Praseodymium-neodymium alloy nondestructive testing method and system based on acoustic characteristic analysis

Abstract

The application relates to the field of alloy defect detection, and specifically discloses a praseodymium-neodymium alloy nondestructive detection method and system based on acoustic feature analysis, which comprehensively captures defect information contained in an original probe signal from two complementary physical perspectives of instantaneous dynamic characteristics and frequency band energy distribution by simultaneously adopting Hilbert-Huang transform and wavelet packet transform. Further, the scheme discards simple feature splicing, and instead utilizes canonical correlation analysis as an information decoupling tool to online decompose two groups of original feature vectors into a shared part describing defect commonality and unique information parts respectively representing the unique resolution capabilities of HHT and wavelet packet. Finally, the three decoupled components are structurally recombined to form a fusion feature vector which can effectively eliminate redundancy, amplify differences and has higher information density, thereby providing a clear structure and highly refined input for a subsequent classification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of alloy defect detection, and more particularly, to a praseodymium-neodymium alloy nondestructive testing method and system based on acoustic feature analysis. BACKGROUND

[0002] Praseodymium-neodymium alloy is the core raw material for preparing high-performance Nd-Fe-B permanent magnet materials, and its internal quality directly determines the magnetic properties, mechanical properties and service stability of the final magnet product. In the production process of praseodymium-neodymium alloy, such as smelting and casting, various internal defects such as cracks and pores will inevitably occur. These microscopic defects will cause discontinuity of the material internal magnetic domain structure and stress concentration, which seriously weakens the performance and reliability of the product. Therefore, accurate and efficient nondestructive testing of praseodymium-neodymium alloy, especially accurate identification and differentiation of different types of defects, is of great significance to ensure product quality, optimize production process and reduce manufacturing cost.

[0003] Acoustic nondestructive testing technology, especially ultrasonic testing, has become the mainstream method for detecting internal defects of metal materials due to its strong penetration, high sensitivity and harmlessness to operators. However, direct application of traditional acoustic testing methods to praseodymium-neodymium alloy faces severe challenges. On the one hand, the polycrystalline heterogeneous structure of praseodymium-neodymium alloy will cause strong scattering and attenuation of ultrasonic waves, resulting in that the defect echo signal is often submerged in strong grain noise, and the signal-to-noise ratio is extremely low. On the other hand, different types of defects (such as cracks and pores) have significant nonlinear and non-stationary characteristics in acoustic response, and the characteristic differences of their echo signals in time domain waveform or traditional Fourier spectrum are very weak, often overlapping, so that the traditional detection method relying on only a single or a few characteristic parameters such as echo amplitude and pulse width cannot effectively distinguish the defect types, resulting in high misjudgment rate and missed detection rate.

[0004] Therefore, an optimized praseodymium-neodymium alloy nondestructive testing scheme based on acoustic feature analysis is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. The embodiments of the present application provide a praseodymium-neodymium alloy nondestructive testing method and system based on acoustic feature analysis.

[0006] According to one aspect of the present application, a praseodymium-neodymium alloy nondestructive testing method based on acoustic feature analysis is provided, which comprises:

[0007] Adaptive band-pass filtering the collected original probe signal to obtain a filtered signal;

[0008] Parallel feature extraction is performed on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector;

[0009] The HHT instantaneous feature vector and the wavelet packet energy feature vector are fused to obtain a fusion feature vector.

[0010] The fusion feature vector is input into a pre-trained SVM model to obtain a defect category label.

[0011] In the above-mentioned acoustic feature analysis-based praseodymium-neodymium alloy nondestructive detection method, the collected original probe signal is subjected to adaptive band-pass filtering to obtain a filtered signal, including: the original probe signal is subjected to defect echo gating processing to obtain a gated signal; the gated signal is subjected to center frequency estimation based on fast Fourier transform to obtain a peak frequency of an energy spectrum; based on the peak frequency of the energy spectrum, the gated signal is subjected to adaptive band-pass filtering to obtain the filtered signal, wherein the center frequency of the adaptive band-pass filtering is the peak frequency of the energy spectrum.

[0012] In the above-mentioned acoustic feature analysis-based praseodymium-neodymium alloy nondestructive detection method, the filtered signal is subjected to parallel feature extraction to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector, including: the filtered signal is subjected to Hilbert-Huang transform instantaneous feature extraction to obtain the HHT instantaneous feature vector; the filtered signal is subjected to wavelet packet transform energy feature extraction to obtain the wavelet packet energy feature vector.

