Method for detecting dissolved gas in transformer oil

Abstract

The invention provides a method for detecting dissolved gas in transformer oil, and belongs to the technical field of dissolved gas in transformer oil, and the method comprises the following steps: firstly, transmitting a plurality of groups of ultrasonic waves with different frequencies into the transformer oil at the same time, obtaining and preprocessing a joint spectrum, and then utilizing a wavelet multi-scale decomposition technology considering a time sliding window to obtain a pre-processed joint spectrum; the signal is disassembled into a stable spectrum and three levels of fluctuation spectrums. Extracting the characteristics of interference, spectrum attenuation, waveform complexity, coupling, energy distribution and the like from each level of fluctuation spectrum, and calculating the influence characteristics of microbubbles on ultrasonic waves, including acoustic attenuation, propagation velocity, scattering, acoustic impedance change and nonlinear acoustic characteristics. By performing weight fusion on the three-level fluctuation spectrum energy ratio, comprehensive characteristics are generated and input into a gas calculation model, so that the state of gas in transformer oil is accurately evaluated and diagnosed, and the technical problem that an existing offline sampling analysis method is difficult to meet the requirement of a power system for evaluating the running state of a transformer in real time is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of dissolved gases in transformer oil, and specifically relates to a method for detecting dissolved gases in transformer oil. Background Art

[0002] Transformers are important electrical equipment in the power system and are widely used in the fields of power transmission and distribution. The safe and reliable operation of transformers is crucial for the stable operation of the power grid. As an insulating medium and a cooling medium, the quality of transformer oil directly affects the operating performance of transformers. The detection of dissolved gases in transformer oil has always been one of the important bases for judging the health status of transformers.

[0003] The existing technologies for detecting dissolved gases in transformer oil mainly include gas chromatography, infrared spectroscopy, etc. Gas chromatography can accurately determine the concentrations of various dissolved gases by separating and analyzing the extracted transformer oil samples. However, this method requires laboratory conditions, the detection process is complex, and the detection cycle is relatively long. Infrared spectroscopy can qualitatively and quantitatively analyze the components of dissolved gases by measuring the absorption characteristics of transformer oil samples in specific infrared bands. However, this method has high requirements for instrument accuracy and sample preparation, and at the same time, it cannot distinguish different types of dissolved gases.

[0004] In addition, with the intelligent development of the power system, higher requirements are put forward for the real-time and continuous monitoring of transformer status. The existing off-line sampling and analysis methods are difficult to meet the needs of the power system for real-time assessment of the operating status of transformers. Therefore, it is urgent to develop a new technology for quickly, accurately, and on-line monitoring the dissolved gases in transformer oil. Summary of the Invention

[0005] In view of this, the present invention provides a method for detecting dissolved gases in transformer oil, which can solve the technical problem that the existing off-line sampling and analysis methods are difficult to meet the needs of the power system for real-time assessment of the operating status of transformers.

[0006] The present invention is implemented as follows:

[0007] The present invention provides a method for detecting dissolved gases in transformer oil, including the following steps:

[0008] S10. Simultaneously emit multiple groups of ultrasonic waves with different frequencies into the transformer oil to obtain a multi-ultrasonic combined spectrum of the transformer oil;

[0009] S20. Preprocess the multi-ultrasonic combined spectrum to remove noise interference signals and obtain a preprocessed combined spectrum;

[0010] S30. Decompose the preprocessed combined spectrum using a wavelet multi-scale decomposition method considering a time-sliding window to obtain a stable spectrum, a first-level fluctuation spectrum, a second-level fluctuation spectrum, and a third-level fluctuation spectrum, where the size of the time-sliding window is adaptively determined according to the frequency characteristics and signal dynamic range of the preprocessed combined spectrum;

[0011] S40. Use a preset set of equations for extracting fluctuation spectrum features to extract features from the three levels of fluctuation spectra respectively, obtaining a set of fluctuation spectrum features corresponding to each level of the fluctuation spectrum. The set of fluctuation spectrum features includes interference features, spectral attenuation features, waveform complexity features, coupling features, and energy distribution features;

[0012] S50. For the set of features corresponding to each level of the fluctuation spectrum, use a preset set of equations for the influence of microbubbles on ultrasonic waves to calculate, obtaining a set of microbubble influence features corresponding to each level of the fluctuation spectrum, including acoustic wave attenuation features, acoustic wave propagation speed features, acoustic wave scattering features, acoustic impedance change features, and nonlinear acoustic features;

[0013] S60. Calculate the energy ratio of each level of the fluctuation spectrum as a weight, and fuse the sets of features corresponding to the energy ratios of the three levels of fluctuations to obtain fused features;

[0014] S70. Input the fused features into a pre-trained gas calculation model in transformer oil to obtain the gas calculation result in transformer oil.

[0015] Based on the above technical solution, a method for detecting dissolved gases in transformer oil according to the present invention can also be improved as follows:

[0016] Among them, the set of equations for extracting fluctuation spectrum features includes an interference feature equation, a spectral attenuation equation, a waveform complexity equation, a nonlinear interaction equation, and an energy distribution equation.

[0017] Further, the interference feature equation is used to describe the superposition and interaction of ultrasonic waveforms with different frequencies;

[0018] The spectral attenuation equation is used to calculate the energy attenuation law of ultrasonic waves when propagating in transformer oil;

[0019] The waveform complexity equation is used to quantify the structural complexity of the waveform;

[0020] The nonlinear interaction equation is used to describe the nonlinear coupling relationship between ultrasonic waves;

[0021] The energy distribution equation is used to analyze the distribution characteristics of energy in transformer oil at different frequencies.

[0022] Furthermore, the system of equations for the influence of microbubbles on ultrasonic waves includes: an acoustic wave attenuation characteristic equation, an acoustic wave propagation velocity characteristic equation, an acoustic wave scattering characteristic equation, an acoustic impedance change characteristic equation, and a nonlinear acoustics characteristic equation.

