Method for identifying liquid hydrogen c-type ball valve internal leakage based on acoustic emission frequency characteristics

Abstract

The application provides a method for identifying internal leakage of a liquid hydrogen C-type ball valve based on acoustic emission frequency characteristics, and belongs to the technical field of internal leakage detection of liquid hydrogen C-type ball valves. In the application, an acoustic emission sensor array is arranged on the outer surface of the valve seat of the liquid hydrogen C-type ball valve to collect original signals. After adaptive multi-scale wavelet packet decomposition, mixed noise reduction and coherent accumulation enhancement are performed. A material dispersion compensation model is established to extract parameters and input a leakage characteristic identification model to obtain a leakage probability value. When the probability value exceeds a threshold value, fast independent component analysis and sparse reconstruction are used to estimate the number of leakage sources. The spatial angle of the leakage source is located by using a multiple signal classification algorithm. The leakage aperture is calculated by thermodynamic inversion calculation. The turbulent noise characteristics extracted by the fractal theory are fused and judged, thereby solving the technical problems that it is difficult to accurately identify a small leakage source and realize quantitative evaluation of the leakage degree in a strong background noise environment during internal leakage detection of the liquid hydrogen C-type ball valve.

Description

Technical Field

[0001] This invention belongs to the field of liquid hydrogen C-type ball valve internal leakage detection technology, specifically, it relates to a method for identifying liquid hydrogen C-type ball valve internal leakage based on acoustic emission frequency characteristics. Background Technology

[0002] The sealing performance of C-type ball valves in liquid hydrogen storage and transportation systems directly affects system safety. Traditional internal leak detection primarily employs acoustic emission technology, using sensor arrays to collect and analyze leak acoustic signals. However, in existing technologies, the 10kHz to 500kHz broadband turbulent noise generated by liquid hydrogen flash evaporation highly overlaps with weak leak signals in the frequency domain. Furthermore, the dispersion effect of materials at low temperatures causes severe distortion of different frequency components during sound wave propagation, making the features extracted by traditional time-frequency analysis methods sensitive to propagation distance and difficult to accurately identify. Simultaneously, existing methods can only qualitatively determine the presence of a leak, failing to establish a quantitative relationship between acoustic measurements and leak orifice size, thus hindering the assessment of leak severity. In other words, existing technologies present a technical challenge in accurately identifying minute leak sources and quantitatively assessing leak severity in liquid hydrogen C-type ball valve internal leak detection under strong background noise conditions. Summary of the Invention

[0003] In view of this, the present invention provides a method for identifying internal leakage of liquid hydrogen C-type ball valve based on acoustic emission frequency characteristics, which can solve the technical problem in the prior art that it is difficult to accurately identify minute leakage sources and achieve quantitative assessment of leakage degree in strong background noise environment when detecting internal leakage of liquid hydrogen C-type ball valve.

[0004] This invention is implemented as follows: It provides a method for identifying internal leakage in a liquid hydrogen C-type ball valve based on acoustic emission frequency characteristics. The method includes: uniformly arranging a circular array of eight broadband acoustic emission sensors on the outer surface of the liquid hydrogen C-type ball valve seat to collect the original acoustic emission signal; performing adaptive multi-scale wavelet packet decomposition on the original acoustic emission signal into frequency band sub-signals; performing hybrid noise reduction using spectral subtraction and Wiener filtering; and coherently accumulating and enhancing the signal using periodic modulation characteristics to obtain the noise-reduced acoustic emission signal; establishing a material dispersion compensation model to perform dispersion compensation on the noise-reduced acoustic emission signal; and extracting the Wigner-Ville distribution ridge parameters as the dispersion. The model inputs invariant feature quantities to the leakage feature identification model and outputs leakage probability values ​​and leakage type classification results. When the leakage probability value is greater than the preset leakage probability threshold, the fast independent component analysis algorithm is used to separate blind sources and the number of leakage sources is estimated by combining sparse reconstruction theory. The spatial spectrum is estimated using a multi-signal classification algorithm to obtain the spatial angular position of each leakage source. The leakage orifice size is estimated by thermodynamically driving leakage sound source intensity inversion for each leakage source. The fractal feature vector is constructed by fractal theory turbulence noise feature extraction of the denoised acoustic emission signal. The fractal feature vector is fused with the leakage probability value and leakage orifice size estimate to determine the leakage state.

[0005] The broadband acoustic emission sensor has a frequency response range of 10kHz to 500kHz.

[0006] The adaptive multi-scale wavelet packet decomposition has 5 decomposition layers, which decomposes the original acoustic emission signal into 32 frequency band sub-signals.

[0007] The hybrid noise reduction process involves performing spectral subtraction on each frequency band sub-signal before Wiener filtering.

[0008] The oversubtraction factor used in the spectral subtraction processing ranges from 1.2 to 1.5.

[0009] The modulation frequency range of the periodic modulation characteristic is 100Hz to 2kHz.

[0010] Among them, material dispersion compensation uses deconvolution technology to reconstruct the original waveform, and the deconvolution technology uses the Tikhonov regularization method.

[0011] Among them, the frequency domain compensation filter of the transfer function adopts a finite impulse response structure with an order of 128 to 256.

[0012] Among them, the ridge parameters of the Wigner-Ville distribution include ridge slope, ridge curvature, and ridge energy concentration.

[0013] The structure of the leakage feature recognition model is a cascaded architecture consisting of an input layer, a convolutional feature extraction layer, an attention weighting layer, a recurrent memory layer, and a fully connected classification layer.

[0014] This invention establishes an adaptive multi-scale wavelet packet decomposition and hybrid denoising signal preprocessing system. Broadband noise and leakage features are separated into different frequency bands and denoised independently. Then, periodic modulation characteristics are used for coherent accumulation to enhance the signal-to-noise ratio, effectively suppressing turbulent noise interference. By constructing a material dispersion compensation model and extracting Wigner-Ville distribution ridge parameters as dispersion-invariant features, the influence of propagation distance on the features is eliminated, making leak identification independent of sensor location. A constitutive relationship between acoustic emission energy and leakage aperture is established through thermodynamically driven sound source intensity inversion. Combined with turbulent noise features extracted using fractal theory, a fusion judgment is made, achieving a leap from qualitative detection to quantitative assessment. In summary, this invention solves the technical problem mentioned in the background art of accurately identifying minute leak sources and achieving quantitative assessment of leakage degree in liquid hydrogen C-type ball valve internal leakage detection under strong background noise environments. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2The Wigner-Ville time-frequency distribution diagram of the acoustic emission signal after noise reduction.

[0017] Figure 3 This is the spatial spectrum estimation diagram for the multiple signal classification algorithm.

[0018] Figure 4 This is a comparison of the distribution of the fractal eigenvectors of the leakage signal and the turbulence noise. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0020] like Figure 1 The diagram shown is a flowchart of a method for identifying internal leakage in a liquid hydrogen C-type ball valve based on acoustic emission frequency characteristics, provided by the present invention. This method includes the following steps:

[0021] S01. An acoustic emission sensor array is uniformly arranged on the outer surface of the C-type ball valve seat of liquid hydrogen. The acoustic emission sensor array consists of 8 broadband acoustic emission sensors and is evenly distributed in a circle. The frequency response range of each broadband acoustic emission sensor is 10kHz to 500kHz.

[0022] S02. Acquire the original acoustic emission signal output by the acoustic emission sensor array, and perform adaptive multi-scale wavelet packet decomposition on the original acoustic emission signal. The decomposition level is 5 levels, and the original acoustic emission signal is decomposed into 32 frequency band sub-signals.

[0023] S03. Perform hybrid noise reduction processing on the frequency band sub-signals. The hybrid noise reduction processing is to first perform spectral subtraction processing on each frequency band sub-signal, then perform Wiener filtering processing, and then use periodic modulation characteristics to perform coherent accumulation enhancement to obtain the noise-reduced acoustic emission signal.

[0024] S04. Establish a material dispersion compensation model for the noise-reduced acoustic emission signal and perform dispersion compensation. Reconstruct the original waveform through deconvolution technology, use a transfer function frequency domain compensation filter for compensation, and extract the Wigner-Ville distribution ridge parameters as dispersion-invariant features.

[0025] S05. Input the dispersion invariant feature quantity into the leakage feature identification model for processing. The leakage feature identification model outputs the leakage probability value and leakage type classification result.

[0026] S06. When the leakage probability value is greater than the preset leakage probability threshold, the fast independent component analysis algorithm is used to perform blind source separation on the denoised acoustic emission signal to obtain independent component signals. The independent component signals are then processed using sparse reconstruction theory to estimate the number of leakage sources. Finally, the spatial spectrum of the denoised acoustic emission signal of the acoustic emission sensor array is estimated using a multiple signal classification algorithm to obtain the spatial angular position of each leakage source.

[0027] S07. Perform thermodynamically driven leakage sound source intensity inversion processing on the leakage source corresponding to the spatial angle position, establish the constitutive relationship between leakage orifice diameter, pressure difference and sound power based on Bernoulli equation and isentropic expansion theory, calculate the flash rate by combining hydrogen Joule-Thomson coefficient, calculate the sound power value by measuring acoustic emission energy, and obtain the leakage orifice diameter estimate by solving the nonlinear equation system using Newton iteration method.

[0028] S08. Extract turbulent noise features from the denoised acoustic emission signal using fractal theory, calculate the box dimension, correlation dimension, and multifractal spectrum of the denoised acoustic emission signal, estimate the fractal dimension using the Higuchi algorithm, calculate the generalized dimension spectrum using the Chhabra-Jensen method, and extract the singularity index and Hurst index to construct the fractal feature vector.

[0029] S09. The fractal feature vector is fused with the leakage probability value and the leakage aperture estimate. When the leakage probability value is greater than or equal to 0.75 and the leakage aperture estimate is greater than 0.1 mm, it is determined to be a confirmed leak. When the leakage probability value is in the range of [0.5, 0.75) and the leakage aperture estimate is in the range of [0.05 mm, 0.1 mm], it is determined to be a suspected leak. When the leakage probability value is less than 0.5, it is determined to be a normal state.