[0013] In the above-mentioned acoustic feature analysis-based praseodymium-neodymium alloy nondestructive detection method, the HHT instantaneous feature vector and the wavelet packet energy feature vector are fused to obtain a fusion feature vector, including: the HHT instantaneous feature vector and the wavelet packet energy feature vector are subjected to feature preprocessing to obtain a shared feature vector, an HHT unique feature vector and a wavelet packet unique feature vector; the shared feature vector, the HHT unique feature vector and the wavelet packet unique feature vector are subjected to structured feature assembly to obtain the fusion feature vector.

[0014] In the above-mentioned acoustic feature analysis-based praseodymium-neodymium alloy nondestructive detection method, the HHT instantaneous feature vector and the wavelet packet energy feature vector are subjected to feature preprocessing to obtain a shared feature vector, an HHT unique feature vector and a wavelet packet unique feature vector, including: obtaining a training set HHT feature matrix and a training set WPT feature matrix; performing canonical correlation analysis on the training set HHT feature matrix and the training set WPT feature matrix to obtain a canonical basis matrix; based on the canonical basis matrix, online feature decoupling is performed on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain the shared feature vector, the HHT unique feature vector and the wavelet packet unique feature vector.

[0015] In the Pr-Nd alloy nondestructive detection method based on acoustic feature analysis, the HHT instantaneous feature vector and the wavelet packet energy feature vector are decoupled based on a canonical basis matrix to obtain a shared feature vector, an HHT unique feature vector and a wavelet packet unique feature vector, including: projecting the HHT instantaneous feature vector and the wavelet packet energy feature vector to a feature space of the canonical basis matrix to obtain an HHT projected feature vector and a wavelet packet projected feature vector; calculating a mean vector of the HHT projected feature vector and the wavelet packet projected feature vector as the shared feature vector; calculating a residual vector between the HHT instantaneous feature vector and the HHT projected feature vector as the HHT unique feature vector; and calculating a residual vector between the wavelet packet energy feature vector and the wavelet packet projected feature vector as the wavelet packet unique feature vector.

[0016] In the Pr-Nd alloy nondestructive detection method based on acoustic feature analysis, the defect category label includes a crack label, a pore label and a defect-free label.

[0017] According to another aspect of the present application, a Pr-Nd alloy nondestructive detection system based on acoustic feature analysis is provided, which includes:

[0018] A band-pass filtering module is configured to perform adaptive band-pass filtering on the collected original probe signal to obtain a filtered signal.

[0019] A parallel feature extraction module is configured to perform parallel feature extraction on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector.

[0020] A feature fusion module is configured to perform feature fusion on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a fusion feature vector.

[0021] A defect detection module is configured to input the fusion feature vector into a pre-trained SVM model to obtain a defect category label.

[0022] Compared with the prior art, the praseodymium-neodymium alloy nondestructive detection method and system based on acoustic feature analysis provided by the application comprehensively capture the defect information contained in the original probe signal from two complementary physical perspectives of instantaneous dynamic characteristics and frequency band energy distribution by parallel use of Hilbert-Huang transform and wavelet packet transform. Further, the scheme discards simple feature splicing and instead uses canonical correlation analysis as an information decoupling tool to online decompose the two groups of original feature vectors into a shared part describing the commonality of defects and unique information parts respectively representing the unique resolution capabilities of HHT and wavelet packet. Finally, the three decoupled components are structurally recombined to form a fusion feature vector that can effectively eliminate redundancy, amplify differences, and has higher information density. The subsequent classification model is provided with a clear structure and highly refined input, and the fusion feature is then used for defect category recognition, thereby fundamentally improving the recognition accuracy and system stability of similar defects such as cracks and pores. BRIEF DESCRIPTION OF DRAWINGS

[0023] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which:

[0024] Figure 1 A flowchart of the praseodymium-neodymium alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the application;

[0025] Figure 2 A data flow diagram of the praseodymium-neodymium alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the application;

[0026] Figure 3 A flowchart of parallel feature extraction on filtered signals to obtain HHT instantaneous feature vectors and wavelet packet energy feature vectors in the praseodymium-neodymium alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the application;

[0027] Figure 4 A flowchart of feature fusion on HHT instantaneous feature vectors and wavelet packet energy feature vectors to obtain a fusion feature vector in the praseodymium-neodymium alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the application;

[0028] Figure 5 A block diagram of the praseodymium-neodymium alloy nondestructive detection system based on acoustic feature analysis according to the embodiments of the application. DETAILED DESCRIPTION

[0029] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of embodiments of the present application, and thus should not be used to limit the present application, and it should be understood that the present application is not limited by the example embodiments described herein.