[0023] Furthermore, the acoustic wave attenuation characteristic equation is used to describe that dissolved gases and microbubbles will significantly increase the attenuation of ultrasonic waves in oil;

[0024] The acoustic wave propagation velocity characteristic equation is used to describe that gases and microbubbles will change the propagation velocity of acoustic waves in oil;

[0025] The acoustic wave scattering characteristic equation is used to describe that microbubbles will cause strong scattering and reflection of acoustic waves;

[0026] The acoustic impedance change characteristic equation is used to describe that dissolved gases will change the acoustic impedance of oil;

[0027] The nonlinear acoustics characteristic equation is used to describe that gases and microbubbles will cause the nonlinear propagation effect of ultrasonic waves.

[0028] Furthermore, the gas calculation model in transformer oil adopts a deep residual attention network structure, which is specifically composed of multiple residual blocks, a self-attention mechanism, and a fully connected layer. The network can adaptively learn the complex characteristics of different-scale fluctuation spectra.

[0029] Furthermore, the method for obtaining the training dataset of the gas calculation model in transformer oil is specifically to simulate transformer oil samples with different gas concentrations and types in the laboratory, and use high-precision ultrasonic detection equipment for multi-spectrum joint acquisition to ensure that the dataset covers a wide range of gas states and concentration changes.

[0030] Furthermore, the transformer oil samples include dissolved gas samples simulating different gas concentrations and types, specifically including: hydrogen (H2), methane (CH4), carbon monoxide (CO), carbon dioxide (CO2), acetylene (C2H2), and ethylene (C2H4).

[0031] Furthermore, the gas concentration range covers 0 - 1000 ppm and is sampled at logarithmic intervals or geometric progressions.

[0032] Furthermore, the high-precision ultrasonic detection equipment has the following technical indicators: (1) Frequency range: 100 kHz - 10 MHz; (2) Dynamic range: > 100 dB; (3) Signal-to-noise ratio: > 80 dB; (4) Time resolution: < 1 ns.

[0033] Furthermore, the steps for training the gas calculation model in transformer oil are specifically:

[0034] 1) Data preprocessing and standardization; 2) Model architecture design and parameter initialization; 3) Adopt a multi-scale cross-entropy loss function; 4) Use the Adam optimizer with an adaptive learning rate; 5) Hierarchical cross-validation and model performance evaluation; 6) Model parameter fine-tuning and generalization ability verification.

[0035] Furthermore, for the wavelet multi-scale decomposition of the multi-ultrasound joint spectrum of transformer oil, the equations for decomposing the stable spectrum and the three-level fluctuation spectrum can be expressed as follows:

[0036] Equations for stable spectrum and fluctuation spectrum decomposition:

[0037]

[0038] In the formula, S decomposition (t) is the original ultrasonic signal; D j,k is the j-th level detail coefficient; ψ j,k (t) is the wavelet basis function; R J (t) is the final approximation coefficient (stable spectrum); J is the maximum decomposition level.

[0039] The specific decomposition method is as follows:

[0040] 1. Equation for obtaining the stable spectrum:

[0041] In the formula, c J,k is the scale coefficient; φ J,k (t) is the scaling function.

[0042] 2. Equation for obtaining the first-level fluctuation spectrum: D 1,k = ∑ n s(n)·h1(2k - n);

[0043] 3. Equation for obtaining the second-level fluctuation spectrum: D 2,k = ∑ n D 1,n ·h2(2k - n);

[0044] 4. Equation for obtaining the third-level fluctuation spectrum: d 3,k = ∑ n D 2,n ·h3(2k - n);

[0045] In the formula, h1, h2, h3 are high-pass filters of different levels; s(n) is the original signal; k is the discrete time index.

[0046] Equations for calculating the energy of the stable spectrum and the fluctuation spectrum:

[0047] E R (J) = ∑ j |R J (k)| 2; E D,j (j) = ∑ k |D j,k | 2 ;

[0048] Wherein, E R (J) is the stable spectrum energy; E D,j (j) is the j-th level fluctuation spectrum energy.

[0049] Fluctuation spectrum inter-level relationship equation:

[0050] Wherein, σ D,j is the energy percentage of the j-th level fluctuation spectrum relative to the stable spectrum.

[0051] These equations are designed considering the following principles: Multi-scale decomposition can reveal the characteristics of the signal at different scales; Wavelet transform can retain both time and frequency information; Gradual decomposition can capture the subtle changes of the signal; The stable spectrum reflects the basic characteristics of the signal; The fluctuation spectrum reflects the dynamic changes of the signal.

[0052] The detailed description of the fluctuation spectrum feature extraction equation set is as follows:

[0053] 1. Interference feature equation:

[0054] Wherein, A i (t) is the amplitude of the i-th ultrasonic wave; φ i is the phase; n is the number of ultrasonic wave frequencies.

[0055] 2. Spectrum attenuation equation: S attenuation (f) = S0(f) · e -α(f)·d ;

[0056] Wherein, S0(f) is the initial spectrum amplitude; α(f) is the frequency-dependent attenuation coefficient; d is the propagation distance.

[0057] 3. Waveform complexity equation:

[0058] Wherein, p i is the probability distribution at signal i; m is the number of discrete intervals.

[0059] 4. Nonlinear interaction equation:

[0060] Wherein, γ k is the nonlinear coupling coefficient; K is the number of frequency modes.

[0061] 5. Energy distribution equation:

[0062] Where X(f,t) is the time-frequency domain signal; T is the total time.

[0063] The equations for the influence of microbubbles on ultrasonic waves are described in detail as follows:

[0064] 1. Acoustic attenuation characteristic equation: A attenuation (f) = A0(f)·(1 + β·C g ) -γ ;

[0065] Where A0(f) is the initial acoustic wave amplitude; C g is the gas concentration; β, γ are empirical coefficients.

[0066] 2. Acoustic wave propagation velocity characteristic equation: v sound (C g ) = v0·(1 - δ·C g ) -1 / 2 ;

[0067] Where v0 is the reference sound speed in oil; δ is the gas influence coefficient.

[0068] 3. Acoustic wave scattering characteristic equation:

[0069] Where σ0 is the initial scattering cross-section; r is the propagation distance; r0 is the reference distance.

