Ultrasonic detection method for leakage rate of fuel control valve of aero-engine

Abstract

The invention relates to the technical field of ultrasonic detection, and provides an aero-engine fuel control valve leakage amount ultrasonic detection method, which forms a coherent technical link through leakage ultrasonic signal acquisition, leakage signal separation, leakage feature extraction, leakage area positioning, offset compensation and cross validation of a sensor, and the links are mutually verified and optimized. A leakage signal is separated through a dynamic threshold value, the identification degree of a deep belief network is enhanced through multi-dimensional features, environment interference is compensated and corrected through working condition parameters, positioning deviation caused by single-link error accumulation in a traditional method is effectively restrained, accurate capturing and positioning of a micro leakage signal are achieved, and the method is suitable for large-scale popularization and application. The leakage thermodynamic diagram integrates leakage rate grading, coarse positioning area and micro leakage point information, the analysis and decision-making time of maintenance personnel is greatly shortened through a visual presentation mode, and rapid interpretation of fault positions and severity is achieved.

Description

Technical Field

[0001] The present application relates to the field of ultrasonic detection technology, and in particular to a method for ultrasonically detecting leakage of an aircraft engine fuel control valve. Background Art

[0002] Leak detection in aircraft engine fuel control valves is crucial to flight safety. Traditional detection methods primarily include pressure drop testing, flow monitoring, and contact ultrasonic testing. These methods, for example, rely on measuring changes in fuel line pressure and flow to determine leaks, or rely solely on single-point sensor signal acquisition, which requires coupling agents and is prone to failure in high-temperature and vibration environments, making omnidirectional leak detection difficult. However, they have not yet addressed the issue of ultrasonic detection of fuel control valve leakage through leak signal separation, hierarchical location optimization, and operating condition compensation. Summary of the Invention

[0003] To address the shortcomings of the prior art, the present application provides an ultrasonic detection method for leakage of an aircraft engine fuel control valve. The method comprises: arranging an ultrasonic sensor array to acquire a leakage ultrasonic signal of the fuel control valve; separating the background noise of the leakage ultrasonic signal by wavelet packet transform; dynamically setting a threshold based on the energy probability of the background noise to separate the leakage signal; and iteratively extracting leakage features by a matching pursuit algorithm. The leakage features include sudden changes in acoustic emission energy, resonant frequency offset, and harmonic distortion rate.

[0004] A mapping relationship between leak characteristics and leakage volume is established based on a deep belief network, and a leakage volume classification is determined. A cross-correlation algorithm is used to determine the coarse localization area of ​​the leak sound source based on the time difference of the ultrasonic sensor array. Within this coarse localization area of ​​the leak sound source, a compressed sensing algorithm is used to reconstruct the leakage sound field distribution inside the fuel control valve from the leakage ultrasonic signal with a 30% undersampling rate to locate the micro-leak. The leakage volume classification is then combined with the coarse localization area of ​​the leak sound source and the micro-leak location to generate a leak heat map.

[0005] Fuel temperature and engine speed are monitored in real time. A bias compensation term is determined based on the fuel temperature and engine speed. The bias compensation term and the leakage feature are concatenated into a joint vector, which is then input into a deep belief network. When the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter is greater than a deviation threshold, cross-validation of the ultrasonic sensor array is triggered.

[0006] As an optional implementation, the iterative extraction logic of the leakage feature includes:

[0007] Based on the structural characteristics of the fuel control valve, an initial composite atom library is constructed;

[0008] Initialize the residual For leakage signal, search the atom with the largest inner product with leakage signal from the initial composite atom library , calculate the residual ;

[0009] The residual Match with the atoms in the initial composite atom library and select the atoms with the residual The atom with the largest inner product , update the composite atom library;

[0010] When the residual energy ratio is less than 5% or the number of iterations is met, the iteration is stopped, the composite atom library is output, and leakage features are extracted from the composite atom library;

[0011] The acoustic emission energy mutation represents the difference in atomic energy between adjacent iterations, the resonant frequency offset represents the change between the atomic resonant frequency and the normal resonant frequency, and the harmonic distortion rate represents the ratio of the atomic harmonic energy to the fundamental wave energy.

[0012] As an optional implementation manner, the leakage signal separation logic includes:

[0013] The acquired leakage ultrasonic signal is decomposed into 3-layer wavelet packets to generate 8 sub-bands;

[0014] Determine the energy entropy of each sub-band, determine the noise frequency band based on the energy entropy, and separate the background noise of the leaked ultrasonic signal;

[0015] The energy probability of the background noise is determined according to the energy mean and energy standard deviation of the noise frequency band, so as to dynamically set the energy threshold, and the leakage ultrasound signal with energy greater than the energy threshold in each sub-band is regarded as a candidate leakage signal;

[0016] Performing short-time Fourier transform on the candidate leakage signals, screening the candidate leakage signals to separate the leakage signals.

[0017] As an optional implementation manner, the logic for establishing the mapping relationship between the leakage characteristics and the leakage amount includes:

[0018] Collecting leakage characteristics and actual leakage amount obtained by the fuel flow meter, normalizing the leakage characteristics and actual leakage amount to obtain a standard data set;

[0019] Construct a deep belief network with multiple layers of RBM stacked together, initialize the weights and biases of each RBM layer, and pre-train the RBM layer by layer through unsupervised learning;

[0020] The deep belief network is fine-tuned in a supervised manner through the back-propagation algorithm, with the mean square error as the loss function, and the weights and biases of the deep belief network are updated through stochastic gradient descent;

[0021] After multiple iterative training, when the loss function converges to the loss threshold, the mapping relationship between leakage features and leakage amount in the deep belief network is output.

[0022] As an optional implementation manner, the sub-logic for determining the coarse positioning area of ​​the leakage sound source includes:

[0023] The time difference of the leakage ultrasonic signals between each pair of sensors is calculated by a cross-correlation algorithm based on the timestamp of the leakage ultrasonic signal received by each sensor in the ultrasonic sensor array;

[0024] The distance between each pair of sensors is determined based on the time difference, and the hyperbola of each pair of sensors is obtained. By combining different sensor pairs, the hyperbola equation is obtained.

[0025] Based on the structural characteristics of the fuel control valve and the physical characteristics of the leakage, the solutions of the hyperbolic equation are screened, and adjacent solutions with similar characteristics are merged into the coarse localization area of ​​the leakage sound source through cluster analysis.

[0026] The time difference is randomly perturbed multiple times by Monte Carlo simulation method, and the solution of the hyperbola equation is re-screened to determine the confidence interval of the rough positioning area of ​​the leakage sound source.

[0027] As an optional implementation manner, the micro-leak location locating sub-logic includes:

[0028] In the coarse location area of ​​the leakage sound source, the ultrasonic sensor array acquires the leakage ultrasonic signal at a 30% undersampling rate to obtain the undersampling signal, and the undersampling signal is preprocessed;

[0029] Based on the structural characteristics of the fuel control valve, the internal acoustic field of the fuel control valve is divided into grid cells. The under-sampled signal is sparsely represented. The sparse coefficient of each grid cell is iteratively updated through a threshold iteration algorithm to reconstruct the leakage acoustic field distribution of the fuel control valve from the under-sampled signal.

[0030] The acoustic energy density of each grid unit of the leakage sound field distribution is calculated, the energy abnormal grids are identified, the energy abnormal grids are clustered, the adjacent energy abnormal grids are merged, and the micro-leak point is determined in combination with the leakage amount classification, and the centroid coordinates of the micro-leak point are calculated as the micro-leak position.