[0030] The adaptive multi-scale wavelet packet decomposition is a method that simultaneously refines the signal in the time and frequency domains. By constructing a decomposition tree structure, both the low-frequency and high-frequency components of the signal are further decomposed, ensuring that each frequency band achieves a time-frequency resolution adapted to its characteristics. A decomposition layer of 5 layers means that the original acoustic emission signal is decomposed into 32 frequency bands of equal bandwidth, each with a bandwidth of approximately 15.6 kHz. This adaptive multi-scale wavelet packet decomposition targets the 10 kHz to 500 kHz broadband noise generated by liquid hydrogen flash evaporation, separating leakage characteristic signals and interference noise of different frequency components into different frequency band sub-signals, providing a foundation for subsequent independent denoising processing. The adaptive aspect is reflected in the construction of the decomposition tree, automatically selecting whether to continue decomposition based on the energy distribution and entropy value of each node signal, avoiding computational redundancy caused by over-decomposition and feature loss caused by under-decomposition.

[0031] The spectral subtraction processing described herein is a noise reduction method that subtracts the estimated noise power spectrum from the power spectrum of a noisy signal in the frequency domain. The steps of spectral subtraction processing are as follows: performing a short-time Fourier transform on the sub-signal of the frequency band to obtain its spectrum; estimating the average power spectrum of the noise band as a noise reference; subtracting the noise power spectrum from the signal power spectrum and multiplying by an over-subtraction factor; and finally performing an inverse Fourier transform to recover the time-domain signal. The over-subtraction factor is set to 1.2 to 1.5 to compensate for noise estimation errors. Spectral subtraction processing has a good suppression effect on stationary noise.

[0032] The Wiener filtering process described herein is based on an optimal linear filter designed using the minimum mean square error criterion. By estimating the power spectral density of the signal and noise, the transfer function of the Wiener filter is constructed. This transfer function assigns different weights to different frequency components in the frequency domain, with higher signal-to-noise ratio (SNR) components receiving larger weights and lower SNR components receiving smaller weights. This Wiener filtering process effectively smooths the musical noise generated by the spectral subtraction process, further improving the SNR.

[0033] The periodic modulation characteristic refers to the periodic amplitude modulation phenomenon of the acoustic emission signal caused by pressure pulsation and turbulent vortex shedding during liquid hydrogen leakage. The modulation frequency is typically in the range of 100Hz to 2kHz. The coherent accumulation enhancement utilizes this periodic modulation characteristic by synchronously superimposing signals from multiple modulation cycles. This linearly accumulates the amplitude of the coherent leakage characteristic signal, while the amplitude of the incoherent random noise increases according to the square root law, thereby improving the signal-to-noise ratio. The more accumulation cycles, the more significant the enhancement effect, but the signal needs to have a sufficient duration.

[0034] The material dispersion compensation model describes the differences in propagation speed and attenuation coefficient of different frequency components when sound waves propagate in cryogenic materials. Around the liquid hydrogen operating temperature of 20K, the sound velocity and attenuation coefficient of the material exhibit nonlinear changes with frequency, with high-frequency components propagating slower and attenuating more rapidly. The material dispersion compensation model obtains the dispersion curve of the material through experimental measurements or theoretical calculations, establishing a functional relationship between frequency, phase velocity, and attenuation coefficient. This model is used for subsequent waveform reconstruction to compensate for waveform distortion caused by dispersion.

[0035] The deconvolution technique involves deconvolving the received distorted signal with the transfer function of the propagation path to recover the original waveform at the leakage source. In the frequency domain, this is represented by dividing the received signal spectrum by the transfer function spectrum. Since the transfer function approaches zero at some frequencies, causing instability in the division, a regularized deconvolution method is employed, introducing a Tikhonov regularization parameter to balance the accuracy and stability of the deconvolution. The original waveform reconstructed by this deconvolution technique eliminates waveform distortion caused by dispersion and restores the true time-domain characteristics of the leaked signal.

[0036] The transfer function frequency domain compensation filter is a digital filter constructed based on the material dispersion compensation model. The frequency response characteristics of the transfer function frequency domain compensation filter are inversely related to the transfer function of the propagation path. The filter applies different gains and phase compensations to different frequency components, providing greater gain to compensate for propagation attenuation in high-frequency components and correcting phase distortion caused by dispersion through phase compensation. The filter is implemented using a finite impulse response structure with orders from 128 to 256, ensuring compensation accuracy and real-time performance.

[0037] The Wigner-Ville distribution is a time-frequency analysis method that obtains the energy distribution of a signal in the time-frequency plane by calculating the Fourier transform of the instantaneous autocorrelation function of the signal. The Wigner-Ville distribution has high time-frequency resolution and can clearly display the instantaneous frequency changes of the signal. The ridge parameters of the Wigner-Ville distribution refer to the curves with the highest energy concentration in the Wigner-Ville distribution, representing the trajectory of the signal's main frequencies over time. The ridge parameters include the ridge slope, ridge curvature, and ridge energy concentration. These parameters are insensitive to dispersion and remain relatively stable at different propagation distances; therefore, they are called dispersion-invariant characteristics. The method for extracting the ridge parameters involves finding the frequency corresponding to the energy peak for each time slice of the Wigner-Ville distribution, connecting these frequencies to form a ridge, and then calculating the geometric and statistical characteristics of the ridge.

[0038] The leakage feature recognition model has a cascaded architecture consisting of an input layer, a convolutional feature extraction layer, an attention-weighted layer, a recurrent memory layer, and a fully connected classification layer. The input layer receives a 32-dimensional dispersion-invariant feature vector. The convolutional feature extraction layer contains three one-dimensional convolutional layers with kernel sizes of 5, 3, and 3, and channel numbers of 64, 128, and 128, respectively, using the ReLU activation function to extract local patterns of features. The attention-weighted layer employs a multi-head attention mechanism with four heads, each with a dimension of 32. It calculates the correlation between different dimensions in the dispersion-invariant feature vector through self-attention, assigning higher weights to important features. The recurrent memory layer uses a bidirectional long short-term memory network structure with a hidden layer dimension of 256 to capture the temporal dependencies of feature sequences. The fully connected classification layer contains two fully connected layers with 128 and 3 neurons, respectively. Finally, a softmax function is used to output probability distributions for three categories, corresponding to normal state, suspected leakage, and confirmed leakage. The leakage probability value is defined as the sum of the output probability of the confirmed leakage category and the output probability of the suspected leakage category. The leakage type classification result is the category corresponding to the maximum probability among the three categories.

[0039] The steps for establishing the training dataset for the leakage feature recognition model include: preparing standard leakage samples with different pore sizes on an experimental bench, the pore size range being 0.01 mm to 5 mm, with an interval of 0.01 mm; conducting leakage experiments on each pore size under different pressure differential conditions, the pressure differential range being 0.1 MPa to 2 MPa, with an interval of 0.1 MPa; collecting acoustic emission signals under each leakage condition; and extracting the dispersion-invariant feature quantity according to steps S02 to S04; and having experts evaluate each sample based on the leakage pore size and actual leakage rate. Category labels were used to label leaks: leaks with an aperture less than 0.05 mm and a leakage rate less than 1 μg / s were labeled as normal; leaks with an aperture between 0.05 mm and 0.1 mm and a leakage rate between 1 μg / s and 10 μg / s were labeled as suspected leaks; and leaks with an aperture greater than 0.1 mm or a leakage rate greater than 10 μg / s were labeled as confirmed leaks. 5000 normal state samples, 3000 suspected leak samples, and 2000 confirmed leak samples were collected and divided into training, validation, and test sets in a ratio of 8:1:1.

[0040] The training steps of the leakage feature recognition model include: using the Adam optimizer for gradient descent optimization, with an initial learning rate of 0.001, which decays to 0.9 times its original value every 10 training epochs; using a weighted combination of cross-entropy loss and focus loss as the loss function, with a weight ratio of 1:2, where the focus loss is used to alleviate class imbalance; setting the batch size to 64 and the total number of training epochs to 200; evaluating the model performance on the validation set after each training epoch, recording the validation set accuracy and loss value; employing an early stopping strategy, stopping training when the validation set loss does not decrease for 20 consecutive epochs; selecting the model parameters with the highest validation set accuracy as the final model; and using data augmentation techniques during training, adding Gaussian white noise and random time-shift perturbations to the training samples, with noise intensity ranging from 5% to 10% of the signal amplitude and time shifts ranging from -5 to +5 sampling points to enhance the robustness of the model.

[0041] The number of attention heads in the multi-head attention mechanism of the leakage feature identification model is determined based on the dimension of the dispersion-invariant feature quantity, the signal-to-noise ratio of the denoised acoustic emission signal, and the valve operating pressure. An attention head number adjustment function is designed, which calculates an adjustment factor based on the normalized value of the feature dimension, the normalized value of the signal-to-noise ratio, and the normalized value of the pressure. The normalized value of the feature dimension is the actual feature dimension divided by 32, the normalized value of the signal-to-noise ratio is the actual signal-to-noise ratio divided by 30 dB, and the normalized value of the pressure is the actual operating pressure divided by 2 MPa. The adjustment factor is equal to the power of 0.4 of the normalized value of the feature dimension multiplied by the power of 0.3 of the normalized value of the signal-to-noise ratio multiplied by the power of 0.3 of the normalized value of the pressure. When the adjustment factor is within [0, 0.7), two attention heads are used; when the adjustment factor is within [0.7, 1.3), four attention heads are used; and when the adjustment factor is greater than or equal to 1.3, eight attention heads are used. The attention head number adjustment function is used to dynamically adjust the complexity of the attention weighting layer according to the actual working conditions. In low signal-to-noise ratio or low-pressure conditions, the number of attention heads is reduced to reduce the risk of overfitting, while in high signal-to-noise ratio or high-pressure conditions, the number of attention heads is increased to improve the feature representation ability.

[0042] The preset leakage probability threshold is set to 0.5, which is used to determine whether to start the multi-leakage source location process.

[0043] The Fast Independent Component Analysis (FIC) algorithm is a blind source separation method used to separate independent source signals from a mixed signal. Based on the principle of maximizing non-Gaussianity, the FIC algorithm assumes that the source signals are statistically independent and that at most one source signal follows a Gaussian distribution. The FIC algorithm iteratively optimizes the unmixing matrix to maximize the non-Gaussianity of the separated signal. The non-Gaussianity is measured using negative entropy or kurtosis. The FIC algorithm employs a fixed-point iterative scheme, which converges faster than the traditional gradient descent method. In this scheme, the denoised acoustic emission signals from eight broadband acoustic emission sensors are used as the mixed observation signal input to the FIC algorithm. The FIC algorithm outputs the separated independent component signals, each corresponding to a leakage source.