[0030] As used in this application and in the claims, the terms "including", "containing", "having", "including", "comprising", "characterized by" and the like are not intended to be limiting, and are understood to mean "comprising." As used in this application, the terms "coupled" and "coupling" mean an indirect, direct, optical, and / or electrical connection. Thus, the

[0031] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0032] Flow diagrams are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in the exact order as shown. Rather, various steps can be handled in reverse order, or at the same time, as desired. Other operations can also be added to, or removed from, these processes, or one or more steps can be removed from these processes.

[0033] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of embodiments of the present application, and thus should not be used to limit the present application, and it should be understood that the present application is not limited by the example embodiments described herein.

[0034] In view of the technical problems that in the prior art, when acoustic signals of Pr-Nd alloy nondestructive detection are processed, defect characteristics such as cracks and pores are mixed and overlapped, and the traditional feature fusion method will introduce information redundancy and bury key classification basis, in the technical scheme of the present application, a Pr-Nd alloy nondestructive detection method based on acoustic feature analysis is proposed. Specifically, first, the collected original probe signal is adaptively band-pass filtered to suppress noise and enhance effective echo; then, the filtered signal is subjected to Hilbert-Huang transform (HHT) and wavelet packet transform in parallel, from the two complementary perspectives of instantaneous dynamic characteristics and frequency band energy distribution, HHT instantaneous feature vectors and wavelet packet energy feature vectors are extracted. Further, the present scheme discards the traditional method of directly splicing features, and instead introduces a feature decoupling mechanism based on canonical correlation analysis. The mechanism uses the canonical basis matrix learned in advance on the training set to project and calculate the residual of the two groups of feature vectors obtained online, thereby accurately decomposing them into three parts: a shared feature vector that condenses the common information of the two features (such as defect macro energy), and two unique feature vectors that capture the unique transient information of HHT (which may correspond to the nonlinear response of crack tip) and the unique frequency band information of wavelet packet (which may correspond to the group resonance of pores). Subsequently, the three decoupled and semantically clear feature components are structured and assembled to form a final fusion feature vector with information redundancy eliminated and key differences significantly amplified. Finally, the high-quality fusion feature vector is input into a pre-trained support vector machine (SVM) model to obtain an accurate defect class label. In this way, the present application effectively solves the problems of feature redundancy and confusion, enabling the classifier to focus on distinguishing the essential differences between defects, thereby significantly improving the accuracy and reliability of the detection.

[0035] In the technical scheme of the present application, a Pr-Nd alloy nondestructive detection method based on acoustic feature analysis is proposed. Figure 1 A flowchart of the Pr-Nd alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the present application. Figure 2 A data flow diagram of the Pr-Nd alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the present application. As shown in Figure 1 and Figure 2 The Pr-Nd alloy nondestructive detection method based on acoustic feature analysis according to the embodiments of the present application includes the steps of: S100, adaptively band-pass filtering the collected original probe signal to obtain a filtered signal; S200, performing parallel feature extraction on the filtered signal to obtain HHT instantaneous feature vectors and wavelet packet energy feature vectors; S300, performing feature fusion on the HHT instantaneous feature vectors and wavelet packet energy feature vectors to obtain a fusion feature vector; S400, inputting the fusion feature vector into a pre-trained SVM model to obtain a defect class label.

[0036] Specifically, in step S100, the collected original probe signal is adaptively band-pass filtered to obtain a filtered signal. It should be understood that, in the process of acoustic detection of praseodymium-neodymium alloy, the collected original probe signal not only contains defect-irrelevant signal components such as initial wave and bottom wave, but also contains a large amount of high-frequency and low-frequency noise originating from material grain scattering and electronic system; meanwhile, the attenuation effect of the sound wave in the material will cause the actual center frequency of the defect echo to deviate from the nominal frequency of the probe. Therefore, in the technical solution of the present application, after the original probe signal is collected, the collected original probe signal is further adaptively band-pass filtered to obtain a filtered signal, so as to accurately isolate the defect echo signal and dynamically adjust the filtering parameters according to the frequency spectrum characteristics of the echo itself, thereby filtering out the noise interference in irrelevant frequency bands to the greatest extent. In this way, the signal-to-noise ratio of the defect echo signal can be significantly improved, laying a high-quality data foundation for subsequent accurate HHT instantaneous feature and wavelet packet energy feature extraction, and avoiding the interference of noise components on feature analysis.