[0031] As an optional implementation manner, the determination logic of the bias compensation term includes:

[0032] Real-time monitoring of fuel temperature and engine speed;

[0033] The difference between the fuel temperature and the standard temperature is weighted and summed with the engine speed to obtain the bias compensation term;

[0034] The bias compensation term and the leakage feature are concatenated into a joint vector.

[0035] As an optional implementation, the cross-validation logic of the ultrasound sensor array includes:

[0036] Calculate in real time the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter;

[0037] Configure a deviation threshold. When the deviation is greater than the deviation threshold, cross-validation of the ultrasonic sensor array is triggered.

[0038] At the same time, the cross-validation of the ultrasonic sensor array is determined based on the changes in fuel temperature and engine speed;

[0039] When the cross-validation of the ultrasonic sensor array is triggered, the spare sensing channel of the ultrasonic sensor array is activated to reacquire the leakage ultrasonic signal to extract a new leakage feature;

[0040] The similarity between the new leakage feature and the joint vector is calculated by Euclidean distance, and the similarity is compared with the similarity threshold to determine whether to retrain the deep belief network.

[0041] As an optional implementation, the generation logic of the leakage heat map includes:

[0042] Integrate the leakage level classification with the coarse localization area of ​​the leakage sound source and the micro-leak position into the data structure, and assign quantitative values ​​of the leakage level classification to the coarse localization area of ​​the leakage sound source and the micro-leak position;

[0043] The three-dimensional coordinates of the fuel control valve are converted into two-dimensional coordinates through orthographic projection. The two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position are calibrated and adjusted according to the arrangement of the ultrasonic sensor array.

[0044] The quantitative values ​​of the leakage level are mapped to different colors, and the two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position and the mapped colors are integrated to generate a leakage heat map.

[0045] As an optional embodiment, an ultrasonic sensor array is arranged in a non-contact manner on the outer wall of the metal pipe section downstream of the fuel control valve. The ultrasonic sensor array includes at least two sensor groups, and the sensor groups include a longitudinal wave sensor group and a shear wave sensor group. The longitudinal wave sensor group is arranged obliquely at a first critical angle of incidence to obtain the longitudinal leakage ultrasonic signal of the fuel control valve, and the shear wave sensor group is arranged obliquely at a second critical angle of incidence to obtain the transverse leakage ultrasonic signal of the fuel control valve.

[0046] Compared with the existing technology, the beneficial effects of the present application are: by acquiring leakage ultrasonic signals, separating leakage signals, extracting leakage features, locating leakage areas, bias compensation and cross-validation of sensors, a coherent technical chain is formed, each link is mutually verified and optimized, leakage signals are separated by dynamic thresholds, the recognition of deep confidence networks is enhanced by multi-dimensional features, and environmental interference is corrected by operating condition parameter compensation, effectively suppressing the positioning deviation caused by the accumulation of errors in a single link in traditional methods, and realizing the accurate capture and positioning of micro-leakage signals. The leakage heat map integrates leakage quantity classification, coarse positioning area and micro-leakage point information, and the visual presentation method greatly shortens the analysis and decision-making time of maintenance personnel, and realizes rapid judgment of fault location and severity.

[0047] By dynamically setting the threshold through wavelet packet transform combined with the energy probability of background noise, leakage signals of different intensities can be adaptively separated. This method can effectively suppress background noise interference such as mechanical vibration and fluid turbulence during aircraft engine operation, accurately retain the high-frequency transient components in the leakage signal, and avoid weak leakage signals being submerged by noise, significantly improving the signal-to-noise ratio and separation reliability; based on the matching pursuit algorithm, iteratively extracting multidimensional features such as acoustic emission energy mutation and resonant frequency offset can fully characterize the physical process of leakage, and the combination of multidimensional features enhances the discrimination of leakage signals.

[0048] The time difference between each pair of sensors is calculated through the cross-correlation algorithm to construct a hyperbola equation, and the compressed sensing algorithm is combined to reconstruct the leakage sound field distribution at a low sampling rate. This combination realizes hierarchical optimization from coarse positioning to precise positioning. The coarse positioning stage narrows the search range and reduces invalid calculations. The micro-positioning stage restores the sound field details through sparse reconstruction, breaking through the traditional method's dependence on high sampling rates, and improving the accuracy of micro-leak positioning while reducing the amount of data processing; the leakage heat map is presented in a visual way, allowing maintenance personnel to quickly identify high-risk leakage areas and their severity without complex data analysis, significantly improving the efficiency and accuracy of fault diagnosis.

[0049] Fuel temperature and engine speed are monitored in real time, and bias compensation terms are constructed through weighted summation and integrated into a deep confidence network. This mechanism can dynamically correct the impact of operating condition changes on leakage characteristics. The compensated deep confidence network can adapt to the differences in leakage signal characteristics under different operating conditions and maintain stable detection performance. Based on a cross-validation mechanism with dual triggers of deviation threshold and operating condition change, combined with redundant design of backup sensing channels, when the prediction deviation of the deep confidence network exceeds the limit, the backup channel is automatically activated to obtain the leakage ultrasonic signal and compare the feature similarity to determine whether the deep confidence network needs to be retrained to avoid false alarms and missed alarms. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be derived from these drawings without inventive work. Among them:

[0051] Figure 1 A flow chart of a method for ultrasonically detecting leakage of an aircraft engine fuel control valve provided in an embodiment of the present application;

[0052] Figure 2 A logic diagram for iterative extraction of leakage features of an ultrasonic detection method for leakage of an aircraft engine fuel control valve provided in an embodiment of the present application;

[0053] Figure 3 A logic diagram for locating micro-leakage positions in the ultrasonic detection method for leakage of an aircraft engine fuel control valve provided in an embodiment of the present application;

[0054] Figure 4 This is a cross-validation logic diagram of the ultrasonic sensor array of the ultrasonic detection method for leakage of an aircraft engine fuel control valve provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application are clearly and completely described below in conjunction with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0056] Example 1

[0057] like Figure 1 As shown, a flow chart of a method for ultrasonically detecting leakage of an aircraft engine fuel control valve is provided for an embodiment of the present application. The method includes:

[0058] S1. Arrange an ultrasonic sensor array to acquire the leakage ultrasonic signal of the fuel control valve. Use wavelet packet transform to separate the background noise of the leakage ultrasonic signal. Dynamically set a threshold based on the energy probability of the background noise to isolate the leakage signal. Simultaneously, use the matching pursuit algorithm to iteratively extract leakage characteristics, including acoustic emission energy mutation, resonant frequency offset, and harmonic distortion rate.

[0059] The ultrasonic signal generated by the leakage of the fuel control valve contains longitudinal wave and shear wave components, and the propagation direction has spatial anisotropy. By orthogonally arranging the longitudinal wave and shear wave sensor groups, the leakage signal can be captured omnidirectionally and the signal integrity can be improved; an ultrasonic sensor array is arranged non-contactly on the outer wall of the metal pipe section downstream of the fuel control valve. The ultrasonic sensor array includes at least two sensor groups, and the sensor groups include a longitudinal wave sensor group and a shear wave sensor group. The longitudinal wave sensor group is arranged obliquely at a first critical angle of incidence to obtain the longitudinal leakage ultrasonic signal of the fuel control valve, and the shear wave sensor group is arranged obliquely at a second critical angle of incidence to obtain the transverse leakage ultrasonic signal of the fuel control valve.

