offset compensation model, and outputs a denoised feature set;

[0229] Feature fusion module: performs multi-scale feature fusion on the denoising feature set to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features, and energy evolution features;

[0230] Feature extraction module: uses an improved random fern classifier to dynamically extract a set of fault sensitive factors from the multidimensional feature tensor. The set of fault sensitive factors includes mechanical deformation sensitivity factor, insulation degradation sensitivity factor, and poor contact sensitivity factor.

[0231] Fault diagnosis module: Constructs a fault pattern recognition model based on a set of fault sensitive factors and outputs the corresponding fault type, fault level and location information.

[0232] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0233] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for detecting a mutual inductor fault, characterized in that: The following steps are involved: S1: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate a multi-physics field raw data set; S2: Analyze the mechanical resonance characteristics of the multi-physics field original data set, implement adaptive noise filtering by constructing a resonance frequency offset compensation model, and output a denoised feature set; S3: Inputting the denoising feature set into a multi-scale feature fusion module to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features, and energy evolution features; S4: dynamically extracting a set of fault sensitive factors from the multidimensional feature tensor based on an improved random fern classifier, wherein the set of fault sensitive factors includes a mechanical deformation sensitive factor, an insulation degradation sensitive factor, and a poor contact sensitive factor; S5: Constructing a fault pattern recognition model according to the set of fault sensitive factors, and outputting corresponding fault type, fault level and location information.

2. A transformer fault detection method according to claim 1, characterized in that: Said S1 comprises: S11: The vibration acceleration signals of the key structural positions of the core are collected by arranging a three-dimensional MEMS acceleration sensor array. The sampling frequency is set to f a =10kHz, generating the vibration acceleration signal A(t), which is expressed as: Among them, a xi (t),a yi (t),a zi (t) are the acceleration signals of the i-th sensor in the x / y / z direction, n is the number of acceleration sensors, t is the sampling time point, t = k / f a ,k∈N + ; S12: High-sensitivity Hall magnetic flux sensors are placed around the core to collect transient leakage magnetic flux intensity signals and use time synchronization markers T s The time stamp is unified with the acceleration signal to generate the leakage flux intensity signal M(t), which is expressed as: M(t)=[m1(t),m2(t),...,m k (t)]; Among them, m j (t) is the output of the jth flux sensor at time t, k is the number of flux sensors; S13: Continuously image the surface temperature field of the transformer through the infrared thermal imaging module, spatially sample the thermal image using the region partition matrix R, and extract the surface temperature field distribution signal T(t), which is expressed as: T(t)={T i,j (t)},i=1,…,h;j=1,…,w; Among them, T i,j (t) represents the surface temperature value of the (i, j)th pixel area at time t, h, w are the height and width pixels of the image; S14: Based on the unified time base T s The vibration acceleration signal, leakage magnetic flux intensity signal, and surface temperature field distribution signal sources are time-aligned and expressed using a linear interpolation function. The asynchronously sampled signals are resampled to generate a multi-physics field raw data set consisting of three channels, which is expressed as: X s (t)=interp(X(t i ),T s ); D(t)={A(t),M(t),T(t)}; Among them, X(t i ) represents the original signal sequence, T s The target unified time axis, interp represents the linear interpolation function, X s (t) is the output signal after alignment, and D(t) is the original data set of the multi-physics field.

3. A transformer fault detection method according to claim 2, characterized in that: The S2 includes: S21: Extract the frequency domain features of the vibration signal through short-time Fourier transform and combine it with temperature field modeling to compensate for the temperature drift of the resonant frequency; S22: Adaptive filtering is used to suppress leakage flux noise interference, and vibration, leakage flux intensity and temperature signals are integrated to generate a multi-dimensional denoising feature set.

4. A transformer fault detection method according to claim 3, characterized in that: The S21 includes: S211, Mechanical Resonance Frequency Domain Analysis: Perform short-time Fourier transform (STFT) on each sensor channel in the vibration acceleration signal A(t) to extract the main resonant frequency f r and its harmonic component set {f r1 ,f r2 ,...,f rm }, satisfying P(f r ,τ)≥α.max(P(f,τ)), α is the energy threshold coefficient, and α∈(0,1), calculate the time-frequency energy spectrum matrix P(f,τ), expressed as: Where w(t-τ) is the Hanning window function, τ is the center position of the time window, f is the frequency component, and i∈{1,2,...,n} represents the sensor number; S212, Resonant frequency temperature drift modeling and compensation: Based on the surface temperature field distribution signal T(t), establish the resonant frequency temperature drift model Δf r (t), expressed as: in, is the average temperature rise, T0 is the reference temperature, β is the static temperature coefficient, and γ is the dynamic temperature coefficient; Calculate the resonant frequency after compensation, expressed as: f r ′(t)=f r (t)-Δf r (t); Among them, f r ′(t) is the resonant frequency after compensation, f r (t) is f r (t) Original main resonant frequency.