[0044] The sparse reconstruction theory is used to estimate the number of leakage sources. This theory assumes that leakage sources are sparsely distributed in the spatial domain, meaning that only a few locations contain leakage sources, while most locations are leak-free. By constructing a spatial grid, the leakage source localization problem is transformed into a sparse signal reconstruction problem. An L1-norm regularized optimization algorithm is used to reconstruct the sparse source distribution in the spatial domain from the independent component signals. The number of non-zero elements in the reconstruction result represents the number of leakage sources, and the positions of these non-zero elements indicate their spatial locations. Compared to traditional methods, this sparse reconstruction theory can achieve super-resolution source number estimation even with a limited number of sensors.

[0045] The described multi-signal classification algorithm is a high-resolution spatial spectrum estimation method used to accurately locate the spatial positions of multiple sound sources. Based on subspace theory, the algorithm decomposes the covariance matrix of the observed signal into eigenvalues, separating a signal subspace and a noise subspace. The signal subspace is spanned by eigenvectors related to the signal, and the noise subspace is spanned by eigenvectors related to the noise; the signal and noise subspaces are orthogonal to each other. By searching for the orthogonality between the spatial steering vector and the noise subspace, a spatial spectrum function is constructed, and the peak position of the spatial spectrum function corresponds to the azimuth angle of the sound source. The resolution of the multi-signal classification algorithm is significantly higher than that of traditional beamforming methods, capable of distinguishing multiple sound sources with angular intervals smaller than the array beamwidth. In this scheme, the circular geometry of the acoustic emission sensor array is used to calculate the steering vectors corresponding to different azimuth angles, and the spatial angular positions of each leakage source are determined through one-dimensional spectral peak search.

[0046] The thermodynamically driven leakage sound source intensity inversion process establishes the physical relationship between leakage parameters and acoustic measurements based on the fundamental laws of fluid mechanics and thermodynamics. The Bernoulli equation describes the energy conservation of fluids along streamlines; for incompressible fluids, the sum of pressure energy, kinetic energy, and potential energy remains constant. The isentropic expansion theory describes the relationship between pressure, temperature, and density of a gas during an adiabatic process, following the Poisson equation. When hydrogen passes through the leak, it undergoes isentropic expansion, with the pressure dropping from the high-pressure side to the low-pressure side, and the temperature decreasing accordingly. Some hydrogen changes from a liquid to a gaseous state, resulting in flash evaporation. The flash evaporation rate is determined by the hydrogen Joule-Thomson coefficient, which describes the rate of temperature change with pressure during isenthalpic expansion. For hydrogen at low temperatures, the hydrogen Joule-Thomson coefficient is negative, meaning that a decrease in pressure leads to an increase in temperature. However, liquid hydrogen flash evaporation is mainly driven by a phase change caused by the pressure drop. The constitutive relationship between the leakage orifice diameter, the pressure difference, and the acoustic power is established using a sound source model, where the leakage orifice is equivalent to a point sound source. The acoustic power is proportional to the square of the leakage mass flow rate and the cube of the jet velocity. The acoustic power value is calculated using the measured acoustic emission energy, and combined with the measured pressure difference value, a set of nonlinear equations is established regarding the leakage orifice diameter and the leakage mass flow rate.

[0047] The Newton-Raphson iteration method is a numerical method for solving nonlinear equation systems, approximating the roots of the equations through iteration. In each iteration, the nonlinear equation system is expanded using Taylor at the current point, retaining the first-order terms to obtain a linearized equation. Solving the linearized equation yields a correction, which is then used to update the current point. The iteration process is repeated until the correction is less than a set threshold or the maximum number of iterations is reached. The Newton-Raphson iteration method has a second-order convergence rate. In this scheme, physical constraints and empirical formulas are used to provide initial value estimates, with an iteration convergence threshold of 0.01% and a maximum number of iterations of 50.

[0048] The box dimension is a fractal dimension describing the complexity of a set, calculated by statistically analyzing the number of boxes of different scales required to cover the set. For time-series signals, the signal is first reconstructed into phase space to form a trajectory. Then, the trajectory is covered with boxes of different side lengths, and the number of boxes required for each side length is statistically analyzed. The number of boxes and the side length have a linear relationship in a log-log coordinate system, and the negative value of the slope is the box dimension. The box dimension reflects the spatial filling capability of the signal. Turbulent noise has a high box dimension due to its chaotic characteristics, typically between 2.5 and 3.5, while periodic or quasi-periodic leakage signals have a lower box dimension, typically between 1.5 and 2.5.

[0049] The correlation dimension is a fractal dimension defined based on the correlation integral, which statistically represents the proportion of point pairs in the phase space whose distance is less than a given radius. The correlation dimension is the scaling exponent of the correlation integral as a function of the radius. This correlation dimension is insensitive to noise and can reveal the intrinsic dimension of the attractor. Calculating the correlation dimension requires selecting an appropriate embedding dimension and time delay to reconstruct the phase space. The embedding dimension is determined using the pseudo-nearest neighbor method, and the time delay is determined by the zeros of the autocorrelation function or the minimum of the mutual information. The correlation dimension of turbulent noise is typically higher than that of leakage signals; this difference in correlation dimension is used to distinguish between noise- and signal-dominated frequency bands.

[0050] The multifractal spectrum describes the local singularity distribution of a signal at different scales and locations. The multifractal spectrum is characterized by a generalized dimension spectrum or a singularity spectrum. The generalized dimension spectrum is a series of fractal dimensions of different orders, ranging from negative infinity to positive infinity, with the zero-order generalized dimension spectrum equal to the box dimension. The singularity spectrum describes the fractal dimension corresponding to different singularity intensities, with the horizontal axis representing the singularity index and the vertical axis representing the fractal dimension corresponding to the singularity index. Turbulent noise exhibits a broad multifractal spectrum due to its multi-scale vortex structure, resulting in a larger spectral width, while the leakage signal exhibits a narrower multifractal spectrum. The width, peak position, and symmetry of the multifractal spectrum serve as classification features.

[0051] The Higuchi algorithm is an efficient method for estimating the fractal dimension, calculated directly in the time domain without phase space reconstruction. The Higuchi algorithm divides the time series into multiple subsequences, calculates the length of each subsequence, and performs regression analysis on the average length of different time intervals to obtain the fractal dimension. The Higuchi algorithm has low computational complexity, is not stringent about data length, and is suitable for real-time applications. The parameters of the Higuchi algorithm include the maximum time interval, typically one-tenth of the signal length, and the range of the regression analysis, which should be selected while ensuring the accuracy of linear fitting.

[0052] The Chhabra-Jensen method is used to calculate the generalized dimension spectrum, extracting generalized dimension spectra of different orders through the power-law scaling behavior of the partition function. The partition function divides the signal into multiple intervals, calculating the sum of the q-th powers of the measure for each interval, where q is the real order. The partition function follows a power-law relationship with the interval scale, and the exponent is related to the generalized dimension spectrum and the order. The Chhabra-Jensen method improves upon the traditional box counting method by introducing measure weights to enhance computational stability and accuracy. The Chhabra-Jensen method can calculate the generalized dimension spectrum of both positive and negative orders, comprehensively characterizing the multifractal properties of the signal.

[0053] The singularity index describes the degree of local irregularity of the signal and is defined as the Holder exponent of the signal in the neighborhood of a certain point. A larger singularity index indicates a smoother signal, while a smaller singularity index indicates a more irregular signal. Due to the intermittent nature of the turbulent jet, leakage signals exhibit different singularity index distributions corresponding to different leakage modes. The singularity index is extracted using the wavelet transform modulus maxima method. The singularity index distribution is concentrated in stable leakage and dispersed in pulsating leakage. The statistical characteristics of the singularity index serve as the characteristic fingerprint of the leakage mode.

[0054] The Hurst exponent quantifies the long-range correlation and self-similarity of the signal, with a value ranging from 0 to 1. A Hurst exponent of 0.5 indicates an uncorrelated random walk signal; a Hurst exponent greater than 0.5 indicates persistence, meaning past growth trends tend to continue in the future; and a Hurst exponent less than 0.5 indicates de-persistence, meaning past growth trends tend to reverse in the future. The Hurst exponent is estimated using rescaled range analysis or detrended fluctuation analysis. Turbulent noise typically has a Hurst exponent close to 0.5, exhibiting short-range correlation, while some leakage signals, due to the hysteresis effect of pressure regulation, have a Hurst exponent greater than 0.5, exhibiting long-range correlation. The difference in Hurst exponents is used to distinguish between leakage signals and turbulent noise.

[0055] The fractal feature vector is a multidimensional vector composed of the box dimension, the correlation dimension, the width of the multifractal spectrum, the mean and variance of the singularity index, and the Hurst exponent, comprehensively characterizing the complexity and irregularity of the signal. The fractal feature vector is input into a pattern recognition classifier for automatic leakage type identification. The pattern recognition classifier employs a support vector machine or random forest algorithm, establishing a mapping relationship between the fractal feature vector and the leakage type through supervised learning. Compared to traditional time-domain or frequency-domain features, the fractal feature vector exhibits stronger robustness, is insensitive to noise and signal amplitude variations, and can capture the essential properties of the signal.

[0056] The thermodynamically driven leak sound source intensity inversion processing provides a quantitative leak assessment capability for the entire scheme. Traditional acoustic emission detection methods can only determine the presence and approximate location of a leak, but cannot accurately quantify its severity. The thermodynamically driven leak sound source intensity inversion processing establishes the constitutive relationship between acoustic measurements and fluid dynamic parameters, enabling inverse calculations from acoustic emission energy to the leak orifice diameter and leak mass flow rate. Compared to purely data-driven empirical models, this thermodynamically driven leak sound source intensity inversion processing has stronger interpretability and generalization ability, and can adapt to leak detection needs under different operating conditions. The Bernoulli equation and the isentropic expansion theory provide strict physical constraints, ensuring that the inversion results conform to the laws of thermodynamics and avoiding non-physical interpretations. The introduction of the hydrogen Joule-Thomson coefficient considers the unique thermal properties of hydrogen, accurately describing the temperature-pressure coupling effect during liquid hydrogen flash evaporation, which is not present in conventional gas leak models. The rapid convergence characteristics of the Newton iteration method ensure the real-time performance of the inversion calculation, meeting the needs of online monitoring. The compensation mechanism of the temperature field correction coefficient addresses the nonlinear influence of latent heat of phase change on the acoustic emission energy, significantly improving inversion accuracy. The Bayesian inference framework extends deterministic inversion to probabilistic inference, providing not only point estimates of the leak aperture but also quantifying the uncertainty of the estimate. Multi-sensor data fusion reduces single-point measurement errors and provides confidence intervals for risk assessment in decision-making. The thermodynamically driven leak sound source intensity inversion process upgrades leak detection from qualitative judgment to quantitative analysis, providing a scientific basis for maintenance decisions and avoiding economic losses or safety hazards caused by misjudgments.