[0060] The calculation formula for the first critical angle is: ,in represents the longitudinal wave speed of the outer wall of the metal pipe section, represents the fuel sound velocity, and the longitudinal leakage ultrasonic signal is obtained by longitudinal wave mode conversion. The calculation formula of the second critical angle is: ,in The transverse wave velocity of the outer wall of the metal pipe section is represented. The transverse leakage ultrasonic signal is captured through the polarization characteristics of the transverse wave. The transverse wave sensor group and the longitudinal wave sensor group are spatially orthogonally distributed at 90 degrees, and the spacing is less than or equal to half of the wavelength corresponding to the center frequency of the leakage ultrasonic signal. This realizes omnidirectional perception of the leakage ultrasonic signal, reduces the signal blind area, improves the subsequent positioning accuracy, and reduces the impact of single sensor failure.

[0061] Specifically, the separation logic of the leakage signal includes:

[0062] The acquired leakage ultrasonic signal is decomposed into 3-layer wavelet packets to generate 8 sub-bands;

[0063] Determine the energy entropy of each sub-band, determine the noise frequency band based on the energy entropy, and separate the background noise of the leaked ultrasonic signal;

[0064] The energy probability of the background noise is determined according to the energy mean and energy standard deviation of the noise frequency band, so as to dynamically set the energy threshold, and the leakage ultrasound signal with energy greater than the energy threshold in each sub-band is regarded as a candidate leakage signal;

[0065] Performing short-time Fourier transform on the candidate leakage signals, screening the candidate leakage signals to separate the leakage signals.

[0066] The leakage ultrasonic signal is a non-stationary signal. The traditional Fourier transform cannot simultaneously characterize the time-frequency characteristics. The wavelet packet transform can perform multi-scale decomposition on the leakage ultrasonic signal to achieve frequency domain separation of noise and effective signal. The Daubechies-8 wavelet basis is used to perform 3-layer wavelet packet decomposition on the original leakage ultrasonic signal to generate 8 sub-bands, where the bandwidth is 1 / 8 of the original leakage ultrasonic signal bandwidth. The frequency interval corresponding to the sub-band is obtained by , Indicates the The frequency interval of the sub-band, represents the minimum frequency, represents the width of the sub-band, ; This decomposes the signal into fine frequency bands, making the background noise and leakage signal present differentiated distribution in the frequency domain, facilitating subsequent screening.

[0067] The signal complexity is measured by energy entropy. The energy entropy value of the noise band corresponding to the background noise is usually greater than the frequency band of the leakage signal, so as to distinguish the effective components; the energy entropy of each sub-band is calculated, and the entropy threshold is set. The entropy threshold is obtained by adding the mean of the energy entropy of all sub-bands to 1.5 times the standard deviation. The sub-band greater than the entropy threshold is judged as the noise band; thus, the noise band is adaptively identified, avoiding misjudgment caused by manual experience in setting the entropy threshold, and improving the separation accuracy.

[0068] The energy of background noise fluctuates with the working conditions, and the dynamic energy threshold can adapt to environmental changes to avoid missed detection or false detection; the energy mean μ and standard deviation σ of the noise frequency band are calculated, and the energy threshold = μ + k × σ, where k is initially set to 3. It needs to be optimized through cross-validation. For each sub-band, the signal segment with energy greater than the energy threshold is marked as a candidate leakage signal; thus, the energy threshold is dynamically adjusted in a complex noise environment to balance the signal extraction sensitivity and anti-interference ability. The candidate leakage signal is used as the short-time Fourier transform input to reduce the amount of calculation and focus on the effective components.

[0069] Through time-frequency analysis, it is verified whether the candidate leakage signal meets the leakage characteristics, such as burst pulses and specific frequency components, to eliminate false signals; the candidate leakage signal is subjected to short-time Fourier transform, and the Hanning window function is selected as the window function. The window length is set to 3 times the signal period. A standard leakage signal is configured, and the similarity between the candidate leakage signal and the standard leakage signal is calculated through the normalized cross-correlation coefficient. The candidate leakage signal with a similarity greater than the similarity threshold is selected as the leakage signal; through secondary screening of the leakage signal, the interference signal is further eliminated to ensure the purity of the final leakage signal. The leakage signal is input into the matching pursuit algorithm to improve the accuracy of feature extraction and reduce redundant calculations.

[0070] Specifically, if Figure 2 As shown, the iterative extraction logic of leakage features includes:

[0071] Based on the structural characteristics of the fuel control valve, an initial composite atom library is constructed;

[0072] Initialize the residual For leakage signal, search the atom with the largest inner product with leakage signal from the initial composite atom library , calculate the residual ;

[0073] The residual Match with the atoms in the initial composite atom library and select the atoms with the residual The atom with the largest inner product , update the composite atom library;

[0074] When the residual energy ratio is less than 5% or the number of iterations is met, the iteration is stopped, the composite atom library is output, and leakage features are extracted from the composite atom library;

[0075] The acoustic emission energy mutation represents the difference in atomic energy between adjacent iterations, the resonant frequency offset represents the change between the atomic resonant frequency and the normal resonant frequency, and the harmonic distortion rate represents the ratio of the atomic harmonic energy to the fundamental wave energy.

[0076] The leakage signal generated by a fuel control valve leak has complex time-frequency characteristics, such as acoustic emission pulses, mechanical resonance, and fluid turbulence harmonics. Therefore, a composite atomic library containing multiple atoms needs to be constructed to adapt to the characteristics of the leakage signal. The atoms include time-domain atoms and frequency-domain atoms. The time-domain atoms include Gaussian envelope sine waves and Chirp linear frequency modulation signals. Gaussian envelope sine waves are used to simulate acoustic emission signals, and Chirp linear frequency modulation signals are used to capture frequency drift. Frequency-domain atoms include bandpass filter group atoms and impulse response atoms. Bandpass filter group atoms are used to match harmonic components, and impulse response atoms are used to locate transient leakage events. The parameters of the atoms are initialized, and the atoms are stored through a tree index. They are hierarchically classified according to frequency-bandwidth-time domain characteristics to accelerate matching retrieval; thus, all types of leakage signal features are covered, leakage features caused by a single atom are avoided, and matching efficiency is improved.

[0077] By iteratively matching atoms with leakage signals, the components of leakage signals are gradually stripped away, and characteristic atoms are extracted to characterize leakage features. Fast Fourier transform is used to calculate time domain convolution, which is the equivalent form of inner product, to reduce the computational complexity of a single match. GPU is used to calculate the inner products of multiple atoms and leakage signals in parallel, while processing other candidate atoms at the same time. The residual update strategy is to select the atom with the largest inner product each time. After that, the residual ,in Indicates the The residual after iterations, the initial residual is , Indicates the The residual after iterations, Indicates the After iterations, select and The atom with the largest inner product is selected, and the atomic parameters, including frequency, amplitude, and phase, are recorded. A minimum energy threshold is set, such as 0.01. If the energy contributed by an atom is less than the minimum energy threshold, it is skipped to reduce invalid iterations. This allows for efficient extraction of leakage signal features and removal of noise components. GPU acceleration shortens the time of a single iteration.