5. A transformer fault detection method according to claim 3, characterized in that: The S22 includes: S221, Adaptive noise suppression based on LMS algorithm: Construct an adaptive filter based on the LMS algorithm, take the leakage flux intensity signal M(t) as the reference noise input, the vibration signal as the desired signal, and output the filtered image, which is expressed as: Among them, w p (t+1)=w p (t)+μ·e(t)·m j (tp) is the weight update formula, L is the filter order, μ is the convergence factor, e(t)=a xi (t)-y(t) is the error signal; S222, multi-physics field feature fusion and denoising set generation: fusion of the denoising results of each physical field, output of the multi-dimensional denoising feature set F(t), expressed as: Among them, M′(t)=M(t)-y(t) is the magnetic flux signal after decorrelation, T′(t)=T(t)·H(f r ′(t)) is the frequency-domain weighted temperature field distribution signal, and H(f) is the resonant frequency matching window function.

6. A transformer fault detection method according to claim 5, characterized in that: The S3 includes: S31, time-frequency feature extraction: vibration acceleration signal after denoising Perform continuous wavelet transform to extract the peak sequence in the time-frequency energy distribution and generate the time-frequency feature vector W(a,b) that represents the dynamic change of mechanical resonance, which is expressed as: Among them, ψ(t) is the mother wavelet function (Morlet wavelet is selected), a is the scale parameter, corresponding to the frequency a0 is the scale reference corresponding to the wavelet center frequency, b is the translation parameter, which represents the time offset, and i∈{1,2,...,n} is the sensor photo number; Extract the time-frequency energy peak sequence E TF (t)=max a |W(a,b)| 2 , forming the time-frequency feature vector F TF (t); S32, spatial distribution feature modeling: perform spatial convolution coding on the temperature field distribution signal T′(t), using the convolution kernel K s Extract local temperature gradient features C ij (t), expressed as: Among them, K s The Sobel operator of (2d+1)×(2d+1) is used to detect the spatial gradient of the temperature field, d=1, the convolution kernel size is defined as 3×3, h, w are the height and width pixels of the temperature field image; Divide the convolution output C(t) by region to calculate the mean and variance, and generate the spatial distribution feature matrix F SP (t)∈R h×w×2 ; S33, energy evolution characteristics analysis: Through sliding window integration and entropy calculation, the energy accumulation and complexity of the leakage flux intensity signal M′(t) are quantified and expressed as: Among them, E M (t) is the sliding energy window integral, N = f a ΔT is the window length, ΔT = 0.1s is the integration time window, and k is the number of magnetic flux sensors; The energy evolution complexity is quantified by the entropy value, and the energy evolution feature vector is generated, which is expressed as: F EN (t)=[E M (t),H E (t)]; Among them, H E (t) is the energy evolution complexity, F EN (t) is the energy evolution characteristic; S34, multi-dimensional feature tensor fusion: the time-frequency feature F TF (t), spatial distribution characteristics F SP (t) and energy evolution characteristics F EN (t) Splicing by channel dimension to form a multi-dimensional feature tensor Expressed as: Flatten(·) means flattening the matrix into a vector, D = dim(F TF )+2hw+2 is the total feature dimension.

7. A transformer fault detection method according to claim 6, characterized in that: The S4 includes: S41, Improved Random Fern Classifier Construction: Define dynamic weight random fern classifier, composed of N f fern structure Composition, each fern From the multidimensional feature tensor Randomly select L feature subsets S from m ={f m1 ,f m2 ,...,f mL }, and calculate its dynamic classification weight ω m (t), expressed as: Among them, λ is the weight adjustment factor (λ>0), Var(S m ) represents the subset S m The characteristic variance, L is the number of features selected for each fern (L≤D, D is The total dimension of ), m is the fern structure index number; S42, dynamic extraction of fault-sensitive factors: For three types of fault factors (mechanical deformation, insulation degradation, and poor contact), a fern classifier is used to output probability mapping; S43, sensitive factor set generation: Fusion of three types of sensitive factors, output of dynamically updated fault sensitive factor set Q(t), expressed as: Q(t)={Q mech (t),Q insul (t),Q contac (t)}; Each type of sensitivity factor is normalized by the threshold and expressed as: Among them, μ type ,σ type are the mean and standard deviation of historical data respectively.