[0057] The fractal theory-based turbulence noise feature extraction provides a mathematical tool for pattern recognition of nonlinear signals. While traditional Fourier analysis and wavelet analysis can extract frequency and time-frequency domain features, linear feature extraction methods struggle to capture the essential properties of chaotic signals like turbulence noise, which possess multi-scale self-similar structures. This fractal theory-based turbulence noise feature extraction, starting from the geometric complexity and scale invariance of the signal, reveals the inherent differences between turbulence noise and leakage signals in their space-filling patterns and local singularity distributions. The box dimension and correlation dimension, as global fractal measures, reflect the dimensional characteristics of the signal trajectory in phase space. Turbulence noise, due to its high-dimensional chaotic attractor, exhibits higher box and correlation dimensions, while leakage signals, constrained by valve geometry, show lower box and correlation dimensions. This dimensional difference forms the basis for signal classification. The multifractal spectrum further refines the fractal description, characterizing the signal's metric distribution at different scales through the generalized dimension spectrum. The multi-scale vortex structure of turbulent noise results in a broad multifractal spectrum, while the relatively singular sound source mechanism of leakage signals leads to a narrower multifractal spectrum. The width of the multifractal spectrum becomes an effective feature for distinguishing between the two types of signals. The singularity index extracts the local irregularity of the signal. The intermittent characteristics of the leakage signal are reflected in the singularity index distribution corresponding to different leakage modes, and different leakage modes have unique singularity fingerprints. The Hurst index reveals the temporal correlation of the signal. The dynamic response of pressure regulation during leakage results in long-range correlation, while the randomness of turbulent noise manifests as short-range correlation. The difference in the Hurst index provides a criterion for time-domain discrimination. The efficiency of the Higuchi algorithm and the Chhabra-Jensen method ensures the real-time calculation of the fractal feature vector, meeting the timeliness requirements of online detection. The fractal feature vector compresses the time series into a low-dimensional parameter space through dimensionality reduction mapping, significantly reducing computational and storage overhead while preserving the essential characteristics of the signal. The fractal theory-based turbulence noise feature extraction enables signal classification to no longer rely on manually set thresholds and empirical rules, but rather on automatic identification based on the inherent geometric structure of the signal. This improves the adaptability and robustness of the detection, and can still accurately distinguish between leakage signals and turbulence noise under strong background noise and time-varying conditions. It significantly reduces the false alarm rate and false negative rate, and improves the reliability and practicality of the leakage detection system.

[0058] In addition, the present invention also provides a method for forming a liquid hydrogen C-type ball valve internal leakage acoustic emission identification system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-mentioned method when running in the computer.

[0059] The specific implementation methods of the above steps are described in detail below.

[0060] The specific implementation of step S01 involves uniformly arranging an array of acoustic emission sensors on the outer surface of the C-type ball valve seat for liquid hydrogen. First, the circumferential centerline of the valve seat is determined. Eight positions are uniformly selected on the plane containing the centerline, with a central angle of 45 degrees between any two adjacent positions. Then, a broadband acoustic emission sensor is installed at each position. The sensor is fixed to the valve seat surface using a low-temperature resistant adhesive to ensure close contact between the sensor and the valve seat surface to guarantee sound wave transmission efficiency. The frequency response range of each broadband acoustic emission sensor is 10kHz to 500kHz, which can cover the broadband acoustic emission signals generated by liquid hydrogen leakage. The circumferentially distributed array structure provides the geometric basis for subsequent spatial spectrum estimation and multi-source localization. Acoustic emission signals from different positions are collected by multiple sensors, providing necessary observation data for blind source separation and time difference of arrival calculation.

[0061] The specific implementation of step S02 involves acquiring the raw acoustic emission signal output from the acoustic emission sensor array and performing adaptive multi-scale wavelet packet decomposition. First, the output signals of eight broadband acoustic emission sensors are simultaneously acquired via a data acquisition card at a sampling frequency of 2MHz, obtaining eight channels of raw acoustic emission signal. Then, the Daubechies wavelet is selected as the mother wavelet, and the decomposition level is set to five levels. Wavelet packet decomposition is performed on the raw acoustic emission signal. During the decomposition process, the Shannon entropy is calculated for each node as an information metric. The decision to continue decomposing the node is based on the magnitude of the Shannon entropy. A Shannon entropy greater than a preset threshold of 0.6 is considered acceptable. The nodes continue to be decomposed, and the decomposition stops when the Shannon entropy is less than or equal to the preset threshold of 0.6. Through a 5-layer decomposition, the 500kHz signal bandwidth is divided into 32 frequency band sub-signals of approximately 15.6kHz. The construction of the adaptive decomposition tree allows for more detailed decomposition of frequency bands with concentrated energy, while maintaining a coarser decomposition of frequency bands with dispersed energy, thus improving the adaptability of time-frequency resolution. Wavelet packet decomposition utilizes the time-frequency localization characteristics of wavelet functions, enabling refined analysis of the signal in both the time and frequency domains. It separates broadband noise and leakage characteristic signals into different frequency band sub-signals, laying the foundation for subsequent independent noise reduction processing.

[0062] The specific implementation of step S03 involves hybrid noise reduction processing of the frequency band sub-signals. First, a short-time Fourier transform is performed on each frequency band sub-signal using a Hamming window with a window length of 256 sampling points and an overlap rate of 50%, yielding the time spectrum. Then, the first 0.5 seconds of the leak-free segment from the original acoustic emission signal is extracted as the noise estimation segment. The average power spectrum of the noise estimation segment in the current frequency band sub-signal is calculated as the noise reference. The power spectrum of each time-frequency point in the time spectrum is subtracted from the noise reference power spectrum and multiplied by an over-subtraction factor of 1.3. When the subtraction result is negative, it is set to zero, completing the spectral subtraction processing. The spectral subtraction is based on the assumption that the power spectrum of stationary noise can be estimated through statistical averaging. Stationary noise components are suppressed by frequency domain subtraction. Then, a Wiener filter is constructed, with the filter gain function being the signal power spectrum estimate divided by the sum of the signal power spectrum estimate and the noise power spectrum estimate. The signal power spectrum estimate is obtained from the power spectrum after spectral subtraction processing. The Wiener filter gain function is multiplied by the time-frequency spectrum after spectral subtraction to complete the Wiener filtering process. Based on the minimum mean square error criterion, Wiener filtering assigns weights related to the signal-to-noise ratio to different frequency components in the frequency domain, effectively smoothing the musical noise generated by spectral subtraction. Then, the amplitude envelope of the denoised acoustic emission signal is detected, and the periodic component of the envelope is found through the autocorrelation function. The modulation frequency range of the periodic modulation characteristic is 100Hz to 2kHz. The denoised acoustic emission signal is segmented according to the detected modulation period, with each segment being one modulation period in length. All segments are time-domain aligned and then averaged. The amplitude of the coherent leakage characteristic signal is linearly accumulated, while the amplitude of the incoherent random noise increases according to the square root law, further improving the signal-to-noise ratio. The hybrid denoising process combines the speed of spectral subtraction, the smoothness of Wiener filtering, and the enhancement of coherent accumulation. The multi-stage denoising strategy effectively suppresses the broadband strong noise generated by liquid hydrogen flash evaporation.

[0063] The specific implementation of step S04 involves establishing a material dispersion compensation model for the noise-reduced acoustic emission signal and performing dispersion compensation. First, the dispersion curve of the valve seat material at a low temperature of 20K is obtained through experimental measurement or literature review. The dispersion curve describes the relationship between sound velocity and attenuation coefficient with frequency. The dispersion curve is fitted as a piecewise linear function or polynomial function, and a mathematical model of frequency, phase velocity, and attenuation coefficient is established as the material dispersion compensation model. Then, the transfer function of the propagation path is calculated based on the material dispersion compensation model. The transfer function describes the amplitude attenuation and phase change of the sound wave from the leakage source to the sensor in the frequency domain, contributing to noise reduction. The emitted acoustic signal is then subjected to a Fourier transform to obtain its spectrum. This spectrum is divided by the transfer function spectrum. To prevent instability caused by the transfer function approaching zero at certain frequency points, a Tikhonov regularization parameter of 0.01 is introduced. This corrects the division operation to multiplying the spectrum by the conjugate of the transfer function, dividing by the square of the transfer function's modulus, and adding the regularization parameter. This completes the frequency domain deconvolution. An inverse Fourier transform is then performed to recover the time-domain waveform, yielding the original waveform after dispersion compensation. The deconvolution technique, based on linear system theory, eliminates the distortion effects of the propagation path on the signal through inverse filtering. A transfer function frequency domain compensation filter is then constructed. The rate response is the reciprocal of the spectrum of the transfer function. A digital filter is designed using a finite impulse response structure with an order of 256. The filter coefficients are designed using the window function method or frequency sampling method. The compensation filter is applied to the original waveform after dispersion compensation. Gain and phase compensation are applied to different frequency components. Higher frequency components receive greater gain to compensate for propagation attenuation, and phase compensation corrects phase distortion caused by dispersion. Finally, the Wigner-Ville distribution is calculated on the compensated signal. The Wigner-Ville distribution is obtained by the Fourier transform of the instantaneous autocorrelation function of the signal and has high time-frequency resolution. It can clearly display the instantaneous frequency changes of the signal, find the frequency corresponding to the maximum energy for each time slice of the Wigner-Ville distribution, connect these frequency points to form ridges, and calculate the slope, curvature and energy concentration of the ridges. These geometric features are insensitive to the dispersion effect and remain relatively stable at different propagation distances. The ridge slope, ridge curvature and ridge energy concentration are combined into a 32-dimensional dispersion-invariant feature vector. Dispersion compensation and invariant feature extraction solve the waveform distortion and frequency domain feature drift problems caused by acoustic wave dispersion in low-temperature materials, providing stable and reliable feature input for subsequent pattern recognition.