[0070] 4. Acoustic impedance change characteristic equation: Z acoustic (C g ) = Z0·(1 + η·C g );

[0071] Where Z0 is the reference acoustic impedance; η is the gas concentration influence factor.

[0072] 5. Nonlinear acoustics characteristic equation:

[0073] Where A0 is the reference acoustic wave amplitude; A threshold is the threshold amplitude; ζ, κ are nonlinear parameters.

[0074] The energy ratio weight sum and feature fusion equation are described in detail as follows:

[0075]

[0076] Where E j is the energy of the j-th level fluctuation spectrum; F j is the corresponding feature group; F fusion is the fusion feature.

[0077] Furthermore, in the step of obtaining the training data set of the gas calculation model in transformer oil, a new type of dissolved gas collection unit is also used for detection, including using two Ni / Zn@SnO2 gas-sensitive sensing elements for collecting dissolved gas. The Ni / Zn@SnO2 gas-sensitive sensing element consists of an Si-based bottom layer, an SiO2 insulating layer, a heating electrode layer, an SiNi isolation layer, a test electrode layer, and an SnO2-based sensing layer from bottom to top in sequence. Among them, the difference between these two elements is that the heating electrode layer of the first element is formed by a single heating channel, and the single heating channel is deployed in a repeated arrangement of U-inverted U shapes, and the distance between the two side arms of adjacent U shapes is 25 μm; the heating electrode layer of the second element is heated by two comb-shaped heating channels arranged oppositely at intervals, and the distance between the comb teeth of the comb shape is 5 μm. During collection, a collection unit is set at multiple positions in the transformer. Each collection unit includes two first elements and two second elements, forming the four corners of a square in an arrangement. The two first elements are diagonal, and the two second elements are also diagonal.

[0078] The present invention also constructs a highly refined Ni / Zn@SnO2 gas sensor array collection unit. Its microstructure stacks an Si-based bottom layer (300 - 500 μm), an SiO2 insulating layer (0.5 - 2 μm), a heating electrode layer, an SiNi isolation layer (50 - 200 nm), a test electrode layer, and an SnO2-based sensing layer from bottom to top. Among them, two types of heterogeneous heating electrodes are innovatively designed: the first type of sensing element uses a U-inverted U-shaped heating channel, and the distance between the two side arms of adjacent U shapes is precisely controlled to be 25 μm; the second type of sensing element uses a comb-shaped heating channel, and the distance between the comb teeth is precisely designed to be 5 μm, and is deployed in a diagonal array at the four corners of a square. Moreover, the SnO2-based sensing layer co-doped with Ni / Zn can highly sensitively and selectively detect key dissolved gases such as hydrogen, methane, carbon monoxide, acetylene, and ethylene in transformer oil within the working temperature range of 200 - 400 °C, providing an innovative gas detection solution at the microscale for the condition diagnosis of power equipment.

[0079] Compared with the prior art, the beneficial effects of a method for detecting dissolved gases in transformer oil provided by the present invention are as follows:

[0080] 1. Fast detection speed: By synchronously transmitting and collecting multiple groups of ultrasonic signals with different frequencies, the multi-ultrasonic combined spectrum of transformer oil can be quickly obtained, providing basic data for subsequent gas analysis and greatly shortening the detection cycle.

[0081] 2. High detection accuracy: Using wavelet multi-scale analysis technology to extract multi-level fluctuation spectrum features and combining with the microbubble physical model to calculate various gas characteristic parameters can more comprehensively and accurately reflect the concentration state of various dissolved gases in transformer oil.

[0082] 3. Strong detection continuity: This method realizes the online monitoring of dissolved gases in transformer oil, can continuously and real-time detect the change of oil quality state, and provides important support for the health state assessment and fault early warning of transformers.

[0083] 4. Low cost: This method does not require complex oil sample treatment and laboratory analysis equipment. Only the ultrasonic sensor and calculation unit installed on the transformer are needed to complete gas detection, and the equipment investment and operation cost are relatively low.

[0084] In summary, the method for detecting dissolved gases in transformer oil based on the multi-frequency ultrasonic combined spectrum proposed by the present invention can overcome the deficiencies of the existing technology, realize fast, accurate and continuous online monitoring, provide valuable basis for the assessment of transformer operation status, and solve the technical problem that the existing off-line sampling analysis method is difficult to meet the real-time assessment requirements of the power system for the operation status of transformers. Brief Description of the Drawings

[0085] Figure 1 is the flowchart of the method provided by the present invention;

[0086] Figure 2 is the schematic diagram of the micro-structure of the acquisition unit;

[0087] Figure 3 is the schematic diagram of the U-inverted U-shaped heating electrode;

[0088] Figure 4 is the schematic diagram of the hair comb-shaped heating electrode;

[0089] Figure 5 is the circuit schematic diagram of the gas sensor array acquisition unit;

[0090] Figure 6 is the schematic diagram of the finished product of the gas sensor array acquisition unit. Detailed Embodiment

[0091] As Figure 1 shown, it is the flowchart of a method for detecting dissolved gases in transformer oil provided by the present invention. The following is a detailed description of the specific implementation of the steps in the present invention:

[0092] The specific implementation of step S10 is: This step aims to obtain the multi-ultrasonic combined spectrum of transformer oil. It specifically includes the following sub-steps:

[0093] 1. Select multiple groups of ultrasonic signals with different frequencies. Let the selected frequencies be f1, f2,..., f n , where f i represents the i-th frequency component.

[0094] 2. Transmit ultrasonic signals of these different frequencies to the transformer oil sample simultaneously. For the i-th frequency component f i , the ultrasonic signal transmitted can be expressed as x i (t) = A i (t)cos(2πf i t + φ i (t)), where A i (t) is the amplitude, and φ i (t) is the phase.

[0095] 3. Collect the response signals of ultrasonic waves of each frequency in the transformer oil. For the i-th frequency component f i , the response signal in the transformer oil can be expressed as y i (t) = B i (t)cos(2πf i t + θ i (t)), where B i (t) is the amplitude, and θ i (t) is the phase.