8. A transformer fault detection method according to claim 7, characterized in that: The S42 includes: S421, Mechanical deformation sensitivity factor: based on time-frequency characteristics F TF (t) and energy evolution characteristics F EN (t), calculate the weighted voting result of the fern classifier, expressed as: Among them, θ mech is the mechanical deformation sensitivity factor, δ(·) is the indicator function (output 1 when the condition is met, otherwise 0), N f The number of random ferns is set to 50 by default; S422, Insulation degradation sensitivity factor: Based on spatial distribution characteristics F SP (t), calculated by convolution feature response, is expressed as: Among them, Q insul (t) is the insulation degradation sensitivity factor, σ(·) is the Sigmoid function, C ij (t) is the temperature field convolution output; S423, poor contact sensitivity factor: combined with the energy entropy H of the leakage flux signal M′(t) E (t), dynamically adjust the fern classifier weight, expressed as: Among them, Q contac (t) is the poor contact sensitivity factor, and k is the number of leakage flux sensors.

9. A transformer fault detection method according to claim 7, characterized in that: The S5 includes: S51, Fault mode feature matrix construction: Fault sensitive factor set Q(t) is integrated with multi-scale feature tensor F(t) to construct fault mode feature matrix G(t)∈R N×M , expressed as: Where W∈R M×D is a trainable weight matrix, D is the total feature dimension, b is the bias vector, N = 3 is the number of fault types (mechanical deformation, insulation degradation, poor contact), M is the feature mapping dimension (M = 64 by default), and || represents the vector concatenation operation; S52, hierarchical identification based on attention mechanism: multi-head attention mechanism is used to calculate the probability distribution P of each fault type type and fault level score S level , expressed as: S level =softmax(W l ·P type +b l ); Among them, W q ,W k ,W v are the weight matrices corresponding to query, key, and value, respectively. d is the dimension of each attention head. The softmax operation is used to normalize the attention weights and output the probability distribution of each fault type. l is the level weight matrix (level 5 corresponds to 3 types of faults), b l is the bias vector; S53, Fault Location Analysis: Analyze spatial positioning and vibration source positioning to obtain positioning information, including: (1) Spatial positioning L heat :For temperature field characteristics F SP (t) performs deconvolution operation, which is expressed as: L heat =Deconv(F SP (t),K d ); Among them, K d is a 3×3 deconvolution kernel; (2) Vibration source positioning L vib :Based on time-frequency features F TF (t) Calculate the energy center of gravity, expressed as: Where n is the number of acceleration sensors, W(a i ,b j ) is the wavelet coefficient; S54, state identification and early warning: Based on the fused fault type probability, fault level score and location information, a comprehensive decision function Y(t) is constructed, which is expressed as: Y(t)=αP type +βS level +γ(L heat +L vib ); Among them, α, β, γ are weight coefficients, satisfying α+β+γ=1; When Y(t)>η (η is the preset threshold, set according to historical data statistics), a graded warning (such as mild, moderate, severe) is triggered, and the fault type, level and location information (temperature abnormality area and vibration source location) are simultaneously output; The preset threshold η is set based on historical data statistics, specifically including: (1) Initial threshold calculation: Based on historical normal data, calculate the mean μ and standard deviation σ of the comprehensive decision function Y(t), and set the initial threshold η0 = μ + 3σ (covering 99.7% normal operating conditions); (2) Dynamic adjustment mechanism: Set a sliding window (1 week of data). Specific adjustment rules include: When the false alarm rate exceeds 5%, the threshold is increased by a step size of Δη = 0.1σ; When the false negative rate exceeds 5%, the threshold is lowered according to the same step length.

10. A transformer fault detection system, used to implement a transformer fault detection method according to any one of claims 1 to 9, characterized in that: Includes the following modules: Data acquisition module: Synchronously collects transformer core vibration acceleration signals, leakage flux intensity signals, and surface temperature field distribution signals to generate multi-physics field raw data sets; Signal processing module: performs mechanical resonance characteristic analysis on the multi-physics field raw data set, implements adaptive noise filtering by constructing a resonance frequency offset compensation model, and outputs a denoised feature set; Feature fusion module: performs multi-scale feature fusion on the denoising feature set to generate a multi-dimensional feature tensor including time-frequency features, spatial distribution features and energy evolution features; Feature extraction module: using an improved random fern classifier to dynamically extract a set of fault sensitive factors from the multidimensional feature tensor, wherein the set of fault sensitive factors includes a mechanical deformation sensitive factor, an insulation degradation sensitive factor, and a poor contact sensitive factor; Fault diagnosis module: constructs a fault pattern recognition model based on the fault sensitive factor set, and outputs the corresponding fault type, fault level and location information.

Images