[0064] The specific implementation of step S05 involves inputting the dispersion-invariant feature vector into the leakage feature recognition model for processing. First, the 32-dimensional dispersion-invariant feature vector is normalized to the interval between 0 and 1, and then input into the input layer of the leakage feature recognition model. The input layer passes the feature vector to the convolutional feature extraction layer, which contains three one-dimensional convolutional layers. The first convolutional layer has a kernel size of 5 and 64 channels, performing convolution operations on the input feature vector to extract patterns between five consecutive local features. The convolution result is then processed by the ReLU activation function to introduce non-linearity. The second convolutional layer has a kernel size of 3 and 1 channel. 28. Further convolution is performed on the output of the first layer to extract higher-level feature patterns. The kernel size of the third convolutional layer is 3, and the number of channels is 128. Abstract features are extracted again. The output of the convolutional feature extraction layer is passed to the attention weighting layer. The attention weighting layer adopts a multi-head attention mechanism. The number of attention heads under the current working condition is calculated according to the attention head number adjustment function. The feature dimension normalization value is 32 divided by 32, which equals 1. The signal-to-noise ratio normalization value is the actual signal-to-noise ratio divided by 30dB. The pressure normalization value is the actual working pressure divided by 2MPa. The head number adjustment factor is equal to 1 raised to the power of 0.4 multiplied by the signal-to-noise ratio normalization value raised to the power of 0.3. Multiplying by the pressure normalization value to the power of 0.3, the number of attention heads is determined according to the interval to which the head number adjustment factor belongs. The multi-head attention mechanism calculates the correlation between different dimensions in the feature vector through self-attention, assigning higher weights to important features. The output of the attention weighting layer is passed to the recurrent memory layer. The recurrent memory layer adopts a bidirectional long short-term memory network structure with a hidden layer dimension of 256. It captures the temporal dependency relationship of the feature sequence through a gating mechanism. The long short-term memory network can learn long-term dependency information and avoid the gradient vanishing problem. The output of the recurrent memory layer is passed to the fully connected classification layer. The first fully connected layer contains 128 neurons, and the second fully connected layer contains 3 neurons, corresponding to the three categories of normal state, suspected leakage, and confirmed leakage, respectively. The output of the second fully connected layer is converted into a probability distribution through the softmax function. The sum of the output probability of the confirmed leakage category and the output probability of the suspected leakage category is calculated as the leakage probability value. The category corresponding to the maximum probability among the three categories is selected as the leakage type classification result. The leakage feature recognition model automatically learns the nonlinear mapping relationship between dispersion invariant features and leakage state through deep learning methods, avoiding the limitations of manually setting thresholds and empirical rules, and improving the accuracy and adaptability of recognition.

[0065] The specific implementation of step S06 is as follows: When the leakage probability value is greater than the preset leakage probability threshold of 0.5, the multi-leakage source localization process is initiated. First, the noise-reduced acoustic emission signals of the eight broadband acoustic emission sensors are organized into an eight-row matrix, with each row representing the signal from one sensor. This is input into a fast independent component analysis (CFI) algorithm. The CFI algorithm first centers the mixed signal by subtracting the mean of each row, then whitens the centered signal. Principal component analysis (PCA) is used to calculate the eigenvalues ​​and eigenvectors of the covariance matrix to remove correlations. Next, a fixed-point iterative scheme is used to optimize the unmixed matrix. During the iteration, negative entropy is calculated as a measure of non-Gaussianity. By maximizing negative entropy, statistically independent components are found. After the iteration converges, the unmixed matrix is ​​obtained. Multiplying the unmixed matrix with the whitened signal yields the independent component signals, each corresponding to a leakage source. Then, a spatial grid is constructed for the independent component signals, covering the valve seat surface with a grid resolution of 1 mm. An orthogonal matching pursuit (ORP) algorithm is used for sparse reconstruction. The ORP algorithm iteratively selects the spatial location with the highest correlation to the residual, updates the coefficients of the selected location, and so on. The difference is updated by subtracting the contribution of the selected position from the current residual. This iteration continues until the residual energy is less than the threshold of 0.01 or the maximum number of iterations (50) is reached. The number of non-zero elements in the reconstruction result represents the number of leakage sources, and the positions of these non-zero elements indicate the initial spatial locations of the leakage sources. Then, the covariance matrix of the denoised acoustic emission signals from the eight broadband acoustic emission sensors is calculated. Eigenvalue decomposition is performed on the covariance matrix. Based on the number of leakage sources, the eigenvectors corresponding to the largest eigenvalues ​​are selected as the signal subspace, and the remaining eigenvectors are used as the noise subspace. A spatial spectrum function is constructed. The spectral function is the reciprocal of the steering vector and the projection of the noise subspace. The steering vector is calculated based on the geometry of the circular array and the candidate azimuth angles. The spectral peaks are searched at 1-degree intervals for azimuth angles from 0 degrees to 360 degrees to find the peak positions of the spatial spectral function. The peak positions correspond to the spatial angular positions of each leakage source. The multi-signal classification algorithm is based on the orthogonality of the subspace and can achieve super-resolution spatial spectrum estimation. The resolution is much higher than that of traditional beamforming methods. The combination of blind source separation, sparse reconstruction and spatial spectrum estimation solves the problem of signal aliasing and localization when multiple leakage sources exist at the same time.

[0066] The specific implementation of step S07 involves performing thermodynamically driven leakage sound source intensity inversion processing on the leakage source corresponding to the spatial angular position. First, the distance from the leakage source to each broadband acoustic emission sensor is determined based on the spatial angular position. The acoustic emission energy received by each sensor is calculated, and the acoustic emission energy of the eight sensors is averaged to obtain the average acoustic emission energy. The sound power value is calculated based on the point source model. The sound power value is equal to the average acoustic emission energy multiplied by 4 times pi multiplied by the square of the propagation distance. Then, the constitutive relationship between the leakage orifice diameter, pressure difference, and sound power is established. According to Bernoulli's equation, pressure energy plus kinetic energy equals a constant. The jet velocity at the leakage orifice is calculated. The square of the jet velocity is equal to twice the pressure difference divided by the hydrogen density. The temperature drop of hydrogen passing through the leakage orifice is calculated based on the isentropic expansion theory. The temperature drop is described by the Poisson equation. The power of the pressure ratio is equal to the power of the temperature ratio. The power exponent is the specific heat ratio minus 1 divided by the specific heat ratio. The Joule-Thomson coefficient is obtained by consulting hydrogen property data. The Joule-Thomson coefficient describes the rate of change of temperature with pressure during isenthalpic expansion. The flash evaporation rate is calculated. The rate is calculated by multiplying the pressure drop by the Joule-Thomson coefficient. According to the sound source model, the sound power is proportional to the square of the leakage mass flow rate and the cube of the injection velocity. The leakage mass flow rate is equal to the square of the leakage orifice diameter multiplied by pi divided by 4 multiplied by the hydrogen density multiplied by the injection velocity. Substituting the sound power value and the pressure difference measurement value into the above relationship, a nonlinear equation system about the leakage orifice diameter and leakage mass flow rate is established. The Newton iteration method is used to solve the nonlinear equation system. The initial value estimation is based on the empirical formula that the initial value of the leakage orifice diameter is the quarter power of the sound power value multiplied by the empirical coefficient 0.001. In each iteration, the nonlinear equation system is expanded by Taylor at the current point, retaining the first-order term to obtain the linearized equation. Solving the linearized equation yields the correction amount, and the leakage orifice diameter and leakage mass flow rate are updated. Iteration continues until the correction amount is less than 0.01% or the maximum number of iterations (50) is reached to obtain the estimated value of the leakage orifice diameter. The thermodynamically driven leakage sound source intensity inversion processing establishes a quantitative relationship between acoustic measurements and fluid dynamic parameters through a physical model, realizing the accurate inversion from acoustic emission energy to leakage orifice diameter, and providing a scientific basis for the quantitative assessment of leakage severity.

[0067] The specific implementation of step S08 involves extracting fractal theory turbulence noise features from the denoised acoustic emission signal. First, the box dimension is calculated. The phase space of the denoised acoustic emission signal is reconstructed. The embedding dimension is determined to be 8 using the pseudo-nearest neighbor method. The time delay is determined by the zero point of the autocorrelation function. After phase space reconstruction, the signal trajectory is obtained. The trajectory is covered with boxes of different side lengths, starting from one-tenth of the signal amplitude range and halving each time. The number of boxes required to cover the trajectory for different side lengths is counted. A linear fit is performed on the side length and the number of boxes in a double logarithmic coordinate system. The negative value of the slope of the fitted line is the box dimension. Then, the correlation dimension is calculated by statistically analyzing the distances between point pairs in the phase space, given a radius starting from... Starting with a 100% threshold of the signal amplitude range, incrementing by 1% each time, the correlation integral is calculated based on the proportion of points within a given radius. A linear fit is then performed on the radius and the correlation integral in a log-log coordinate system; the slope of the fitted line represents the correlation dimension. The generalized dimension spectrum is then calculated using the Chhabra-Jensen method. The signal is divided into multiple intervals of length 100 sampling points. The measure of each interval is calculated; the measure is the normalized value of the signal amplitude within the interval. For different orders q, ranging from -5 to +5 with a step size of 0.5, the partition function is calculated. The partition function is the sum of the q-th powers of all interval measures. The partition function is calculated for different interval scales in a log-log coordinate system. Linear fitting is performed on the degree and partition function. The generalized dimension spectrum is calculated based on the fitting slope and order q. The width of the multifractal spectrum is the maximum value minus the minimum value of the generalized dimension spectrum. The singularity index is extracted using the wavelet transform modulus maxima method. Continuous wavelet transform is performed on the denoised acoustic emission signal, with the Gaussian wavelet selected as the wavelet function. Local maxima of the wavelet transform coefficients are found at different scales. The Holder exponent is calculated as the singularity index based on the relationship between the maxima and the scale. The mean and variance of the singularity index are statistically analyzed. The Hurst exponent is estimated using the rescaled range analysis method. The denoised acoustic emission signal is divided into several segments, and the mean of each segment is calculated. The mean is subtracted from the signal within each segment. The cumulative deviation is obtained, the range of the cumulative deviation is calculated, the standard deviation of each signal segment is calculated, and the range divided by the standard deviation is used to obtain the rescaled range. The rescaled range is calculated for different segment lengths. Linear fitting is performed on the segment length and the rescaled range in a double logarithmic coordinate system. The slope of the fitted line is the Hurst exponent. The box dimension, correlation dimension, width of the multifractal spectrum, mean of the singularity index, variance of the singularity index, and Hurst exponent are combined to form a 6-dimensional fractal feature vector. Fractal theory turbulence noise feature extraction uses multiple fractal dimensions and singularity measures to characterize the essential difference between turbulence noise and leakage signals from the perspective of signal geometric complexity and scale invariance, providing robust feature representation for pattern recognition.