[0096] 4. Perform spectral analysis on the collected multi-spectrum signals. Convert the time-domain signal y i (t) into a frequency-domain spectral shape Y i (f) through fast Fourier transform (FFT) to obtain the multi-ultrasonic joint spectrum {Y1(f), Y2(f),..., Y n (f)} of the transformer oil.

[0097] In summary, in step S10, by simultaneously transmitting multiple groups of ultrasonic waves of different frequencies and collecting the response signals, the multi-ultrasonic joint spectrum {Y1(f), Y2(f),..., Y n (f)} of the transformer oil is obtained. This lays a foundation for subsequent signal processing and feature extraction.

[0098] The specific implementation manner of step S20 is as follows: The purpose of this step is to preprocess the obtained multi-ultrasonic joint spectrum, remove the noise interference signals, and obtain a reliable preprocessed joint spectrum. It specifically includes the following sub-steps:

[0099] 1. Perform window function processing on the collected multi-spectrum signals. Use the Hamming window function where n is the discrete time point and N is the signal length.

[0100] 2. Apply a median filter to remove impulse noise.

[0101] 3. Use a band-pass filter to eliminate low-frequency and high-frequency noises.

[0102] 4. Normalize the preprocessed multi-spectrum signal. The normalization formula is where x min and x max are the minimum and maximum values of the signal respectively.

[0103] Through the above preprocessing steps, the noise interference in the multi-ultrasound joint spectrum can be effectively eliminated, and a clean and reliable preprocessed joint spectrum can be obtained.

[0104] The specific implementation of step S30 is as follows: This step aims to perform wavelet multi-scale decomposition on the preprocessed joint spectrum to extract the feature information at different time scales. It specifically includes the following sub-steps:

[0105] 1. Select the Daubechies wavelet (db4) as the basis function. The Daubechies wavelet has good time-frequency locality and orthogonality, and can effectively decompose and reconstruct signals.

[0106] 2. Determine the maximum level J of wavelet decomposition. The approximation coefficient R J (t) at the Jth level is the stable spectrum, which reflects the basic characteristics of the signal. The detailed coefficient D j,k at the jth level constitutes the wave spectra at different levels.

[0107] 3. Perform wavelet decomposition using a time-sliding window. The wavelet decomposition formula is where ψ j,k (t) is the wavelet basis function.

[0108] 4. Calculate the energy of the wave spectra at each layer. The energy of the stable spectrum is E R (J) = ∑ k |R J (k)| 2 , and the energy of the wave spectrum at the jth level is E D,j (j) = ∑ k |D j,k | 2 .

[0109] 5. Evaluate the relative energy proportion of the wave spectra at each layer. The percentage of the energy of the wave spectrum at the jth level relative to the stable spectrum is

[0110] Through wavelet multi-scale decomposition, the feature information of the multi-ultrasound joint spectrum of transformer oil at different time scales can be effectively extracted, including the stable spectrum and multi-level wave spectra. This lays a foundation for subsequent feature extraction and fusion.

[0111] The specific implementation of step S40 is as follows: The purpose of this step is to extract features from the extracted stable spectrum and fluctuation spectrum to obtain the feature groups corresponding to the fluctuation spectra at each level. It specifically includes the following sub-steps:

[0112] 1. Extract the interference feature. The interference feature equation is where A i (t) and φ i are the amplitude and phase of the i-th frequency component respectively.

[0113] 2. Extract the spectral attenuation feature. The spectral attenuation equation is S attenuation (f) = S0(f)·e -α(f)·d , where S0(f) is the initial spectral amplitude, α(f) is the frequency-dependent attenuation coefficient, and d is the propagation distance.

[0114] 3. Extract the waveform complexity feature. The waveform complexity equation is where p i is the probability distribution at signal i, and m is the number of discrete intervals.

[0115] 4. Extract the non-linear interaction feature. The non-linear interaction equation is where γ k is the non-linear coupling coefficient, and K is the number of frequency modes.

[0116] 5. Extract the energy distribution feature. The energy distribution equation is where X(f,t) is the time-frequency domain signal, and T is the total time.

[0117] Through the above 5 types of feature extraction equations, the interference features, spectral features, waveform features, non-linear features, and energy features of the multi-ultrasonic combined spectrum of transformer oil at different time scales can be comprehensively characterized. These feature groups reflect multiple dimensions of information on the gas state in transformer oil.

[0118] The specific implementation of step S50 is as follows: The purpose of this step is to further calculate the microbubble influence feature groups corresponding to the fluctuation spectrum features by using the physical model of the influence of microbubbles on ultrasonic waves. It specifically includes the following sub-steps:

[0119] 1. Calculate the acoustic wave attenuation feature. The acoustic wave attenuation feature equation is A attenuation (f) = A0(f)·(1 + β·C g ) -γ , where A0(f) is the initial acoustic wave amplitude, C g is the gas concentration, and β and γ are empirical coefficients.

[0120] 2. Calculate the characteristics of the acoustic wave propagation velocity. The characteristic equation of the acoustic wave propagation velocity is v sound (C g ) = v0·(1 - δ·C g ) -1 / 2 , where v0 is the reference sound velocity in oil and δ is the gas influence coefficient.

[0121] 3. Calculate the characteristics of the acoustic wave scattering. The characteristic equation of the acoustic wave scattering is where σ0 is the initial scattering cross-section, r is the propagation distance, and r0 is the reference distance.

[0122] 4. Calculate the characteristics of the acoustic impedance change. The characteristic equation of the acoustic impedance change is Z acoustic (C g ) = Z0·(1 + η·C g ), where Z0 is the reference acoustic impedance and η is the gas concentration influence factor.

[0123] 5. Calculate the characteristics of the nonlinear acoustics. The characteristic equation of the nonlinear acoustics is where A0 is the reference acoustic wave amplitude, A threshold is the threshold amplitude, and ζ and κ are the nonlinear parameters.

[0124] Through the above five types of microbubble influence models, the microbubble influence characteristic groups corresponding to the wave spectrum characteristics can be calculated from multiple angles such as acoustic wave attenuation, propagation velocity, scattering, acoustic impedance, and nonlinearity. These characteristic groups provide an important physical basis for subsequent gas calculations.