[0078] A single match is difficult to cover all the characteristics of the leakage signal, and the composite atom library needs to be dynamically expanded to adapt to complex leakage patterns. Based on the spectral peak of the residual, a new atom with the center frequency as the peak frequency is generated, and the bandwidth is set to the peak 3dB bandwidth. For the residual containing frequency modulation components, the chirp atom slope is generated as an estimate of the spectral slope. The new atom is compared with the existing composite atom library for similarity. When the cosine similarity between the new atom and the existing composite atom library is greater than 0.9, it is merged to avoid redundancy. Atoms that have not been matched for five consecutive times are eliminated, and the capacity of the composite atom library is maintained within 200 atoms. Incremental storage is used, and only the parameters and indexes of the new atoms are recorded to reduce storage overhead. In this way, the composite atom library can be adaptively expanded to cover the long-tail characteristics, keep the scale of the composite atom library controllable, and avoid the dimensionality disaster.

[0079] Set the convergence condition to avoid excessive iterations, quantify the key leakage features from the final composite atom library, and calculate the residual energy ratio, where the residual energy ratio is calculated as , it is terminated when the residual energy ratio is less than 0.05 or the number of iterations reaches 50. The acoustic emission energy mutation is obtained by calculating the difference in atomic energy in adjacent iterations, and the maximum difference is taken as the acoustic emission energy mutation. The resonant frequency offset is the change between the atomic resonant frequency (that is, the atomic center frequency) and the normal resonant frequency, that is, the absolute value of the difference between the atomic resonant frequency and the normal resonant frequency. The harmonic distortion rate is obtained by calculating the ratio of the atomic harmonic energy to the fundamental wave energy. Then, the atomic features with energy contribution less than 3% are eliminated, and the dominant leakage features are retained, thereby accurately quantifying the leakage features and reducing the data dimension. The standardized leakage features are directly adapted to the deep belief network input, reducing the data preprocessing overhead.

[0080] S2. Based on a deep belief network, a mapping relationship between leakage characteristics and leakage amount is established, and the leakage amount classification is determined. The coarse positioning area of ​​the leakage sound source is determined based on the time difference of the ultrasonic sensor array through a cross-correlation algorithm. Within the coarse positioning area of ​​the leakage sound source, the leakage sound field distribution inside the fuel control valve is reconstructed from the leakage ultrasonic signal with a 30% undersampling rate through a compressed sensing algorithm to locate the micro-leak. The leakage amount classification is combined with the coarse positioning area of ​​the leakage sound source and the micro-leak location to generate a leakage heat map.

[0081] Specifically, the logic for establishing the mapping relationship between leakage characteristics and leakage amount includes:

[0082] Collecting leakage characteristics and actual leakage amount obtained by the fuel flow meter, normalizing the leakage characteristics and actual leakage amount to obtain a standard data set;

[0083] Construct a deep belief network with multiple layers of RBM stacked together, initialize the weights and biases of each RBM layer, and pre-train the RBM layer by layer through unsupervised learning;

[0084] The deep belief network is fine-tuned in a supervised manner through the back-propagation algorithm, with the mean square error as the loss function, and the weights and biases of the deep belief network are updated through stochastic gradient descent;

[0085] After multiple iterative training, when the loss function converges to the loss threshold, the mapping relationship between leakage features and leakage amount in the deep belief network is output.

[0086] The dimensional differences of leakage features are eliminated through normalization, and high-amplitude features are avoided from dominating the training, so that the deep belief network can learn the influence of each leakage feature equally. Simulated noise and frequency offset are added to the leakage features to simulate actual interference scenarios and improve the anti-interference generalization ability of the deep belief network. There is a high-order nonlinear coupling between leakage features and leakage amount, and traditional shallow models are difficult to capture complex mapping relationships. The stacking structure of the deep belief network is adopted, and a three-layer restricted Boltzmann machine stack is set. The top fully connected layer of the deep belief network outputs the leakage amount, and the weights and bias of each layer of restricted Boltzmann machine are initialized. Unsupervised pre-training is used to mine the hidden correlations between leakage features, such as the coordinated changes in acoustic emission energy and harmonic distortion rate. In the supervised fine-tuning stage, the actual leakage amount of the fuel flow meter is used as a benchmark, and the pre-training parameters are corrected through back propagation to achieve leakage features. End-to-end mapping of leakage volume, using mean square error as the loss function, updates the weights and biases of the deep belief network through stochastic gradient descent, sets a dynamic learning rate strategy, initializes the learning rate to 0.001, and adjusts the weights of the restricted Boltzmann machine through the AdamW optimizer. When the loss function converges to the loss threshold during training, the mapping relationship between leakage features and leakage volume in the deep belief network is output. Rapid convergence is achieved in the early stage of training, and fine-tuning is performed in the later stage to balance underfitting and overfitting. The deep representation of the leakage signal is effectively extracted through the hierarchical feature learning mechanism, so that the deep belief network maintains prediction stability. The trained deep belief network outputs the leakage volume, and then sets the leakage volume threshold (minimum micro leakage volume and maximum micro leakage volume) to classify the leakage volume, including high leakage volume, medium leakage volume and micro leakage volume, providing a decision-making basis for subsequent positioning accuracy requirements.

[0087] Specifically, the generation logic of the leakage heat map includes:

[0088] Integrate the leakage level classification with the coarse localization area of ​​the leakage sound source and the micro-leak position into the data structure, and assign quantitative values ​​of the leakage level classification to the coarse localization area of ​​the leakage sound source and the micro-leak position;

[0089] The three-dimensional coordinates of the fuel control valve are converted into two-dimensional coordinates through orthographic projection. The two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position are calibrated and adjusted according to the arrangement of the ultrasonic sensor array.

[0090] The quantitative values ​​of the leakage level are mapped to different colors, and the two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position and the mapped colors are integrated to generate a leakage heat map.

[0091] The abstract leakage level is associated with the spatial position and converted into an intuitive visual expression. The leakage data is stored in a three-dimensional array. Each element is associated with the position coordinate, leakage level and confidence level. For example, a small leakage is level 1, a medium leakage is level 2, and a high leakage is level 3. This forms a basic data structure for visualization. The structured data uniformly encodes multi-source information (leakage level, position coordinates and confidence level) and provides a standardized interface for thermal map rendering.

[0092] The three-dimensional structure of the fuel control valve is complex and needs to be reduced to a two-dimensional plane for display. Parallel orthographic projection is used to preserve the spatial proportional relationship to ensure that the relative distance of the leakage position in the two-dimensional image is not deformed. An adaptive cropping algorithm is introduced to retain the details of leakage-sensitive areas, such as sealing interfaces. A laser target is placed on the surface of the fuel control valve to obtain the actual coordinates, compensate for the positioning deviation caused by the non-ideal arrangement of the sensor array, and calibrate and adjust the two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position. The calibrated two-dimensional coordinates are in line with the reading habits of engineers and ensure the geometric accuracy of the spatial position. The precise coordinate information enables the color distribution of the thermal map to accurately match the actual leakage area, thereby improving the reliability of diagnosis.

[0093] The quantitative values ​​of the leakage level are converted into color codes, and the leakage distribution is displayed in the form of an intuitive heat map. Through the CIELab* color gamut, low leakage (level 1) is mapped to cool colors such as blue, and high leakage (level 3) is mapped to warm colors such as red. Color interpolation algorithms (such as bilinear interpolation) are introduced to smooth the transition. WebGL is used to accelerate rendering, and data is loaded in blocks to reduce memory usage. Interactive functions are added, such as clicking on the heat map node to display detailed information such as the leakage level, location coordinates, and confidence level, allowing operators to quickly identify high-risk leakage areas and shorten maintenance response time.