[0068] The specific implementation of step S09 involves fusing the fractal feature vector with the leakage probability value and the leakage aperture estimate for judgment. First, the fractal feature vector is input into a support vector machine (SVM) classifier. The SVM classifier establishes a mapping relationship between the fractal feature vector and the leakage type through supervised learning. The classifier outputs a judgment result indicating whether turbulence noise dominates or leakage signal dominates. Then, the leakage probability value, leakage aperture estimate, and classifier output are combined for multi-level judgment. When the leakage probability value is greater than or equal to 0.75, the leakage aperture estimate is greater than 0.1 mm, and the classifier output indicates that the leakage signal dominates, it is determined to be a leak. Leakage is identified by a fusion method. When the leakage probability value is between 0.5 and 0.75, the estimated leakage orifice diameter is between 0.05 mm and 0.1 mm, and the classifier output is dominated by a leakage signal, it is considered a suspected leak. When the leakage probability value is less than 0.5 or the classifier output is dominated by turbulence noise, it is considered a normal state. The fusion judgment combines the leakage probability value from the deep learning model, the leakage orifice diameter estimate from physical inversion, and the turbulence noise discrimination result from fractal features. The fusion of multi-source information improves the reliability of the judgment, reduces the risk of misjudgment by a single indicator, and achieves accurate identification of internal leakage in liquid hydrogen C-type ball valves.

[0069] It should be noted that one of the key technical ideas of this invention is the combination of adaptive multi-scale wavelet packet decomposition and hybrid denoising strategy. Traditional fixed-decomposition wavelet decomposition cannot adapt to the energy distribution differences of broadband noise generated by liquid hydrogen flash evaporation in different frequency bands, resulting in over-decomposition of low-energy frequency bands introducing redundant calculations and under-decomposition of high-energy frequency bands losing detailed features. This invention adaptively determines whether to continue decomposition by calculating the Shannon entropy of each decomposition node, so that frequency bands with concentrated energy obtain more detailed time-frequency resolution, while frequency bands with dispersed energy maintain coarser resolution, thus achieving optimized allocation of computing resources. The hybrid denoising strategy organically integrates the fast noise suppression capability of spectral subtraction, the smoothing capability of Wiener filtering, and the weak signal enhancement capability of coherent accumulation. Addressing the challenge that the noise energy of liquid hydrogen flash evaporation can be tens of times that of the real leakage signal, frequency band independent denoising avoids the masking of useful features by broadband noise, and coherent accumulation utilizes the periodic modulation characteristics of the leakage signal to further improve the signal-to-noise ratio, enabling the weak acoustic emission signal generated by a small leak to be effectively extracted from strong background noise, significantly improving the detection sensitivity of early leaks.

[0070] The second key technical idea of ​​this invention is a dispersion-invariant feature extraction method based on material dispersion compensation. Under low-temperature conditions, the sound velocity and attenuation coefficient of materials exhibit nonlinear changes with frequency, with high-frequency components attenuating faster, leading to severe waveform distortion as the propagation distance increases. Traditional time-domain or frequency-domain features drift at different propagation distances, affecting the consistency of identification. This invention establishes a material dispersion compensation model, uses deconvolution technology and a transfer function frequency-domain compensation filter to reconstruct the original waveform at the leakage source, eliminating waveform distortion caused by dispersion. Then, ridge parameters are extracted through the Wigner-Ville distribution. The geometric features of the ridges are insensitive to dispersion and remain stable at different propagation distances, serving as dispersion-invariant feature inputs to the identification model. This avoids feature inconsistencies caused by differences in propagation paths, improves the generalization ability and robustness of the identification model, and enables signals collected by sensors at different locations from the same leakage source to be accurately identified as belonging to the same category.

[0071] The third key technical approach of this invention is the synergistic application of thermodynamically driven leakage sound source intensity inversion and fractal theory turbulent noise feature extraction. Traditional leakage detection methods can only qualitatively determine the presence or absence of a leak, but cannot quantify its severity. This invention establishes a physical constitutive relationship between leakage aperture and sound power based on Bernoulli's equation and isentropic expansion theory. It combines the Joule-Thomson coefficient of hydrogen to describe the thermodynamic characteristics of the liquid hydrogen flash evaporation process and uses the Newton iteration method to solve the nonlinear equations, achieving quantitative inversion from acoustic emission energy to leakage aperture. This provides a scientific basis for assessing the severity of leaks. At the same time, fractal theory turbulent noise feature extraction uses multi-dimensional fractal measures such as box dimension, correlation dimension, and multifractal spectrum to distinguish turbulent noise from leakage signals from the perspective of signal geometric complexity and scale invariance. This avoids misjudgments caused by non-leakage acoustic emission events such as thermal stress release. Physical inversion provides quantitative leakage aperture estimation, while fractal features provide qualitative signal type discrimination. The two complement each other to form a comprehensive judgment system that combines quantitative and qualitative methods.

[0072] The synergistic effect of the three key technical approaches mentioned above lies in constructing a complete detection chain from signal preprocessing to feature extraction and then to fusion judgment. Adaptive multi-scale wavelet packet decomposition and hybrid noise reduction provide high-quality denoised acoustic emission signals for subsequent processing. Material dispersion compensation and dispersion-invariant feature extraction solve the special problems of sound wave propagation in low-temperature environments, ensuring the stability and reliability of features. Thermodynamic-driven inversion and fractal theory feature extraction provide complementary judgment criteria from both physical model and signal geometry dimensions. Multi-source information fusion significantly improves the accuracy and reliability of judgment. The three technical approaches are interlocked and form a synergistic effect, enabling the present invention to achieve early and accurate identification of minute leaks in the complex low-temperature and high-noise environment of liquid hydrogen C-type ball valves. Compared with traditional methods, it has significant advantages in detection sensitivity, positioning accuracy, and quantitative assessment capabilities, providing reliable technical protection for the safe operation of liquid hydrogen systems.

[0073] It should be noted that this invention also solves the following technical problem: the difficulty in independently locating and distinguishing the intensity of each leakage source when there are multiple leakage sources in a liquid hydrogen C-type ball valve. This invention uses a fast independent component analysis algorithm to perform blind source separation on the denoised acoustic emission signal. Based on the principle of maximizing non-Gaussianity, the mixed signal is decomposed into mutually independent source signals. An optimization problem is constructed in the spatial domain using sparse reconstruction theory. The sparse distribution characteristics of leakage sources are used to achieve super-resolution source number estimation. Then, a multi-signal classification algorithm is used to search for spatial spectrum peaks through the orthogonality of the signal subspace and noise subspace, accurately determining the spatial angular position of each leakage source. For each located leakage source, thermodynamic-driven inversion is performed independently to obtain its respective leakage aperture estimate. This achieves independent identification and quantitative assessment of multiple leakage sources, avoiding the missed detection or misjudgment caused by traditional methods confusing multiple leakage sources into a single strong leakage source.

[0074] Specifically, the principle of this invention is as follows: Targeting the broadband noise characteristics of liquid hydrogen flash evaporation, this invention employs adaptive multi-scale wavelet packet decomposition to separate different frequency components and then independently reduce noise, avoiding the submergence of leakage features caused by traditional full-band processing. In hybrid noise reduction, spectral subtraction suppresses stationary noise while Wiener filtering smooths residual musical noise. Coherent accumulation of periodic modulation characteristics causes the leakage signal amplitude to increase linearly while random noise only increases by the square root. Multi-level processing synergistically improves the signal-to-noise ratio. Material dispersion compensation reconstructs the original waveform through deconvolution and extracts the Wigner-Ville distribution ridge parameters. The geometric characteristics of these parameters remain stable regardless of dispersion changes, ensuring consistency of features across different propagation distances. Thermodynamic inversion establishes the physical relationship between leakage aperture and acoustic power based on Bernoulli's equation and isentropic expansion theory, avoiding the poor generalization problem of purely data-driven models. Fractal feature extraction utilizes the essential differences between turbulent noise and leakage signals in spatial filling complexity and temporal correlation, providing a discrimination criterion independent of amplitude and frequency. Multi-dimensional feature fusion judgment improves detection reliability, thus achieving accurate identification and quantitative assessment of minute leaks under strong noise conditions.

[0075] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0076] The specific implementation of step S01 is to uniformly arrange an acoustic emission sensor array on the outer surface of the liquid hydrogen C-type ball valve seat. The acoustic emission sensor array consists of 8 broadband acoustic emission sensors and is evenly distributed in a circle. The frequency response range of each broadband acoustic emission sensor is 10kHz to 500kHz. The circumferential angle between the sensors is 45 degrees. The installation position is 50 to 100mm away from the valve seat end face. A high-temperature coupling agent is used to ensure good acoustic contact.

[0077] The specific implementation of step S02 is to acquire the original acoustic emission signal output by the acoustic emission sensor array, set the sampling frequency to 1MHz to satisfy the Nyquist sampling theorem, and the sampling duration to 10s to 30s. Adaptive multi-scale wavelet packet decomposition is performed on the original acoustic emission signal, with 5 decomposition layers, decomposing the original acoustic emission signal into 32 frequency band sub-signals, each with a bandwidth of approximately 15.6kHz. The decomposition uses the Daubechies wavelet basis function, and the adaptive criterion is the information entropy and energy distribution of each node, where the information entropy represents the uncertainty of the signal and is calculated using the Shannon entropy formula. Decomposition continues when the node information entropy is greater than the threshold of 0.6 or the energy proportion is greater than 5%.

[0078] The specific implementation of step S03 involves performing hybrid noise reduction processing on the frequency band sub-signals. First, spectral subtraction is performed on each frequency band sub-signal. Then, a short-time Fourier transform is performed on the frequency band sub-signals to obtain the spectrum. A Hamming window is used as the window function, with a window length of 256 sampling points and a frame shift of 128 sampling points. The average power spectrum of the noise segment in the first 1 second is estimated as the noise reference. The formula for power spectrum subtraction is as follows:

[0079] ;

[0080] In the formula, The output signal power spectrum, in units of ; The power spectrum of the input signal, in units of ; This is an estimate of the noise power spectrum, in units of... ; The reference power spectrum value is set to 1. ; It is a subtraction factor, dimensionless, with an empirical value of 1.2 to 1.5; The spectral floor limitation factor is dimensionless and has an empirical value of 0.1. Here is the frequency, in Hz. Then, Wiener filtering is performed. The formula for the Wiener filter transfer function is as follows:

[0081] ;

[0082] In the formula, The transfer function of the Wiener filter is dimensionless. This is the estimated power spectrum of the signal, in units of... ,pass The calculations were performed. Coherent accumulation enhancement was achieved using periodic modulation characteristics. The modulation period was determined by the peak interval of the autocorrelation function, and the number of accumulations ranged from 10 to 50, resulting in the noise-reduced acoustic emission signal.