[0125] The specific implementation manner of step S60 is: the purpose of this step is to perform weighted fusion on the wave spectrum characteristics under different time scales to obtain the final fusion characteristics. It specifically includes the following sub-steps:

[0126] 1. Calculate the energy ratio weights of each layer of the wave spectrum. The calculation formula for the energy ratio weights is where E j is the energy of the j-th layer of the wave spectrum.

[0127] 2. Use the weighted fusion equation to linearly weight and fuse the characteristic groups F j corresponding to different layers of the wave spectrum according to the energy ratio weights W energy,j to obtain the final fusion characteristic F fusion .

[0128] Through this step, the five types of characteristics extracted from the stable spectrum and each layer of the wave spectrum are fused according to their energy distribution in the signal. This weighted fusion method can adaptively extract the most representative characteristic combination and provide a more comprehensive and reliable input for subsequent gas calculations.

[0129] The specific implementation of step S70 is as follows: The purpose of this step is to use a pre-trained Deep Residual Attention Network (ResANet) model to input the fusion features and obtain the calculation results of the gases in the transformer oil. It specifically includes the following sub-steps:

[0130] 1. The structure of the ResANet model can be expressed as where is the mapping function of the network, and Θ is the parameter to be learned. The network consists of multiple residual blocks, self-attention mechanisms, and fully connected layers.

[0131] 2. The loss function uses multi-scale cross-entropy loss where y ij is the actual gas concentration label, is the model prediction result.

[0132] 3. The Adam optimizer with an adaptive learning rate is used for model training, and the optimization objective is The Adam algorithm can automatically adjust the learning rate of each parameter, improving the convergence speed and robustness.

[0133] 4. Hierarchical K-fold cross-validation is adopted to evaluate the model performance. The main indicators include prediction accuracy, generalization ability, etc. Ensure that the model can adapt to the gas detection requirements of different transformer oil samples.

[0134] 5. For specific transformer oil samples, fine-tune the model parameters and verify their generalization performance. This can further improve the prediction accuracy of the model in practical applications.

[0135] In summary, step S70 uses the pre-trained ResANet model, takes the fusion features extracted in step S60 as the input, and outputs the calculation results of the concentrations of various gases in the transformer oil. This model has an adaptive learning ability and can accurately and reliably detect the gases in the transformer oil.

[0136] Specifically, the principle of the present invention is as follows: By using the propagation characteristics of multi-frequency ultrasonic waves in transformer oil, combining wavelet multi-scale analysis and microbubble physical models, characteristic parameters that can accurately reflect the dissolved gas state in transformer oil are extracted.

[0137] First, this method simultaneously emits 5-10 groups of ultrasonic signals with different frequencies to the transformer oil sample to obtain the multi-ultrasonic joint spectrum of the transformer oil. When these ultrasonic waves with different frequencies propagate in the oil, they will be affected by the dissolved gases and microbubbles, manifested as amplitude attenuation, phase change, spectral characteristic change, etc. By synchronously collecting these ultrasonic signals, a multi-ultrasonic joint spectrum reflecting the complex physical characteristics of the transformer oil can be obtained.

[0138] Secondly, the joint spectrum is disassembled using wavelet multi-scale analysis technology. Wavelet transform can retain both time and frequency information simultaneously. By gradually decomposing, the dynamic change characteristics of the signal at different time scales can be extracted, including the stable spectrum and multi-level fluctuation spectra. These characteristics can more comprehensively characterize the multi-dimensional information of the gas state in transformer oil.

[0139] Furthermore, for the extracted fluctuation spectrum characteristics, combined with the influence model of microbubbles on ultrasonic wave propagation, a series of physical parameters such as acoustic wave attenuation, propagation speed, scattering, acoustic impedance change, and nonlinear effects are calculated. These parameters can directly reflect the concentration and type of dissolved gases in transformer oil, providing a reliable physical basis for subsequent gas calculations.

[0140] Finally, using the deep residual attention network (ResANet) model, taking the above fusion features as input, the calculation results of the concentrations of various gases in transformer oil are output. This model has the ability of adaptive learning, can fully explore the complex relationships between multi-scale features, and realize the accurate prediction of the gas state in transformer oil.

[0141] The following provides an embodiment of a specific application scenario of the present invention: The 110 kV main transformer of a substation of an electric power enterprise in a certain province has been in operation for many years. The transformer oil is sampled and detected regularly to evaluate the health status of the transformer. According to past detection data, the content of dissolved gases such as hydrogen and carbon monoxide in this main transformer has exceeded the standard, which needs attention. In order to realize the online continuous monitoring of dissolved gases in transformer oil, this substation decides to adopt the method for detecting dissolved gases in transformer oil based on the joint spectrum of multi-frequency ultrasonic waves proposed by the present invention.

[0142] In August 2023, according to the technical solution of the present invention, a special ultrasonic detection module was installed on the outer shell of the protection device of the 110 kV main transformer in this substation. This module includes 5 ultrasonic transmitting transducers with different frequencies and 1 broadband ultrasonic receiving transducer, with frequencies of 20 kHz, 30 kHz, 40 kHz, 50 kHz, and 60 kHz respectively. At the same time, a signal acquisition, preprocessing, and analysis and calculation unit are integrated inside the module. The main parameters of the detection module are shown in Table 1:

[0143] Table 1 Main parameters of the ultrasonic detection module

[0144]

[0145]

[0146] This detection module automatically starts the detection program every 30 minutes, and the specific steps are as follows:

[0147] 1. Synchronously transmit 5 groups of ultrasonic signals with different frequencies. The transmission duration of each frequency signal is 50 ms, the amplitude is the sound pressure level of 110 dB, and the phase is random.

[0148] 2. Collect the ultrasonic response signals reflected and transmitted in the transformer oil. These signals are captured by a broadband ultrasonic sensor, with a sampling rate of 500 kHz and a quantization bit number of 16 bits.

[0149] 3. Preprocess the 5 channels of time-domain signals collected respectively. First, apply a Hamming window function for windowing, then use median filtering to remove pulse noise, and finally apply a band-pass filter (10 kHz - 100 kHz) to filter out low-frequency and high-frequency interference.