[0037] More specifically, in the embodiment of the present application, adaptively band-pass filtering the collected original probe signal to obtain a filtered signal comprises: performing defect echo gating processing on the original probe signal to obtain a gated signal; performing center frequency estimation based on fast Fourier transform on the gated signal to obtain a peak frequency of an energy spectrum; and adaptively band-pass filtering the gated signal based on the peak frequency of the energy spectrum to obtain the filtered signal, wherein the center frequency of the adaptive band-pass filtering is the peak frequency of the energy spectrum.

[0038] Correspondingly, the original probe signal is subjected to defect echo gating processing to obtain a gated signal. It should be understood that, since the collected original probe signal is a complete time sequence, it not only contains defect information reflecting defect echo, but also contains a strong initial pulse, i.e., initial wave, when the probe is emitted, and a bottom echo, i.e., bottom wave, formed by the reflection of the sound wave to the workpiece bottom surface. Therefore, in the technical solution of the present application, the original probe signal is further subjected to defect echo gating processing to obtain a gated signal, so as to accurately focus the analysis range on the effective signal segment containing defect information and exclude the interference of irrelevant signal components such as initial wave and bottom wave. In this way, it can be ensured that all analysis steps such as center frequency estimation and feature extraction are performed on pure defect echo signals, thereby avoiding the false guidance of strong-energy initial wave or bottom wave to defect spectrum characteristics and ensuring the accuracy and effectiveness of subsequent processing.

[0039] Specifically, in one specific example of the present application, the defect echo gating process is based on the principle of sound path to set a time window to intercept the valid signal. First, according to the known thickness of praseodymium-neodymium alloy workpiece, material sound speed and the incident position of the ultrasonic probe, the effective time range of sound wave propagation inside the workpiece is calculated. Specifically, the starting time point of the gate is set to avoid the initial wave signal area emitted by the probe, and the termination time point of the gate is set before the bottom surface echo arrives. Then, this time gate with a determined starting and ending time point is applied to the collected original probe signal. Finally, only the signal data segment falling within the time gate is extracted and retained, while all signal data outside the gate is discarded, thereby generating a gated signal containing only defect echo information for subsequent spectral analysis and filtering processing.

[0040] Correspondingly, the center frequency estimation of the gated signal based on fast Fourier transform is performed to obtain the peak frequency of the energy spectrum. It should be understood that, due to the frequency-dependent attenuation of ultrasonic waves in the praseodymium-neodymium alloy, high-frequency components attenuate faster than low-frequency components, and different types and depths of defects also have different scattering responses to sound waves, resulting in that the center frequency of the actual received defect echo signal deviates from the nominal frequency of the probe. Therefore, in the technical solution of the present application, the center frequency estimation of the gated signal based on fast Fourier transform is further performed to obtain the peak frequency of the energy spectrum, so as to accurately and data-drivenly determine the frequency point at which the energy of the specific defect echo signal is most concentrated. In this way, an accurate and adaptive center frequency parameter can be provided for the subsequent band-pass filtering step, ensuring that the filter can accurately match the spectral characteristics of the current signal, rather than relying on a fixed and possibly inaccurate preset frequency.

[0041] Specifically, in one specific example of the present application, the center frequency estimation process is to transform the gated signal in time domain to frequency domain for analysis. First, the input gated signal, i.e. a time domain waveform data representing the defect echo, is applied with the fast Fourier transform algorithm. The transformation converts the signal from time domain to frequency domain, generating a complex number array, where each element corresponds to the amplitude and phase of a specific frequency component. Next, to obtain energy information, the square of the modulus of each element in the complex number array is calculated, thereby obtaining the energy spectrum of the signal, which intuitively shows the distribution of signal energy at different frequencies. Finally, peak search is performed on the energy spectrum, i.e. all frequency points are traversed to find the point with the maximum energy value, and the frequency value corresponding to the maximum energy point is determined as the peak frequency of the energy spectrum, which is subsequently used as the center frequency of the adaptive band-pass filter.