[0083] The specific implementation of step S04 involves establishing a material dispersion compensation model for the noise-reduced acoustic emission signal and performing dispersion compensation. The relationship between phase velocity and frequency in the material dispersion compensation model is expressed as follows:

[0084] ;

[0085] In the formula, For frequency The corresponding phase velocity, in m / s; The speed of sound at the reference frequency is expressed in m / s and is obtained through an ultrasonic test experiment. The ultrasonic test experiment includes the following steps: Step 1: Perform ultrasonic transmission test on the valve seat material sample at liquid nitrogen temperature; Step 2: Measure the propagation time of the ultrasonic wave in the sample and the sample thickness; Step 3: Calculate the speed of sound by dividing the sample thickness by the propagation time. The empirical value is 5000 m / s. This is a reference frequency, in Hz, with a value of 100kHz. The dispersion coefficient is dimensionless and obtained by fitting experimental data. The empirical values ​​are 0.02 and 0.005, respectively. The dispersion index is dimensionless, with empirical values ​​of 1.5 and 2.5. The relationship between the attenuation coefficient and frequency is expressed as follows:

[0086] ;

[0087] In the formula, For frequency The corresponding attenuation coefficient is expressed in dB / m. The attenuation coefficient at the reference frequency is expressed in dB / m and is obtained through experimental measurement. The experimental measurement includes the following steps: 1. Placing acoustic emission sensors at different propagation distances; 2. Exciting a standard acoustic emission signal; 3. Measuring the signal amplitude attenuation at different distances; 4. Fitting the relationship between the attenuation coefficient and the frequency. The empirical value is 0.5 dB / m. The attenuation exponent is dimensionless and has an empirical value of 1.8. The original waveform is reconstructed using deconvolution. The frequency domain expression of deconvolution is as follows:

[0088] ;

[0089] In the formula, To reconstruct the signal spectrum, the unit is V; The received signal spectrum, in volts (V). The propagation path transfer function is dimensionless. Here, is the Tikhonov regularization parameter, dimensionless, with an empirical value of 0.01. A frequency-domain compensation filter with a transfer function is used for compensation; the filter's transfer function is expressed as follows:

[0090] ;

[0091] In the formula, The transfer function of the compensation filter is dimensionless. The imaginary unit satisfies ; For frequency-dependent time delay compensation, the unit is seconds, through calculate; This is the normalized time constant, in seconds, with a value of 1 second. The propagation distance, in meters, is determined by the sensor installation location. The Wigner-Ville distribution ridge parameters are extracted as dispersion-invariant features. The ridge slope, ridge curvature, and ridge energy concentration are calculated using the first derivative, second derivative, and normalized energy integral of the ridge trajectory, respectively.

[0092] The specific implementation of step S05 is to input the dispersion invariant feature quantity into the leakage feature identification model for processing. The leakage feature identification model outputs the leakage probability value and the leakage type classification result. The leakage probability value is defined as the sum of the probability of confirmed leakage category and the probability of suspected leakage category. The leakage type classification result is the category corresponding to the maximum value among the three category probabilities.

[0093] The specific implementation of step S06 is as follows: When the leakage probability value is greater than the preset leakage probability threshold of 0.5, the fast independent component analysis algorithm is used to perform blind source separation on the denoised acoustic emission signal to obtain independent component signals. The fast independent component analysis algorithm is iteratively optimized by maximizing negative entropy, and the independent component signals are processed by sparse reconstruction theory to estimate the number of leakage sources. The sparse reconstruction is optimized using L1 norm regularization. The spatial spectrum of the denoised acoustic emission signal of the acoustic emission sensor array is estimated using a multi-signal classification algorithm to obtain the spatial angular position of each leakage source. The spatial spectrum function is expressed as follows:

[0094] ;

[0095] In the formula, For spatial spectral functions, the unit is . ; This is the azimuth angle, in degrees. Azimuth The corresponding guide vector is expressed as follows: ,in For the first The radial distance of each sensor relative to the center of the array, in meters. The wavelength corresponding to the center frequency, in meters (m), superscript. Indicates transpose; The noise subspace feature matrix, in units of m, is constructed by extracting noise-related eigenvectors after eigenvalue decomposition of the received signal covariance matrix; superscript This indicates the conjugate transpose. The spectral peak positions correspond to the spatial angular positions of each leakage source.

[0096] The specific implementation of step S07 involves performing thermodynamically driven leakage sound source intensity inversion processing on the leakage source corresponding to the spatial angular position. Based on Bernoulli's equation and the isentropic expansion theory, a constitutive relationship between the leakage orifice diameter, pressure difference, and sound power is established. The leakage mass flow rate is expressed as follows:

[0097] ;

[0098] In the formula, Leakage mass flow rate, in kg / s; For reference mass flow rate, the unit is kg / s, and the value is 0.001 kg / s; The flow coefficient is dimensionless and has an empirical value of 0.6 to 0.8. The reference flow coefficient is dimensionless and has a value of 1. The cross-sectional area of ​​the leakage orifice, in units of ; For reference cross-sectional area, the unit is... The value is ; This refers to the density of hydrogen on the high-pressure side, in units of... ; For reference density, the unit is... The value is 1 ; This is the pressure difference, expressed in Pa. The reference pressure difference is expressed in Pa, with a value of 1 Pa. The relationship between acoustic power and leakage mass flow rate is as follows:

[0099] ;

[0100] In the formula, Sound power, measured in W; The reference sound power is expressed in W, and its value is 1W. The acoustic radiation efficiency is dimensionless and has an empirical value of 0.001 to 0.01. For reference acoustic radiation efficiency, it is dimensionless and takes a value of 1; The jet velocity is expressed in m / s. calculate; The reference velocity is in m / s, with a value of 1 m / s. The acoustic power is calculated from the measured acoustic emission energy. The estimated leakage orifice diameter is obtained by solving the aforementioned nonlinear equations using the Newton-Raphson iteration method. The iteration convergence threshold is 0.01%, and the maximum number of iterations is 50.

[0101] The specific implementation of step S08 involves extracting fractal theory turbulence noise features from the denoised acoustic emission signal, calculating the box dimension, correlation dimension, and multifractal spectrum of the denoised acoustic emission signal, estimating the fractal dimension using the Higuchi algorithm, calculating the generalized dimension spectrum using the Chhabra-Jensen method, and extracting the singularity index and Hurst index to construct a fractal feature vector. The fractal feature vector consists of the box dimension, correlation dimension, multifractal spectrum width, mean of the singularity index, variance of the singularity index, and Hurst index.

[0102] The specific implementation of step S09 is to fuse the fractal feature vector with the leakage probability value and the leakage aperture estimate for judgment. When the leakage probability value is greater than or equal to 0.75 and the leakage aperture estimate is greater than 0.1 mm, it is determined to be a confirmed leak. When the leakage probability value is between 0.5 and 0.75 and the leakage aperture estimate is between 0.05 mm and 0.1 mm, it is determined to be a suspected leak. When the leakage probability value is less than 0.5, it is determined to be a normal state.

[0103] It should be noted that the variables involved in this embodiment 1 are explained in detail in Table 1.

[0104] Table 1. Variable Explanation Table

[0105]

[0106] To better understand and implement this invention, the following is a specific application scenario example 2: A technical team is responsible for the safety monitoring of a liquid hydrogen storage and transportation facility. This facility is equipped with DN150 C-type ball valves for liquid hydrogen flow control. The valves operate continuously under cryogenic conditions of 20K and a pressure differential of 1.5MPa. Due to the shrinkage of the valve seat material and wear of the sealing surface under cryogenic conditions, there is a risk of minor internal leakage. The technical team decided to use acoustic emission technology to establish an online internal leakage monitoring system.

[0107] The technical team evenly installed eight broadband acoustic emission sensors (model PAC-WD) on the circumference of the outer surface of the ball valve seat. The sensors have a frequency response range of 10kHz to 500kHz and a sensitivity of 65dB. The sensors are fixed using a cryogenic coupling agent, with the coupling layer thickness controlled to within 0.2mm. The data acquisition system's sampling frequency was set to 2MHz, with each channel synchronously acquiring data for 5 seconds, corresponding to 10 million sampling points. During monitoring, the upstream pressure of the valve was 1.52MPa, the downstream pressure was 0.18MPa, the pressure difference was 1.34MPa, and the liquid hydrogen temperature was 20.3K.

[0108] The acquired raw acoustic emission signal amplitude ranged from -0.8V to +0.8V, containing broadband noise from 10kHz to 500kHz generated by liquid hydrogen flash evaporation and low-frequency interference from valve mechanical vibration. The technical team performed a 5-level adaptive multi-scale wavelet packet decomposition on the raw signal, selecting the Daubechies 8 wavelet as the mother wavelet, and adaptively determining the decomposition tree structure based on the Shannon entropy of each node signal. After decomposition, 32 frequency band sub-signals were obtained, each with a bandwidth of 15.625kHz. The first band covered 0 to 15.625kHz, and the 32nd band covered 484.375kHz to 500kHz. Energy analysis showed that the leakage characteristic signal was mainly concentrated in the 8th to 18th frequency bands, corresponding to a frequency range of 109.375kHz to 281.25kHz.

[0109] The 32 frequency band sub-signals were sequentially subjected to hybrid noise reduction processing. First, spectral subtraction was used to estimate the noise power spectrum of each frequency band, with the first 0.5 seconds of data selected as the noise segment, and the over-subtraction factor set to 1.35. After spectral subtraction, the signal-to-noise ratio (SNR) improved from the initial 8.2 dB to 14.6 dB, but residual musical noise remained. Subsequently, a Wiener filter was applied for secondary noise reduction. The transfer function of the Wiener filter was adaptively adjusted based on the prior SNR, further improving the SNR to 21.3 dB, and essentially eliminating the musical noise. The technical team detected periodic amplitude modulation in the denoised signal using Hilbert transform, with a modulation frequency of 486 Hz, corresponding to the pressure pulsations during liquid hydrogen injection. Based on this modulation frequency, coherent accumulation enhancement was performed, accumulating 120 modulation cycles of the signal. After coherent accumulation, the amplitude of the leakage characteristic signal was increased to 2.8 times its original value, and the SNR reached 28.7 dB.