[0150] 4. Perform fast Fourier transform on the 5 channels of preprocessed signals to obtain the 5-channel spectral responses of the transformer oil, forming a multi-ultrasonic joint spectrum. As shown in Table 2, the amplitude and phase information of each frequency component in the joint spectrum are retained.

[0151] Table 2 Multi-ultrasonic joint spectrum of transformer oil

[0152] Frequency Amplitude Phase 20 kHz 0.85 0.12 rad 30 kHz 0.72 0.28 rad 40 kHz 0.61 0.41 rad 50 kHz 0.53 0.59 rad 60 kHz 0.48 0.72 rad

[0153] 5. Perform 3-layer wavelet decomposition on the joint spectrum using Daubechies wavelet (db4) to obtain a stable spectrum and a 3-level fluctuation spectrum. Taking the 20 kHz frequency component as an example, its wavelet decomposition results are shown in Table 3:

[0154] Table 3 Wavelet decomposition results of 20 kHz frequency component

[0155]

[0156]

[0157] 6. Calculate the energy of each layer of the fluctuation spectrum and evaluate its energy proportion relative to the stable spectrum. Taking the 20 kHz frequency as an example, the energy of each layer of the fluctuation spectrum and its proportion are shown in Table 4:

[0158] Table 4 Fluctuation spectrum energy of 20 kHz frequency component

[0159] Level Fluctuation spectrum energy Energy proportion Layer 1 0.0324 5.23％ Layer 2 0.0484 7.81％ Layer 3 0.0729 11.76％

[0160] 7. Extract 5 types of fluctuation spectrum features: interference feature, spectral attenuation feature, waveform complexity feature, non-linear interaction feature, and energy distribution feature. Taking the 20 kHz frequency as an example, the calculation results of each type of feature are shown in Table 5:

[0161] Table 5 Fluctuation spectrum features of 20 kHz frequency component

[0162] Feature type Calculation result Interference feature 0.64 Spectrum attenuation feature 0.78 dB / m Waveform complexity feature 4.12 Nonlinear interaction feature 0.032 Energy distribution feature 0.58

[0163] 8. Combining the influence model of microbubbles on ultrasonic waves, physical parameters such as acoustic attenuation, propagation velocity, scattering, changes in acoustic impedance, and nonlinear effects corresponding to the above-mentioned fluctuation spectrum characteristics are calculated. Taking the frequency of 20 kHz as an example, the calculation results of various physical parameters are shown in Table 6:

[0164] Table 6 Physical Parameters of the 20 kHz Frequency Component

[0165] Parameter name Calculation result Acoustic wave attenuation 0.82 dB / m Acoustic wave propagation speed 1432 m / s Acoustic wave scattering coefficient 0.024 Acoustic impedance change 1.18 MPa·s / m Nonlinear acoustic effect 1.14

[0166] 9. The different-level fluctuation spectrum characteristics and their corresponding physical parameters are weighted and fused to obtain the final fused characteristics. The fusion weights are determined according to the energy proportion of each layer of the fluctuation spectrum, as shown in Table 7:

[0167] Table 7 Calculation of Fused Characteristics

[0168] Level Energy proportion Fusion weight Layer 1 5.23％ 0.18 Layer 2 7.81％ 0.27 Layer 3 11.76％ 0.41

[0169] 10. The fused characteristics are input into a pre-trained deep residual attention network model to calculate the concentrations of various dissolved gases in transformer oil. The training dataset of this model includes more than 400 groups of transformer oil samples with different gas concentrations and types simulated in the laboratory, covering main gases such as hydrogen, carbon monoxide, ethane, and acetylene. Through stratified cross-validation, the prediction accuracy of this model on new samples reaches more than 92%, meeting the actual application requirements.

[0170] After the above 10-step detection process, the monitoring results of the dissolved gases in the 110 kV main transformer oil of this substation are shown in Table 8:

[0171] Table 8 Detection Results of Dissolved Gases in Transformer Oil

[0172] Gas type Concentration Hydrogen (H2) 120 ppm Carbon monoxide (CO) 350 ppm Ethane (C2H6) 45 ppm Acetylene (C2H2) 12 ppm Methane (CH4) 28 ppm

[0173] It can be seen from the detection results that the concentrations of hydrogen and carbon monoxide in the main transformer oil slightly exceed the standard, while the contents of other gases are normal. Combining the analysis of historical trend data, this may indicate signs of partial discharge in the transformer winding or a ground fault in the iron core.

[0174] Next, this substation will further strengthen the patrol inspection and regular maintenance of this main transformer, and at the same time will continuously monitor the online monitoring data of the dissolved gases in the transformer oil. If the concentrations of hydrogen and carbon monoxide continue to rise, or other abnormal situations occur, measures will be taken in a timely manner to avoid serious faults in the transformer.

[0175] Meanwhile, this substation also applies the detection method of the present invention to other 110 kV and 220 kV transformers to achieve comprehensive online monitoring of transformer status. Through the analysis of a large amount of transformer operation data, the diagnostic algorithm of this method is further improved to improve the accuracy and reliability of fault prediction, providing strong support for the safe and stable operation of the power grid.

[0176] Example 2: For the training of the gas calculation model in transformer oil of the present invention, a highly refined Ni / Zn@SnO2 gas sensor array acquisition unit was developed in order to obtain more comprehensive and rich training data. The microstructure of this acquisition unit is as Figure 2 and Table 9 shows:

[0177] Table 9 Microstructure of the Ni / Zn@SnO2 Gas Sensor Array Acquisition Unit

[0178] Level Material Thickness Si-based bottom layer Si 300 - 500 μm SiO2 insulating layer SiO2 0.5 - 2 μm Heating electrode layer Au / Pt 1 - 2 μm SiNi isolation layer SiNi 50 - 200 nm Test electrode layer Au / Pt 1 - 2 μm SnO2-based sensing layer Ni / Zn-doped SnO2 300 - 500 nm

[0179] This sensor array adopts a novel heating electrode design, including two types of heterogeneous structures:

[0180] 1. U-inverted U-shaped heating channel: The distance between the adjacent two side arms of the U-shaped is precisely controlled at 25 μm, which can achieve efficient and uniform local heating, as Figure 3 shown.