[0042] Accordingly, the gating signal is adaptively band-pass filtered based on the peak frequency of the energy spectrum to obtain the filtered signal, wherein the center frequency of the adaptive band-pass filtering is the peak frequency of the energy spectrum. It can be understood that, in addition to the main energy component of the defect echo, random noise introduced by material grain scattering and electronic systems is still distributed in the entire frequency band in the gating signal, which will seriously interfere with the accuracy of subsequent feature extraction. Therefore, in the technical solution of the present application, the gating signal is further adaptively band-pass filtered based on the peak frequency of the energy spectrum to obtain the filtered signal, wherein the center frequency of the adaptive band-pass filtering is the peak frequency of the energy spectrum, so as to construct a filter whose passband accurately covers the core energy area of the current defect echo. In this way, the effective information component of the defect echo can be retained to the greatest extent, while the noise interference outside the frequency band is filtered out, thereby significantly improving the signal-to-noise ratio of the signal and providing pure data input for subsequent high-precision feature extraction.

[0043] Specifically, in one specific example of the present application, the adaptive band-pass filtering process is to dynamically design and apply a digital filter. First, the peak frequency of the energy spectrum calculated in the previous step is taken as the center frequency of the adaptive band-pass filter. At the same time, the bandwidth of the filter is set, for example, the upper and lower cutoff frequencies of the bandwidth are set to 0.5 times and 1.5 times of the center frequency to ensure complete coverage of the main energy area of the signal. Then, based on the dynamically determined center frequency and bandwidth parameters, a digital Butterworth band-pass filter of a specific order is designed and generated, and the coefficients of its transfer function are calculated. Finally, the designed filter is applied to the gating signal, and each data point in the gating signal is processed through convolution operation. The signal sequence output after the operation is the filtered signal, the energy of the signal is concentrated around the peak frequency, and the noise component outside the frequency band has been effectively suppressed.

[0044] Specifically, in step S200, parallel feature extraction is performed on the filtered signal to obtain the HHT instantaneous feature vector and the wavelet packet energy feature vector. It should be understood that, due to the essential difference in the physical mechanism of different types of defects (such as cracks and pores) when interacting with sound waves, these differences will be reflected in the echo signal in a complex and complementary form. Specifically, the diffraction and nonlinear effects of the crack tip manifest as a dramatic change in the instantaneous characteristics of the signal, while the geometric shape of the pore may cause energy resonance in a specific frequency band. Therefore, in the technical solution of the present application, further parallel feature extraction is performed on the filtered signal to obtain the HHT instantaneous feature vector and the wavelet packet energy feature vector, so as to comprehensively and deeply characterize the defect echo signal from the two mutually orthogonal dimensions of instantaneous dynamic characteristics and frequency band energy distribution. In this way, a multi-dimensional feature description with higher resolution than a single feature set can be constructed, providing a rich and complementary information basis for subsequent accurate differentiation of similar defect signals with different physical causes.

[0045] Figure 3 A flowchart of parallel feature extraction on the filtered signal to obtain the HHT instantaneous feature vector and the wavelet packet energy feature vector according to the praseodymium-neodymium alloy non-destructive testing method based on acoustic feature analysis of the embodiments of the present application. As shown in Figure 3 S200, it includes: S210, performing Hilbert-Huang transform instantaneous feature extraction on the filtered signal to obtain the HHT instantaneous feature vector; S220, performing wavelet packet transform energy feature extraction on the filtered signal to obtain the wavelet packet energy feature vector.

[0046] Correspondingly, in step S210, Hilbert-Huang transform instantaneous feature extraction is performed on the filtered signal to obtain the HHT instantaneous feature vector. It should be understood that, due to the sharp edges and complex geometry of crack-like defects inside praseodymium-neodymium alloys, diffraction, scattering and nonlinear effects will occur when they interact with sound waves, resulting in significant non-stationary and nonlinear characteristics of their echo signals, i.e. the frequency and amplitude of the signal will change dramatically in a very short time. Therefore, in the technical solution of the present application, further Hilbert-Huang transform instantaneous feature extraction is performed on the filtered signal to obtain the HHT instantaneous feature vector, so as to adaptively decompose the complex defect echo into a series of intrinsic vibration modes and accurately track the instantaneous trajectory of energy and frequency change over time in each mode. In this way, it can effectively capture and quantify the transient dynamic information that traditional Fourier analysis cannot reveal, which is closely related to the physical nature of the defect (such as the dynamic opening and closing of the crack), thereby generating a feature description that is highly sensitive and specific to this type of defect.