[0110] To address the dispersion characteristics of 304 stainless steel valve seat material at 20K, the technical team established a material dispersion compensation model. The phase velocity dispersion curve of this material at 20K was measured using ultrasonic pulse-echo experiments. The phase velocity at 100kHz was 5680m / s, and at 300kHz it was 5520m / s, with the attenuation coefficient increasing from 0.18dB / m at 100kHz to 0.52dB / m at 300kHz. Based on the measured dispersion curve, a transfer function frequency domain compensation filter was constructed. The filter adopted a 256th-order finite impulse response structure, applying phase and amplitude compensation to different frequency components. The compensated signal was then reconstructed using Tikhonov regularized deconvolution to reconstruct the original waveform. The regularization parameter was set to... The Wigner-Ville distribution was calculated for the reconstructed waveform. The distribution plot shows that the energy is mainly concentrated in the 195kHz to 215kHz frequency band, forming a distinct ridge. Figure 2 As shown. The extracted ridge parameters include the ridge slope (2.3 kHz / ms) and ridge curvature. kHz / The three parameters, namely, the ridge energy concentration of 0.73, constitute the first three dimensions of the dispersion-invariant feature, which, together with other time-frequency features, form a 32-dimensional feature vector.

[0111] The technical team input 32-dimensional dispersion-invariant features into a pre-trained leak feature recognition model, which achieved a classification accuracy of 96.8% on the validation set. The model outputs a probability distribution for three categories: normal state (0.08), suspected leak (0.27), and confirmed leak (0.65). The calculated leak probability value is 0.92, exceeding the preset threshold of 0.5. The leak type classification result is determined to be a confirmed leak, and the multi-leakage source localization process is initiated.

[0112] Fast independent component analysis (ICA) was performed on the denoised acoustic emission signals from eight sensors. The analysis converged after 35 iterations, resulting in five independent component signals. An L1-norm regularized sparse reconstruction algorithm was used to process these independent component signals. The spatial grid was divided into 72 azimuth nodes with 5-degree intervals. The reconstruction results showed that non-zero elements appeared in two nodes, estimating two leakage sources. To accurately locate the spatial angular positions of these two leakage sources, the team employed a multi-signal classification algorithm for spatial spectrum estimation. First, the covariance matrix of the eight sensor signals was calculated. Eigenvalue decomposition was then performed on the covariance matrix. The eigenvalues, arranged from largest to smallest, were 4.82, 3.67, 0.31, 0.28, 0.23, 0.19, 0.16, and 0.14. The first two eigenvalues ​​corresponded to the signal subspace, and the last six corresponded to the noise subspace. A spatial spectrum function was constructed using the noise subspace eigenvectors, searching for spectral peaks at 1-degree intervals within the range of 0 to 360 degrees. Figure 3 As shown, the spatial spectrum shows two distinct peaks at 72 degrees and 238 degrees, with peak heights of 32.6 dB and 28.4 dB, respectively. This determines the spatial angular positions of the two leakage sources at 72 degrees and 238 degrees, corresponding to the northeast and southwest directions of the valve seat.

[0113] Thermodynamically driven leakage sound source intensity inversion was performed on the main leakage source at 72 degrees, with sensor number 3 corresponding to this direction. The leakage orifice diameter was established based on Bernoulli's equation and the isentropic expansion theory. Pressure difference Sound power The constitutive relation of liquid hydrogen is given. The Joule-Thomson coefficient for liquid hydrogen at 20 K is -0.034 K / MPa, and the flash rate at a pressure difference of 1.34 MPa is 0.18. The acoustic emission energy measured by sensor 3 is... J, based on the sound source model, the calculated sound power is 7.2 mW. Substituting the pressure difference of 1.34 MPa and the sound power of 7.2 mW into the constitutive equations, the Newton-Raphson iteration method is used for solving. The initial value is set with a leakage orifice diameter of 0.15 mm, an iteration convergence threshold of 0.01%, and a maximum number of iterations of 50. After 12 iterations, convergence is achieved, yielding an estimated leakage orifice diameter of 0.132 mm for the 72-degree leakage source and an estimated leakage mass flow rate of [value missing]. kg / s. Using the same method to process the secondary leakage source at 238 degrees, the estimated leakage orifice diameter was calculated to be 0.089 mm, and the estimated leakage mass flow rate was... kg / s.

[0114] To further verify the reliability of the leak detection, the technical team extracted turbulent noise features from the denoised acoustic emission signal using fractal theory. The Higuchi algorithm was used to calculate the box-dimensionality, with the maximum time interval set to one-tenth of the signal length, i.e., 500,000 sampling points. Regression analysis was then performed to assess the goodness of linear fit. The box dimension was found to be 2.16, with an embedding dimension of 0.987. The correlation dimension was calculated using the Grassberger-Procaccia algorithm. The embedding dimension was determined to be 7 using the pseudo-nearest neighbor method, and the time delay was determined to be 23 sampling points using the mutual information method, resulting in a correlation dimension of 1.94. The generalized dimension spectrum was calculated using the Chhabra-Jensen method, with the order q ranging from -5 to +5 and a step size of 0.5. The generalized dimension spectrum curve showed a monotonically decreasing trend, with the zero-order generalized dimension equal to the box dimension of 2.16. The multifractal spectrum width was defined as the difference between the +5 and -5 order generalized dimensions, calculated to be 0.68. The singularity index was extracted using the wavelet transform modulus maxima method. The singularity index ranged from 0.42 to 0.88, with a mean of 0.61 and a variance of 0.053. The Hurst exponent was estimated using detrended fluctuation analysis with window sizes ranging from 10 to 10,000 sampling points. In a double logarithmic coordinate system, the fluctuation function exhibited a linear relationship with the window size, with a slope of 0.68, resulting in a Hurst exponent of 0.68. The box dimension (2.16), correlation dimension (1.94), multifractal spectrum width (0.68), singularity index mean (0.61), singularity index variance (0.053), and Hurst exponent (0.68) were combined to form a 6-dimensional fractal eigenvector, as shown below. Figure 4 As shown.

[0115] The technical team fused the fractal feature vector with the leakage probability value and the estimated leakage orifice diameter. The leakage probability value (0.92) was greater than or equal to 0.75, and the estimated orifice diameter of the main leakage source (0.132 mm) was greater than 0.1 mm. Based on the fusion judgment rules, this was determined to be a confirmed leak. The box dimension (2.16) and correlation dimension (1.94) were both within the typical range of 1.5 to 2.5 for leakage signals, significantly lower than the 2.5 to 3.5 range for turbulent noise. The Hurst exponent (0.68) was greater than 0.5, indicating long-range correlation, consistent with the pressure regulation hysteresis effect during leakage. The consistency of the fractal features further verified the accuracy of the confirmed leak judgment. Based on the test results, the technical team arranged for the ball valve to be shut down for maintenance. After disassembly and inspection, wear defects were found on the valve seat at both the 72-degree and 238-degree directions, with defect sizes basically matching the estimated values.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying internal leakage in a liquid hydrogen C-type ball valve based on acoustic emission frequency characteristics, characterized in that, A circular array of eight broadband acoustic emission sensors is uniformly arranged on the outer surface of the C-type ball valve seat of liquid hydrogen to collect the original acoustic emission signal. The original acoustic emission signal is then subjected to adaptive multi-scale wavelet packet decomposition into frequency band sub-signals, followed by hybrid noise reduction using spectral subtraction and Wiener filtering. The noise-reduced acoustic emission signal is then enhanced by coherent accumulation using periodic modulation characteristics. A material dispersion compensation model is established to compensate for the dispersion of the noise-reduced acoustic emission signal, and the Wigner-Ville distribution ridge parameters are extracted as dispersion-invariant features and input into the leakage feature identification model. The model outputs leakage probability values ​​and leakage type classification results. When the leakage probability value is greater than a preset leakage probability threshold, a fast independent component analysis algorithm is used for blind source separation, and the number of leakage sources is estimated using sparse reconstruction theory. A multi-signal classification algorithm is used for spatial spectrum estimation to obtain the spatial angular position of each leakage source. The leakage source corresponding to the spatial angular location is subjected to thermodynamic-driven leakage sound source intensity inversion processing. Based on Bernoulli's equation and isentropic expansion theory, a constitutive relationship between leakage orifice diameter, pressure difference, and sound power is established. The flash rate is calculated by combining the Joule-Thomson coefficient of hydrogen. The sound power value is calculated by measuring the acoustic emission energy. The leakage orifice diameter is estimated by solving the nonlinear equation system using the Newton iteration method. The fractal theory turbulence noise features are extracted from the denoised acoustic emission signal. The box dimension, correlation dimension, and multifractal spectrum of the denoised acoustic emission signal are calculated. The fractal dimension is estimated by the Higuchi algorithm. The generalized dimension spectrum is calculated by the Chhabra-Jensen method. The singularity index and Hurst index are extracted to construct fractal feature vectors. The fractal feature vectors are fused with the leakage probability value and the leakage orifice diameter estimate to determine the leakage state.

2. The method according to claim 1, characterized in that, The frequency response range of the broadband acoustic emission sensor is 10kHz to 500kHz.

3. The method according to claim 2, characterized in that, The adaptive multi-scale wavelet packet decomposition has 5 decomposition layers, which decomposes the original acoustic emission signal into 32 frequency band sub-signals.

4. The method according to claim 3, characterized in that, The hybrid noise reduction process involves first performing spectral subtraction on each frequency band sub-signal and then Wiener filtering.

5. The method according to claim 4, characterized in that, The oversubtraction factor used in the spectral subtraction processing ranges from 1.2 to 1.

5.

6. The method according to claim 5, characterized in that, The modulation frequency range of the periodic modulation characteristic is from 100Hz to 2kHz.

7. The method according to claim 6, characterized in that, Material dispersion compensation employs deconvolution technology to reconstruct the original waveform, and the deconvolution technology uses the Tikhonov regularization method.

8. The method according to claim 7, characterized in that, The transfer function frequency domain compensation filter adopts a finite impulse response structure with orders ranging from 128 to 256.

9. The method according to claim 8, characterized in that, The ridge parameters of the Wigner-Ville distribution include ridge slope, ridge curvature, and ridge energy concentration.

10. The method according to claim 9, characterized in that, The structure of the leakage feature recognition model is a cascaded architecture consisting of an input layer, a convolutional feature extraction layer, an attention weighting layer, a recurrent memory layer, and a fully connected classification layer.

Images