[0181] 2. Comb-shaped heating channel: The distance between the comb teeth is precisely designed at 5 μm. Through the diagonal array deployment at the four corners of the square, it can provide more delicate temperature control, as Figure 4 shown.

[0182] Figure 5 and Figure 6 are the circuit schematic diagram and the finished circuit diagram of the gas sensor array acquisition unit.

[0183] The Ni / Zn co-doped SnO2-based sensing layer can detect key dissolved gases such as hydrogen, methane, carbon monoxide, acetylene, and ethylene in transformer oil with high sensitivity and high selectivity within the working temperature range of 200 - 400 °C. This provides an innovative gas detection solution at the microscale for the status diagnosis of power equipment.

[0184] Based on the above Ni / Zn@SnO2 gas sensor array acquisition unit, a dataset dedicated to training the gas calculation model in transformer oil was constructed. The specific process is as follows:

[0185] 1. Sample preparation: First, under laboratory conditions, simulated transformer oil samples were prepared using mixed solutions of hydrogen, methane, carbon monoxide, acetylene, and ethylene with different concentrations. The concentration ranges of these dissolved gases cover various states that may occur during the actual operation of transformers, as shown in Table 10:

[0186] Table 10 Dissolved gas concentration range of simulated transformer oil samples

[0187] Gas type Concentration range Hydrogen (H2) 50 - 500 ppm Methane (CH4) 20 - 300 ppm Carbon monoxide (CO) 100 - 800 ppm Acetylene (C2H2) 5 - 100 ppm Ethylene (C2H4) 10 - 150 ppm

[0188] 2. Gas sensor array detection: The prepared transformer oil samples are sequentially placed into the Ni / Zn@SnO2 gas sensor array acquisition unit to perform gas detection in the temperature range of 200-400°C. The array synchronously collects the response signals of each gas at different temperatures, including key parameters such as resistance change, response time, and recovery time. Through a large number of experimental accumulations, a sensor array data set covering a wide range of gas concentrations and types has been obtained.

[0189] 3. Data preprocessing: Perform preprocessing operations such as outlier detection and filtering, dimensionality consistency, and numerical normalization on the collected raw sensor data. These steps ensure the reliability and applicability of the data set.

[0190] 4. Dataset division: The preprocessed dataset is randomly divided into training set, validation set and test set in a ratio of 6:2:2. The training set is used for model parameter optimization learning, the validation set is used for hyperparameter tuning and performance evaluation, and the test set is used for objective testing of the final model performance.

[0191] 5. Data augmentation: For the training set, some common data augmentation techniques are used, such as translation, rotation, scaling, and noise addition, to further enrich the diversity of training data. This can improve the generalization ability of the model and reduce the risk of overfitting.

[0192] Through the above five steps, a high-quality transformer oil dissolved gas detection dataset is established, which provides a reliable foundation for subsequent computational model training. The main features of this dataset are as follows:

[0193] 1. Wide coverage: The sample gas concentration and type cover various states that may occur in the actual operation of the transformer, and can fully reflect the complexity of the on-site environment.

[0194] 2. Accurate measurement: The Ni / Zn@SnO2 gas sensor array acquisition unit has high sensitivity and high selectivity, and can accurately capture the response characteristics under slight concentration changes.

[0195] 3. Multidimensional features: In addition to gas concentration, the dataset also includes multidimensional features such as temperature, response time, and recovery time, providing richer input for subsequent modeling.

[0196] 4. High quality: Through strict data preprocessing and amplification strategies, the reliability and representativeness of the data set are ensured, providing a good foundation for model training.

[0197] With such a highly refined and widely covered dataset of dissolved gas in transformer oil as the support, the deep residual attention network (ResANet) model in the present invention was designed and trained specifically. The specific steps are as follows:

[0198] 1. Model architecture design: The ResANet network consists of multiple residual blocks, self-attention mechanisms, and fully connected layers. Residual blocks can effectively alleviate the degradation problem of deep networks, while self-attention mechanisms can adaptively extract the importance of multi-scale features. After repeated experiments, appropriate network depth and width parameters were determined.

[0199] 2. Parameter initialization: According to the network structure and data characteristics, methods such as Xavier initialization and He initialization were used to reasonably initialize the model parameters. This helps to improve the convergence speed and generalization performance of the model.

[0200] 3. Loss function design: Considering the existence of multiple dissolved gases in transformer oil, a multi-scale cross-entropy loss function combining different gas concentration indicators was designed. This can promote the model to learn the characteristics of various gases and improve the prediction accuracy.

[0201] 4. Optimizer selection: During the training process, the Adam optimizer with an adaptive learning rate was adopted. The Adam algorithm can automatically adjust the learning rate of each parameter according to the first and second moments of the gradient, improving the convergence speed and robustness of the model.

[0202] 5. Cross-validation: To comprehensively evaluate the model performance, stratified K-fold cross-validation was adopted. By performing cross-validation on transformer oil samples with different gas concentrations and types, the indicators of the model in terms of prediction accuracy, generalization ability, etc. can be objectively evaluated.

[0203] 6. Parameter fine-tuning: For the transformer oil samples collected from the actual substation site, the model parameters were further fine-tuned to adapt to the characteristics of different transformer oils. At the same time, the generalization performance of the model on new samples was also verified to ensure that it can meet the requirements of transformer operation status monitoring.

[0204] After the above detailed model training process, a gas calculation model for transformer oil with excellent performance was finally obtained. On the test set, the average prediction accuracy of this model reached over 95%, meeting the actual application requirements of the power system for monitoring dissolved gases in transformer oil.

[0205] Specifically, this model can accurately identify 5 key dissolved gases in transformer oil and give the accurate concentration calculation results of each gas. Taking a certain on-site detection as an example, the output of the model is shown in Table 11:

[0206] Table 11 Detection Results of Dissolved Gases in Transformer Oil

[0207] Gas type Concentration Hydrogen (H2) 168 ppm Methane (CH4) 86 ppm Carbon monoxide (CO) 412 ppm Acetylene (C2H2) 38 ppm Ethylene (C2H4) 63 ppm

[0208] As can be seen from the data in Table 11, the concentrations of hydrogen and carbon monoxide in the transformer oil are slightly on the high side, while the contents of other gases are within the normal range. This provides an important basis for the subsequent assessment of the transformer status.