[0047] Specifically, in one specific example of the present application, the Hilbert-Huang Transform instantaneous feature extraction process first processes the input filtered signal by an Empirical Mode Decomposition algorithm. This algorithm adaptively decomposes the filtered signal into a finite number of Intrinsic Mode Function (IMF) components by iterative sifting, where each IMF component satisfies certain time-domain oscillation conditions. Then, the first few IMF components with the largest energy contribution are selected, and the Hilbert transform is applied to each selected IMF component to transform it into an analytic signal. Then, based on the analytic signal, the instantaneous amplitude sequence and the instantaneous frequency sequence are calculated. Finally, in order to form a fixed-dimension feature vector, the statistical feature parameters, such as mean, standard deviation, variance, kurtosis and skewness, are calculated for the instantaneous amplitude sequence and the instantaneous frequency sequence of each IMF component, respectively, and all the calculated statistical parameters are combined into a one-dimensional vector, which is the HHT instantaneous feature vector.

[0048] Correspondingly, in step S220, the wavelet packet transform energy feature extraction is performed on the filtered signal to obtain the wavelet packet energy feature vector. It should be understood that, due to the specific geometric structure and size distribution of the volume defects such as pores and loose in the praseodymium-neodymium alloy, they will be acoustically manifested as resonance absorption or scattering of acoustic waves in specific frequency bands, resulting in a unique distribution pattern of the energy of the defect echo signal in the frequency domain. Therefore, in the technical solution of the present application, the wavelet packet transform energy feature extraction is further performed on the filtered signal to obtain the wavelet packet energy feature vector, so as to finely decompose the signal in the full frequency band and accurately quantify the distribution of the energy of the signal in each orthogonal sub-band. In this way, a feature fingerprint reflecting the geometric structure or resonance characteristics of the defect can be constructed, which provides a key basis for identifying the defect types that mainly manifest their existence by changing the energy structure of the signal spectrum, and forms an effective supplement to the HHT instantaneous feature.

[0049] Specifically, in one specific example of the present application, the wavelet packet transform energy feature extraction process first is parameter setting and signal decomposition. First, a pre-set wavelet basis function is selected, such as Daubechies wavelet, and a decomposition layer number is determined, such as 3 layers, which determines the degree of precision of frequency band division. Then, the wavelet packet decomposition is performed on the input filtered signal. This process decomposes the entire frequency band of the signal into low and high frequency parts layer by layer and equally until the pre-set 3-layer decomposition depth is reached, finally generating 8 sub-band signals covering the entire spectrum and being orthogonal to each other. Then, for each final sub-band obtained after decomposition, the energy of the wavelet packet coefficients contained therein is calculated, i.e. the sum of squares of all coefficients in the sub-band is calculated. Finally, the energy values calculated for all 8 sub-bands are normalized (for example, their sum is 1), and these normalized energy values are arranged in order from low to high frequency band to form an 8-dimensional vector, which is the wavelet packet energy feature vector.

[0050] Specifically, in step S300, the HHT instantaneous feature vector and the wavelet packet energy feature vector are fused to obtain a fusion feature vector. It can be understood that since the HHT instantaneous feature vector and the wavelet packet energy feature vector are derived from the same physical signal, there must be shared information and collinearity describing the macroscopic characteristics between them, while each of them contains unique information reflecting different physical mechanisms. If simple splicing fusion is used, the redundant shared information will dilute and drown those unique features that are crucial for accurate classification, thereby reducing the performance of the classification model. Therefore, in the technical solution of the present application, the HHT instantaneous feature vector and the wavelet packet energy feature vector are further fused to obtain a fusion feature vector, so as to actively decouple the two feature sets into shared information components, HHT unique information components and wavelet packet unique information components, and then structure and reorganize these components with clear physical meaning. In this way, a fusion feature vector with internal redundancy removed, information density improved, and key difference features significantly strengthened can be constructed, providing a clear structure and more discriminant input for the subsequent classifier, thereby fundamentally solving the feature confusion problem.

[0051] Figure 4 The flowchart of the fusion of the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a fusion feature vector according to the method for non-destructive testing of praseodymium-neodymium alloy based on acoustic feature analysis of the embodiments of the present application. As shown in FIG. 4, the fusion process includes the following steps: Figure 4As shown, step S300 includes: S310, performing feature preprocessing on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a shared feature vector, an HHT-specific feature vector, and a wavelet packet-specific feature vector; S320, performing structured feature assembly on the shared feature vector, the HHT-specific feature vector, and the wavelet packet-specific feature vector to obtain the fused feature vector.