[0094] Furthermore, the sub-logic for determining the coarse location area of ​​the leakage sound source includes:

[0095] The time difference of the leakage ultrasonic signals between each pair of sensors is calculated by a cross-correlation algorithm based on the timestamp of the leakage ultrasonic signal received by each sensor in the ultrasonic sensor array;

[0096] The distance between each pair of sensors is determined based on the time difference, and the hyperbola of each pair of sensors is obtained. By combining different sensor pairs, the hyperbola equation is obtained.

[0097] Based on the structural characteristics of the fuel control valve and the physical characteristics of the leakage, the solutions of the hyperbolic equation are screened, and adjacent solutions with similar characteristics are merged into the coarse localization area of ​​the leakage sound source through cluster analysis.

[0098] The time difference is randomly perturbed multiple times by Monte Carlo simulation method, and the solution of the hyperbola equation is re-screened to determine the confidence interval of the rough positioning area of ​​the leakage sound source.

[0099] During the propagation of leakage ultrasonic signals, the different distances from the leakage sound source to different sensors will cause differences in the time it takes for the signals to reach each sensor. By calculating the time difference of the leakage ultrasonic signals between each pair of sensors, important information about the location of the leakage sound source can be obtained; the timestamp of the arrival of the leakage ultrasonic signal obtained by each ultrasonic sensor is recorded, and the leakage ultrasonic signals of each pair of sensors are processed by the cross-correlation algorithm. The basic principle of the cross-correlation algorithm is to calculate the correlation between two leakage ultrasonic signals under different time delays. The time delay corresponding to the maximum correlation is the time difference between the two signals. In actual calculations, fast Fourier transform is used to accelerate the cross-correlation calculation; thus, the time difference of the leakage ultrasonic signals between each pair of sensors can be accurately calculated, which provides key input parameters for subsequent hyperbolic positioning, improves the accuracy of time difference calculation, and reduces the influence of noise and interference on the calculation results.

[0100] Based on the propagation speed of the leakage ultrasonic signal and the calculated time difference, the distance difference from the leakage sound source to each pair of sensors can be determined. On a plane, the locus of a point whose distance difference to two fixed points is a constant is a hyperbola. Therefore, by determining the distance difference between each pair of sensors, a hyperbola with this pair of sensors as the focus can be obtained. The combination of multiple sensor pairs can obtain a set of hyperbola equations, and the intersection of these hyperbolas is the location of the leakage sound source. Given the propagation speed of the leakage ultrasonic signal in the medium, the distance difference from the leakage sound source to each pair of sensors can be calculated by multiplying the time difference by the propagation speed. With the position of each pair of sensors as the focus and the distance difference as a constant, the hyperbola of each pair of sensors is obtained. By combining different sensor pairs, a set of hyperbola equations is obtained, and these equations are solved jointly. In this way, the time difference information is converted into a geometric hyperbola equation, and the intersection of multiple hyperbolas is used to narrow the location range of the leakage sound source, providing a basis for subsequent screening and clustering.

[0101] Since there are multiple solutions to the hyperbola equation, some of them do not conform to the actual structure of the fuel control valve and the physical characteristics of the leakage. For example, the solution is located outside the fuel control valve or in a location where leakage is unlikely to occur. Therefore, these solutions need to be screened and unreasonable solutions removed. At the same time, cluster analysis can be used to merge adjacent solutions with similar characteristics into one area, so that the coarse positioning area of ​​the leakage source can be determined more accurately. According to the actual structural drawing of the fuel control valve, its internal and external boundaries and the location where the leakage occurs are determined. The solutions of the hyperbola equation located outside the fuel control valve or in a location where leakage is unlikely to occur are eliminated. The basic idea of ​​the spatial clustering algorithm is to cluster according to the density of data points. Density-connected data points are divided into the same cluster. When performing cluster analysis, an appropriate neighborhood radius and minimum number of points are set, and adjacent solutions with similar characteristics (such as close distance and time difference) are merged into one cluster. Each cluster represents a coarse positioning area for the leakage sound source; thereby eliminating unreasonable solutions and reducing the workload and error of subsequent positioning. By merging adjacent solutions into regions through cluster analysis, the coarse positioning of the leakage sound source is made more accurate and intuitive, facilitating subsequent further positioning. The coarse positioning area of ​​the leakage sound source obtained after screening and clustering provides a search range for the positioning of micro-leakage positions. Subsequent micro-leakage position positioning will be carried out within these coarse positioning areas, narrowing the search range and improving the efficiency and accuracy of positioning.

[0102] In actual measurements, there are certain errors in the calculation of time difference, which will affect the accuracy of the solution of the hyperbolic equation and the coarse positioning area of ​​the leakage sound source. The Monte Carlo simulation method is used to perform multiple random perturbations on the time difference, which can simulate the impact of measurement errors, thereby evaluating the uncertainty of the coarse positioning area of ​​the leakage sound source and determining its confidence interval; according to the measurement error range of the time difference, a random number within the measurement error range is added to the time difference to determine the range of random perturbation. For the time difference after random perturbation, the distance between each pair of sensors is recalculated to obtain a new hyperbolic equation, and the solution of the hyperbolic equation is re-screened. The above steps are repeated multiple times. , for example, 1000 times, and the coarse positioning area of ​​the leakage sound source obtained in each simulation is counted. According to the statistical results, the confidence interval of the coarse positioning area of ​​the leakage sound source is determined, such as the 95% confidence interval; thereby taking into account the impact of measurement error on the positioning result, providing an uncertainty assessment for the coarse positioning area of ​​the leakage sound source. The determination of the confidence interval can enable users to understand the reliability of the positioning result and provide a reference for subsequent decision-making. The confidence interval of the coarse positioning area of ​​the leakage sound source can be used as reference information for micro-leak position positioning. In the process of micro-leak position positioning, the search strategy and algorithm parameters are adjusted according to the size and position of the confidence interval to improve the positioning accuracy.

[0103] Furthermore, if Figure 3As shown, the micro-leak location positioning sub-logic includes:

[0104] In the coarse location area of ​​the leakage sound source, the ultrasonic sensor array acquires the leakage ultrasonic signal at a 30% undersampling rate to obtain the undersampling signal, and the undersampling signal is preprocessed;

[0105] Based on the structural characteristics of the fuel control valve, the internal acoustic field of the fuel control valve is divided into grid cells. The under-sampled signal is sparsely represented. The sparse coefficient of each grid cell is iteratively updated through a threshold iteration algorithm to reconstruct the leakage acoustic field distribution of the fuel control valve from the under-sampled signal.

[0106] The acoustic energy density of each grid unit of the leakage sound field distribution is calculated, the energy abnormal grids are identified, the energy abnormal grids are clustered, the adjacent energy abnormal grids are merged, and the micro-leak point is determined in combination with the leakage amount classification, and the centroid coordinates of the micro-leak point are calculated as the micro-leak position.