[0209] Compared with traditional laboratory sampling and analysis, the online gas detection scheme based on the Ni / Zn@SnO2 gas sensor array has obvious advantages. First of all, this scheme can realize the real-time dynamic monitoring of the dissolved gases in the transformer oil, greatly improving the timeliness and continuity of the monitoring. Secondly, the high sensitivity and selectivity of the sensor array ensure the accuracy of the detection results and provide a reliable basis for fault warning. Moreover, the equipment investment and operation cost of this scheme are relatively low, providing favorable conditions for the application and promotion by power grid enterprises.

[0210] It should be noted that the variable explanations involved in the present invention are shown in Table 12 below.

[0211] Table 12 Variable Explanation Table

[0212]

[0213]

[0214] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for detecting dissolved gas in transformer oil, characterized in that: The following steps are involved: S10, simultaneously emitting multiple groups of ultrasonic waves of different frequencies into the transformer oil to obtain a multi-ultrasonic joint spectrum of the transformer oil; S20, preprocessing the multiple ultrasonic joint spectrum to remove noise interference signals to obtain a preprocessed joint spectrum; S30, decomposing the preprocessed joint spectrum by a wavelet multi-scale decomposition method considering a time sliding window to obtain a stable spectrum, a primary fluctuation spectrum, a secondary fluctuation spectrum and a tertiary fluctuation spectrum, wherein the size of the time sliding window is adaptively determined according to the frequency characteristics of the preprocessed joint spectrum and the signal dynamic range; S40, using a preset wave spectrum feature extraction equation group to extract features of the three levels of wave spectra respectively, to obtain a wave spectrum feature group corresponding to each level of wave spectrum, wherein the wave spectrum feature group includes interference features, spectrum attenuation features, waveform complexity features, coupling features, and energy distribution features; S50, for each level of wave spectrum corresponding to the feature group, using the preset microbubble to calculate the ultrasonic wave influence equation group, to obtain each level of wave spectrum corresponding to the microbubble influence feature group, including sound wave attenuation characteristics, sound wave propagation velocity characteristics, sound wave scattering characteristics, acoustic impedance change characteristics and nonlinear acoustic characteristics; S60, calculating the energy ratio of the fluctuation spectrum at each level, taking it as a weight, and fusing the feature groups corresponding to the fluctuation energy ratios at three levels to obtain a fusion feature; S70, inputting the fusion feature into a pre-trained transformer oil gas calculation model to obtain a transformer oil gas calculation result.

2. A method for detecting dissolved gas in transformer oil according to claim 1, characterized in that: The wave spectrum feature extraction equation family includes an interference feature equation, a spectrum attenuation equation, a waveform complexity equation, a nonlinear interaction equation and an energy distribution equation.

3. A method for detecting dissolved gas in transformer oil according to claim 2, characterized in that: The interference characteristic equation is used to describe the superposition and interaction of ultrasonic waveforms of different frequencies; The spectrum attenuation equation is used to calculate the energy attenuation law of ultrasonic waves when propagating in transformer oil; The waveform complexity equation is used to quantify the structural complexity of the waveform; The nonlinear interaction equation is used to describe the nonlinear coupling relationship between ultrasonic waves; The energy distribution equation is used to analyze the distribution characteristics of energy in transformer oil at different frequencies.

4. A method for detecting dissolved gas in transformer oil according to claim 3, characterized in that: The equation group of the influence of microbubbles on ultrasonic waves includes: a characteristic equation of sound wave attenuation, a characteristic equation of sound wave propagation speed, a characteristic equation of sound wave scattering, a characteristic equation of acoustic impedance change, and a nonlinear acoustic characteristic equation.

5. A method for detecting dissolved gas in transformer oil according to claim 4, characterized in that: The acoustic wave attenuation characteristic equation is used to describe that dissolved gas and microbubbles will significantly increase the attenuation of ultrasonic waves in oil; The characteristic equation of sound wave propagation velocity is used to describe that gas and microbubbles will change the propagation velocity of sound waves in oil; The characteristic equation of acoustic wave scattering is used to describe the strong scattering and reflection of sound waves caused by microbubbles; The characteristic equation for acoustic impedance change is used to describe that dissolved gas will change the acoustic impedance of oil; The nonlinear acoustic characteristic equation is used to describe the nonlinear propagation effect of ultrasound caused by gas and microbubbles.

6. A method for detecting dissolved gas in transformer oil according to claim 5, characterized in that: The transformer oil gas calculation model adopts a deep residual attention network structure, which is specifically composed of multiple layers of residual blocks, self-attention mechanism and fully connected layers. The network can adaptively learn the complex features of fluctuation spectra of different scales.

7. A method for detecting dissolved gas in transformer oil according to claim 6, characterized in that: The training data set of the transformer oil gas calculation model is obtained by simulating transformer oil samples with different gas concentrations and types in the laboratory, and using high-precision ultrasonic detection equipment to perform multi-spectrum joint acquisition to ensure that the data set covers a wide range of gas states and concentration changes.

8. A method for detecting dissolved gas in transformer oil according to claim 7, characterized in that: The transformer oil samples included dissolved gas samples simulating different gas concentrations and types, including: hydrogen (H2), methane (CH4), carbon monoxide (CO), carbon dioxide (CO2), acetylene (C2H2) and ethylene (C2H4).

9. A method for detecting dissolved gas in transformer oil according to claim 8, characterized in that: The gas concentration range covers 0-1000 ppm and is sampled according to logarithmic intervals or geometric series.

10. A method for detecting dissolved gas in transformer oil according to claim 9, characterized in that: The high-precision ultrasonic testing equipment has the following technical indicators: (1) Frequency range: 100kHz-10MHz; (2) Dynamic range: >100dB; (3) Signal-to-noise ratio: >80dB; (4) Time resolution: <1ns.

Images