[0107] When locating the micro-leakage position in the coarse positioning area, if the full sampling rate is used to obtain the signal, a large amount of data will be generated, which will increase the burden and time cost of data processing. The leakage ultrasonic signal has a certain sparsity. According to the compressed sensing theory, the signal is obtained at a sampling rate far lower than the Nyquist rate, and the signal is restored through the subsequent reconstruction algorithm. Therefore, using a 30% undersampling rate can reduce the amount of data and improve processing efficiency while ensuring signal information. Preprocessing the undersampling signal can remove noise and interference, improve the quality of the signal, and provide a better basis for subsequent sparse representation and reconstruction. In the coarse positioning area of ​​the leakage sound source, the ultrasonic sensor array is controlled to undersample by 30%. The sampling rate is used to obtain the leakage ultrasonic signal, and undersampling is achieved by setting the sampling interval or adopting random sampling. The undersampling signal is preprocessed, including denoising and filtering. Denoising is performed through the wavelet denoising method. Its basic principle is to decompose the leakage ultrasonic signal into different wavelet scales, remove the noise component through threshold processing, and then reconstruct it. The filtering is performed through a bandpass filter. According to the frequency range of the leakage ultrasonic signal, the appropriate passband frequency is set to remove the interference signals of other frequencies; thereby reducing the amount of data, reducing the burden and time cost of data processing, and improving the quality of the signal after preprocessing, which is conducive to subsequent sparse representation and reconstruction, and improving the accuracy of micro-leakage location.

[0108] The internal sound field of the fuel control valve is a complex three-dimensional space. Dividing it into grid units can discretize the continuous sound field, which is convenient for mathematical modeling and calculation. The leakage ultrasonic signal is sparse in space, that is, there are leakage signals in only a few grid units. Therefore, the under-sampled signal is sparsely represented. The sparse coefficient of each grid unit is continuously updated through the threshold iteration algorithm. The complete leakage sound field distribution can be reconstructed from the under-sampled signal, thereby determining the spatial distribution of the leakage signal; according to the structural characteristics of the fuel control valve, its internal sound field is divided into three-dimensional grid units of uniform size. The size of the grid unit is adjusted according to the requirements of positioning accuracy. Generally speaking, the smaller the grid unit, the higher the positioning accuracy, but the amount of calculation will also increase accordingly. Select a suitable sparse basis, such as through wavelet The undersampled signal is sparsely represented by the basis. The goal of sparse representation is to find a set of sparse coefficients so that the leakage ultrasonic signal can be represented as a linear combination of the sparse basis. Through the threshold iteration algorithm, such as the orthogonal matching pursuit algorithm, the basic idea of ​​the algorithm is to select the atom with the largest inner product with the residual, that is, a vector in the sparse basis, update the sparse coefficient, and update the residual in each iteration until the stopping condition is met, that is, the residual energy ratio is less than 5%. By continuously iteratively updating the sparse coefficient of each grid unit, the leakage sound field distribution of the fuel control valve is reconstructed from the undersampled signal; thereby discretizing the complex internal sound field to facilitate mathematical modeling and calculation. The use of sparse representation and reconstruction algorithm can accurately restore the leakage sound field distribution in the case of undersampling, thereby improving the accuracy of micro-leak location positioning.

[0109] The spatial distribution of leakage signals will result in different acoustic energy densities in different grid cells. The acoustic energy density of the grid cell where the leakage point is located will be significantly higher than that of the surrounding grid cells. Therefore, by calculating the acoustic energy density of each grid cell, the energy abnormal grids can be identified. These grid cells are the locations of the leakage points. Clustering and merging the energy abnormal grids can merge adjacent leakage points into one area, and more accurately determine the micro-leak points. Combined with the leakage amount classification, the severity of the leakage can be further judged, providing a reference for subsequent maintenance and treatment. Calculating the centroid coordinates of the micro-leak point can obtain a specific location, which is convenient for accurately locating the micro-leak position. According to the reconstructed leakage sound field distribution, the acoustic energy density of each grid cell is calculated. The calculation formula of the acoustic energy density is: ,in represents the density of the medium, represents the propagation speed of the leaked ultrasonic signal, The amplitude of the leak ultrasonic signal is represented by an acoustic energy threshold. Grid cells with acoustic energy density greater than the threshold are identified as energy anomaly grids. The energy anomaly grids are clustered using the K-Means clustering algorithm, and adjacent energy anomaly grids are merged into a cluster. Each cluster represents a leak point. The leakage level of each leak point is determined, and micro-leak points are extracted. The centroid coordinates of each micro-leak point are calculated by multiplying the coordinates of each grid cell corresponding to the micro-leak point and the acoustic energy density by the sum of the acoustic energy densities of all grid cells. The centroid coordinates of each micro-leak point are accurately identified through acoustic energy density calculation and cluster analysis. Combined with the leakage level classification, more comprehensive leakage information is provided. The centroid coordinates of the micro-leak points are calculated to obtain the specific micro-leak location, providing precise positioning for subsequent repair and treatment. The determined micro-leak location can provide maintenance personnel with clear maintenance targets and maintenance priorities. Maintenance personnel can take appropriate maintenance measures based on this information to improve maintenance efficiency and quality. This information is also used for performance evaluation and fault diagnosis of fuel control valves, providing a basis for equipment optimization and improvement.

[0110] S3. Monitor the fuel temperature and engine speed in real time, determine the bias compensation term based on the fuel temperature and engine speed, concatenate the bias compensation term with the leakage feature into a joint vector, and input the joint vector into a deep belief network. When the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter is greater than the deviation threshold, cross-validation of the ultrasonic sensor array is triggered.

[0111] Specifically, the determination logic of the bias compensation term includes:

[0112] Real-time monitoring of fuel temperature and engine speed;

[0113] The difference between the fuel temperature and the standard temperature is weighted and summed with the engine speed to obtain the bias compensation term;

[0114] The bias compensation term and the leakage feature are concatenated into a joint vector.

[0115] When aircraft engines operate under different operating conditions, fuel temperature and engine speed vary significantly. These changes can affect the leakage characteristics of the fuel control valve. For example, increased fuel temperature reduces fuel viscosity, leading to increased leakage. Changes in engine speed can cause fluctuations in fuel pressure and flow, which in turn affect the characteristics of the leakage signal. Therefore, real-time monitoring of these two parameters is necessary to accurately capture the impact of operating condition changes on leak detection. For fuel temperature monitoring, a temperature sensor is installed in the fuel pipeline near the control valve to ensure accurate measurement of the actual fuel temperature. The sensor's sampling frequency is set to 10Hz to track temperature changes in real time. For engine speed monitoring, a magnetoelectric speed sensor is used, installed near the engine's crankshaft or flywheel. This sensor measures speed by sensing changes in the magnetic field. The sampling frequency is also set to 10Hz to ensure timely response to speed changes. To ensure data accuracy and reliability, the sensor data is filtered. By accurately monitoring fuel temperature and engine speed in real time, changes in operating conditions can be captured promptly, providing a reliable data foundation for the subsequent calculation of bias compensation terms, thereby improving the accuracy and stability of leak detection.

[0116] The difference between the fuel temperature and the standard temperature and the change in engine speed will have different degrees of impact on the leakage characteristics. Through weighted summation, the influence of these two factors can be comprehensively considered to obtain a bias compensation term that can reflect the impact of operating condition changes on the leakage characteristics. The weight setting is adjusted according to actual experimental data and experience to ensure that the bias compensation term can accurately compensate for the impact of operating condition changes. First, the standard temperature is determined according to the design operating conditions of the engine and the characteristics of the fuel. For example, the average fuel temperature of the aircraft engine during normal operation is taken, and the difference between the fuel temperature and the standard temperature is calculated. In order to determine the weight, a large number of experimental studies are carried out. Under different fuel temperatures and engine speed conditions, data on leakage characteristics and actual leakage amounts are obtained. Through multivariate linear regression analysis, the weight coefficients of the difference between the fuel temperature and the standard temperature and the influence of the engine speed on the leakage characteristics are obtained, and the bias compensation term is determined by weighted summation. This can more accurately reflect the impact of operating condition changes on the leakage characteristics and effectively compensate for leakage detection errors caused by operating condition changes.

[0117] When predicting leakage, the deep belief network needs to comprehensively consider the influence of leakage characteristics and operating condition changes. Concatenating the bias compensation term and the leakage characteristics into a joint vector can integrate the operating condition information into the leakage characteristics, enabling the deep belief network to learn the mapping relationship between leakage characteristics and leakage under different operating conditions, thereby improving the accuracy of the prediction. In practical applications, in order to ensure the dimensional consistency of the joint vector, the bias compensation term needs to be normalized so that it has the same order of magnitude as the leakage characteristic. The bias compensation term is normalized to the interval [0,1] through the Min-Max normalization method. Concatenating the bias compensation term and the leakage characteristics into a joint vector can make full use of the operating condition information, improve the deep belief network's ability to predict leakage under different operating conditions, and reduce misjudgments caused by operating condition changes. The joint vector, as the input of the deep belief network, directly affects the output result of the deep belief network. An accurate joint vector can enable the deep belief network to predict leakage more accurately and provide a reliable basis for subsequent cross-validation.

[0118] Specifically, if Figure 4 As shown, the cross-validation logic of the ultrasound sensor array includes:

[0119] Calculate in real time the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter;

[0120] Configure a deviation threshold. When the deviation is greater than the deviation threshold, cross-validation of the ultrasonic sensor array is triggered.

[0121] At the same time, the cross-validation of the ultrasonic sensor array is determined based on the changes in fuel temperature and engine speed;

[0122] When the cross-validation of the ultrasonic sensor array is triggered, the spare sensing channel of the ultrasonic sensor array is activated to reacquire the leakage ultrasonic signal to extract a new leakage feature;

[0123] The similarity between the new leakage feature and the joint vector is calculated by Euclidean distance, and the similarity is compared with the similarity threshold to determine whether to retrain the deep belief network.

[0124] During operation, the deep belief network will be affected by various factors, such as sensor noise and changes in operating conditions, which may cause a deviation between the leakage output and the actual leakage. Real-time calculation of the deviation between the two can timely discover the accuracy of the network prediction and provide a basis for subsequent cross-validation. The output of the deep belief network is the predicted leakage, and the fuel flow meter reading is the actual leakage. The deviation is the absolute value of the difference between the leakage output of the deep belief network and the actual leakage read by the fuel flow meter. In order to smooth the deviation data and reduce the influence of noise, a sliding window is set with a window size of N, such as N=10, through the moving average filtering method. The values ​​of N consecutive deviations are averaged to obtain the smoothed deviation. Real-time calculation of the deviation and smoothing processing can timely and accurately reflect the prediction accuracy of the deep belief network and provide a reliable indicator for judging whether cross-validation is needed.

[0125] In order to ensure the accuracy and reliability of leak detection, a reasonable deviation threshold is set. When the deviation of the deep belief network output is greater than the deviation threshold, it means that there is a large error in the network's prediction result, and cross-validation of the ultrasonic sensor array is needed to further confirm the leakage situation. The determination of the deviation threshold must comprehensively consider multiple factors, such as the measurement error of the sensor, the training error of the deep belief network, and the allowable error in actual application. Through a large amount of experimental data, the deviation distribution of the deep belief network output under different working conditions is analyzed. Through statistical methods, such as calculating the mean and standard deviation of the deviation, the deviation threshold is determined according to a certain confidence level (such as 95%). When the real-time monitoring deviation is greater than the deviation threshold, the cross-validation mechanism of the ultrasonic sensor array is triggered. Reasonable configuration of the deviation threshold can effectively judge the prediction accuracy of the deep belief network, trigger the cross-validation mechanism in time, and improve the reliability and accuracy of leak detection.

[0126] Drastic changes in fuel temperature and engine speed can cause significant changes in leakage characteristics, thereby affecting the prediction accuracy of the deep belief network. Therefore, in addition to considering the deviation threshold, it is also necessary to determine whether cross-validation is needed based on the changes in fuel temperature and engine speed to further improve the reliability of leak detection. By calculating the rate of change of fuel temperature and the rate of change of engine speed, that is, the ratio of the difference in fuel temperature (engine speed) within a fixed period to the fixed period, and then setting the fuel temperature change rate threshold and the engine speed change rate threshold, these change rate thresholds are determined based on the engine design parameters and actual operating experience. When the fuel temperature change rate is greater than the fuel temperature change rate threshold or the engine speed change rate is greater than the engine speed change rate threshold, the cross-validation mechanism of the ultrasonic sensor array is also triggered. Comprehensively considering the changes in fuel temperature and engine speed can more comprehensively judge the prediction accuracy of the deep belief network, timely discover prediction errors caused by drastic changes in operating conditions, and improve the reliability of leak detection.

[0127] When the deviation predicted by the deep confidence network is greater than the deviation threshold or the operating conditions change drastically, it is because the main sensing channel is interfered with or the leakage characteristics have changed significantly. Activating the backup sensing channel can obtain an independent leakage ultrasonic signal. By re-extracting the leakage characteristics and comparing them with the previous results, the prediction accuracy of the deep confidence network is verified. The backup sensing channel of the ultrasonic sensor array is in standby state under normal circumstances. When the cross-validation mechanism is triggered, the backup sensing channel is activated. The sensor of the backup sensing channel has the same performance and parameters as the sensor of the main sensing channel to ensure that the acquired leakage ultrasonic signal is comparable. After re-acquiring the leakage ultrasonic signal, new leakage features are extracted according to the above-mentioned leakage signal separation and leakage feature iterative extraction logic. Activating the backup sensing channel can obtain an independent leakage ultrasonic signal. By re-extracting the leakage characteristics, the prediction accuracy of the deep confidence network can be effectively verified, and misjudgment caused by sensor failure or interference can be reduced.

[0128] By comparing the similarity between the new leakage feature and the joint vector, it is possible to determine whether the leakage feature has changed significantly. If the similarity is low, it means that the leakage feature has changed significantly, the prediction ability of the deep belief network will be affected, and the network needs to be retrained to adapt to the new leakage feature; the similarity between the new leakage feature and the leakage feature in the joint vector is calculated by Euclidean distance, and a similarity threshold is set. When the similarity is less than the similarity threshold, it is determined that the deep belief network needs to be retrained. The similarity threshold is determined by experimental data and experience to balance the cost of retraining and the accuracy of network prediction; by calculating the similarity by Euclidean distance and comparing it with the similarity threshold, it is possible to objectively determine whether the leakage feature has changed significantly and decide whether the deep belief network needs to be retrained, thereby ensuring the prediction accuracy of the deep belief network under different working conditions. If it is determined that the deep belief network needs to be retrained, the leakage ultrasonic signal needs to be reacquired, and the deep belief network needs to be trained and optimized to improve the adaptability of the deep belief network to the new leakage feature. If retraining is not required, the current deep belief network continues to be used for leakage prediction.

Claims

1. Ultrasonic detection method for leakage of fuel control valve of aircraft engine, characterized in that: include: An ultrasonic sensor array is deployed to acquire the ultrasonic leakage signal of the fuel control valve. The background noise of the leakage ultrasonic signal is separated by wavelet packet transform. A threshold is dynamically set based on the energy probability of the background noise to isolate the leakage signal. At the same time, a matching pursuit algorithm is used to iteratively extract leakage characteristics, including sudden changes in acoustic emission energy, resonant frequency offset, and harmonic distortion rate. A mapping relationship between leak characteristics and leakage volume is established based on a deep belief network, and a leakage volume classification is determined. A cross-correlation algorithm is used to determine the coarse localization area of ​​the leak sound source based on the time difference of the ultrasonic sensor array. Within this coarse localization area of ​​the leak sound source, a compressed sensing algorithm is used to reconstruct the leakage sound field distribution inside the fuel control valve from the leakage ultrasonic signal with a 30% undersampling rate to locate the micro-leak. The leakage volume classification is then combined with the coarse localization area of ​​the leak sound source and the micro-leak location to generate a leak heat map. Fuel temperature and engine speed are monitored in real time. A bias compensation term is determined based on the fuel temperature and engine speed. The bias compensation term and the leakage feature are concatenated into a joint vector, which is then input into a deep belief network. When the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter is greater than a deviation threshold, cross-validation of the ultrasonic sensor array is triggered.

2. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 1, wherein: The iterative extraction logic of the leakage feature includes: Based on the structural characteristics of the fuel control valve, an initial composite atom library is constructed; Initialize the residual For leakage signal, search the atom with the largest inner product with leakage signal from the initial composite atom library , calculate the residual ; The residual Match with the atoms in the initial composite atom library and select the atoms with the residual The atom with the largest inner product , update the composite atom library; When the residual energy ratio is less than 5% or the number of iterations is met, the iteration is stopped, the composite atom library is output, and leakage features are extracted from the composite atom library; The acoustic emission energy mutation represents the difference in atomic energy between adjacent iterations, the resonant frequency offset represents the change between the atomic resonant frequency and the normal resonant frequency, and the harmonic distortion rate represents the ratio of the atomic harmonic energy to the fundamental wave energy.

3. The ultrasonic detection method for leakage of an aircraft engine fuel control valve according to claim 2, characterized in that: The separation logic of the leakage signal includes: The acquired leakage ultrasonic signal is decomposed into 3-layer wavelet packets to generate 8 sub-bands; Determine the energy entropy of each sub-band, determine the noise frequency band based on the energy entropy, and separate the background noise of the leaked ultrasonic signal; The energy probability of the background noise is determined according to the energy mean and energy standard deviation of the noise frequency band, so as to dynamically set the energy threshold, and the leakage ultrasound signal with energy greater than the energy threshold in each sub-band is regarded as a candidate leakage signal; Performing short-time Fourier transform on the candidate leakage signals, screening the candidate leakage signals to separate the leakage signals.

4. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 3, wherein: The logic for establishing the mapping relationship between the leakage characteristics and the leakage amount includes: Collecting leakage characteristics and actual leakage amount obtained by the fuel flow meter, normalizing the leakage characteristics and actual leakage amount to obtain a standard data set; Construct a deep belief network with multiple layers of RBM stacked together, initialize the weights and biases of each RBM layer, and pre-train the RBM layer by layer through unsupervised learning; The deep belief network is fine-tuned in a supervised manner through the back-propagation algorithm, with the mean square error as the loss function, and the weights and biases of the deep belief network are updated through stochastic gradient descent; After multiple iterative training, when the loss function converges to the loss threshold, the mapping relationship between leakage features and leakage amount in the deep belief network is output.

5. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 4, wherein: The sub-logic for determining the coarse positioning area of ​​the leakage sound source includes: The time difference of the leakage ultrasonic signals between each pair of sensors is calculated by a cross-correlation algorithm based on the timestamp of the leakage ultrasonic signal received by each sensor in the ultrasonic sensor array; The distance between each pair of sensors is determined based on the time difference, and the hyperbola of each pair of sensors is obtained. By combining different sensor pairs, the hyperbola equation is obtained. Based on the structural characteristics of the fuel control valve and the physical characteristics of the leakage, the solutions of the hyperbolic equation are screened, and adjacent solutions with similar characteristics are merged into the coarse localization area of ​​the leakage sound source through cluster analysis. The time difference is randomly perturbed multiple times by Monte Carlo simulation method, and the solution of the hyperbola equation is re-screened to determine the confidence interval of the rough positioning area of ​​the leakage sound source.

6. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 5, wherein: The micro-leakage location positioning sub-logic includes: In the coarse location area of ​​the leakage sound source, the ultrasonic sensor array acquires the leakage ultrasonic signal at a 30% undersampling rate to obtain the undersampling signal, and the undersampling signal is preprocessed; Based on the structural characteristics of the fuel control valve, the internal acoustic field of the fuel control valve is divided into grid cells. The under-sampled signal is sparsely represented. The sparse coefficient of each grid cell is iteratively updated through a threshold iteration algorithm to reconstruct the leakage acoustic field distribution of the fuel control valve from the under-sampled signal. The acoustic energy density of each grid unit of the leakage sound field distribution is calculated, the energy abnormal grids are identified, the energy abnormal grids are clustered, the adjacent energy abnormal grids are merged, and the micro-leak point is determined in combination with the leakage amount classification, and the centroid coordinates of the micro-leak point are calculated as the micro-leak position.

7. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 6, wherein: The determination logic of the bias compensation term includes: Real-time monitoring of fuel temperature and engine speed; The difference between the fuel temperature and the standard temperature is weighted and summed with the engine speed to obtain the bias compensation term; The bias compensation term and the leakage feature are concatenated into a joint vector.

8. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 7, wherein: The cross-validation logic of the ultrasound sensor array includes: Calculate in real time the deviation between the leakage amount output by the deep belief network and the actual leakage amount read by the fuel flow meter; Configure a deviation threshold. When the deviation is greater than the deviation threshold, cross-validation of the ultrasonic sensor array is triggered. At the same time, the cross-validation of the ultrasonic sensor array is determined based on the changes in fuel temperature and engine speed; When the cross-validation of the ultrasonic sensor array is triggered, the spare sensing channel of the ultrasonic sensor array is activated to reacquire the leakage ultrasonic signal to extract a new leakage feature; The similarity between the new leakage feature and the joint vector is calculated by Euclidean distance, and the similarity is compared with the similarity threshold to determine whether to retrain the deep belief network.

9. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 8, wherein: The generation logic of the leakage heat map includes: Integrate the leakage level classification with the coarse localization area of ​​the leakage sound source and the micro-leak position into the data structure, and assign quantitative values ​​of the leakage level classification to the coarse localization area of ​​the leakage sound source and the micro-leak position; The three-dimensional coordinates of the fuel control valve are converted into two-dimensional coordinates through orthographic projection. The two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position are calibrated and adjusted according to the arrangement of the ultrasonic sensor array. The quantitative values ​​of the leakage level are mapped to different colors, and the two-dimensional coordinates of the coarse positioning area of ​​the leakage sound source and the micro-leak position and the mapped colors are integrated to generate a leakage heat map.

10. The method for ultrasonically detecting leakage of an aircraft engine fuel control valve according to claim 9, wherein: An ultrasonic sensor array is arranged in a non-contact manner on the outer wall of the metal pipe section downstream of the fuel control valve. The ultrasonic sensor array includes at least two sensor groups, and the sensor groups include a longitudinal wave sensor group and a shear wave sensor group. The longitudinal wave sensor group is arranged obliquely at a first critical angle of incidence to obtain a longitudinal leakage ultrasonic signal of the fuel control valve. The shear wave sensor group is arranged obliquely at a second critical angle of incidence to obtain a transverse leakage ultrasonic signal of the fuel control valve.

Images