[0052] Specifically, in step S310, the HHT instantaneous feature vector and wavelet packet energy feature vector are preprocessed to obtain shared feature vectors, HHT-specific feature vectors, and wavelet packet-specific feature vectors. It should be understood that when processing acoustic signals for nondestructive testing of praseodymium-neodymium alloys, the Hilbert-Huang transform instantaneous features and wavelet packet transform energy features extracted from the same original echo are essentially descriptions of the same physical process from different mathematical perspectives. Mathematically, they inevitably contain common information describing the macroscopic physical properties of defects (such as overall energy and size), leading to high correlation and information redundancy between features. At the same time, they each contain key unique information characterizing different physical mechanisms. That is, there is inherent information coupling and collinearity between them. Using a simple feature splicing strategy is tantamount to repeatedly introducing redundant information describing the same physical phenomenon (such as the macroscopic size of the defect) into the model. This not only increases the dimensionality of the feature space but also dilutes the more refined and unique information truly used to distinguish the nature of defects. Furthermore, the computational noise of the two transformation methods (such as the endpoint effect of HHT and the low-energy band noise of WPT) is indiscriminately mixed, forming a high-dimensional, high-noise, and structurally ambiguous feature vector. When such a contaminated and redundant vector is fed into a subsequent classifier, it will seriously interfere with its learning process, making it difficult to focus on subtle but crucial acoustic feature differences such as splitting patterns and pores, thus leading to a decline in classification performance and a lack of model stability. Therefore, in the technical solution of this application, the HHT instantaneous feature vector and wavelet packet energy feature vector are further preprocessed to obtain shared feature vectors, HHT-specific feature vectors, and wavelet packet-specific feature vectors. In this way, through a projection and residual calculation mechanism based on canonical correlation analysis, the original, mutually coupled feature space is decomposed into a shared subspace representing commonalities and two mutually orthogonal unique subspaces representing their respective characteristics. This enables effective decoupling of multi-source feature information, laying the foundation for constructing a fusion feature vector with a clear structure, minimized information redundancy, and highlighted key differences, thus providing the classifier with semantically clear and highly refined input.

[0053] More specifically, in the embodiments of the present application, the HHT instantaneous feature vectors and the wavelet packet energy feature vectors are preprocessed to obtain shared feature vectors, HHT unique feature vectors and wavelet packet unique feature vectors, including: obtaining a training set HHT feature matrix and a training set WPT feature matrix; performing canonical correlation analysis on the training set HHT feature matrix and the training set WPT feature matrix to obtain a canonical basis matrix; based on the canonical basis matrix, the HHT instantaneous feature vectors and the wavelet packet energy feature vectors are decoupled online to obtain the shared feature vectors, the HHT unique feature vectors and the wavelet packet unique feature vectors.

[0054] That is, specifically, first, the off-line learning of shared-unique information bases is performed to obtain the training set HHT feature matrix and the training set WPT feature matrix, which is in the off-line learning stage of the entire intelligent detection system, and lays a data foundation for subsequent establishment of information decoupling model. Specifically, in the off-line stage, a large number of praseodymium-neodymium alloy standard test blocks containing known defect types (such as explicit cracks, pores) and defect-free areas are subjected to ultrasonic detection, and a complete signal processing procedure is performed on each signal sample collected, until the HHT instantaneous feature vectors and the wavelet packet energy feature vectors thereof are generated, respectively. Subsequently, the HHT feature vectors of all samples are stacked by rows to form the training set HHT feature matrix; similarly, the WPT energy feature vectors of all samples are stacked by rows to form the training set WPT feature matrix. The two matrices completely record the mathematical expression set of different defects under two characteristic perspectives in this specific detection scenario.

[0055] Then, canonical correlation analysis is performed on the training set HHT feature matrix and the training set WPT feature matrix to obtain a canonical basis matrix, the purpose of which is to learn and solidify the internal correlation rules between the two features from a large amount of historical data. That is, considering that if the mathematical description of this common information is not established in advance, any online processing will be blind. It is necessary to learn from historical data to find and solidify the strongest correlation axis between the two feature spaces. Since the HHT feature and the WPT feature are derived from the same physical signal, there must be common information between them. Canonical correlation analysis, as a statistical tool, is applied to the two feature matrices, and the core task is to solve a transformation base that can maximize the correlation of the two features after projection, that is, the canonical basis matrix. The two bases constitute a mapping bridge from the original feature space to the shared information subspace, which is expressed in the formula as follows:

[0056] ,

[0057] ,

[0058] ,

[0059] wherein, and denote a transform basis vector or a projection vector for the HHT feature space and a transform basis vector or a projection vector for the WPT feature space, respectively, which are column vectors of the transform matrices A and B, A and B are transform matrices, also called canonical basis matrices, and are the training set HHT feature matrix and the training set WPT feature matrix, respectively, denotes the inner covariance matrix of the training set HHT feature matrix , denotes the inner covariance matrix of the training set WPT feature matrix , denotes the cross covariance matrix between the training set HHT feature matrix and the WPT feature matrix , denotes the cross covariance matrix between the training set WPT feature matrix and the HHT feature matrix , denotes the k-th canonical basis vector obtained by solving the eigen equation, which is the k-th column vector of the canonical basis matrix A, denotes the k-th canonical basis vector obtained by solving the eigen equation, which is the k-th column vector of the canonical basis matrix B, denotes the eigenvalue corresponding to the pair of canonical basis vectors and , denotes the k-th canonical correlation coefficient, denotes the canonical correlation coefficient, i.e. the Pearson correlation coefficient between the two new variables obtained by linearly projecting the feature matrices and , denotes the k-th canonical correlation coefficient, denotes the k-th canonical correlation coefficient, denotes the k-th canonical correlation coefficient, denotes the canonical correlation coefficient, i.e. the Pearson correlation coefficient between the two new variables obtained by linearly projecting the feature matrices and , denotes the maximum value of . This canonical basis matrix is physically equivalent to an information converter in the context of Pr-Nd alloy defect detection, which defines a common feature subspace. Once the HHT and WPT features are projected into this subspace, the common information they contain will be maximally extracted and aligned, so that the common and most stable information components can be efficiently extracted from any subsequent input HHT and WPT features.

[0060] This canonical basis matrix is physically equivalent to an information converter in the context of Pr-Nd alloy defect detection, which defines a common feature subspace. Once the HHT and WPT features are projected into this subspace, the common information they contain will be maximally extracted and aligned, so that the common and most stable information components can be efficiently extracted from any subsequent input HHT and WPT features. ​

[0061] Finally, based on the canonical basis matrix, the HHT instantaneous eigenvector and the wavelet packet energy eigenvector are decoupled online, that is, the projection and residual calculation can be understood. When a new acoustic signal to be tested is processed, this step uses the canonical basis matrix learned in the first step to perform online information decomposition on the newly generated HHT and WPT eigenvectors. This is the core link in the actual detection process for real-time purification and decomposition of unknown defect signal features. That is, the new eigenvector also carries mixed entangled information. In order to achieve accurate classification, it is necessary to purify and decompose it in real time before the classifier intervenes, and to separate the shared information and unique information. When a signal to be tested from a workpiece in the field is processed to obtain its HHT and WPT eigenvectors, the system calls the canonical basis matrix trained in the offline stage to decompose the two entangled eigenvectors into three mutually orthogonal and physically meaningful components: shared eigenvectors, HHT unique eigenvectors and wavelet packet unique eigenvectors. In this way, the transformation from chaos to order is realized. The input of two eigenvectors is effectively decoupled. The shared eigenvectors condense the consensus information from two perspectives, such as the approximate energy response of the defect. More importantly, the HHT unique eigenvector and the wavelet packet unique eigenvector as the residual exactly amplify the weak signals that exist only in a single perspective, such as the crack tip nonlinear response captured in HHT or the specific frequency band resonance of the pore group reflected in WPT. These information is easily submerged in the original mixed vector.

[0062] More specifically, in the embodiments of the present application, based on the canonical basis matrix, the HHT instantaneous feature vector and the wavelet packet energy feature vector are decoupled to obtain a shared feature vector, an HHT unique feature vector and a wavelet packet unique feature vector, including: projecting the HHT instantaneous feature vector and the wavelet packet energy feature vector to the feature space of the canonical basis matrix to obtain an HHT projected feature vector and a wavelet packet projected feature vector. This step is to map the new feature vector through the canonical basis matrix converter to the common subspace that can maximize the commonality, and the two projection vectors obtained are the mathematical expression of the common information part in the original feature. Subsequently, the mean vector of the HHT projected feature vector and the wavelet packet projected feature vector is calculated as the shared feature vector, aiming to fuse the two projection results to obtain a more stable and robust unified description of the common information (such as defect macro size or energy). Finally, the residual vector between the HHT instantaneous feature vector and the HHT projected feature vector is calculated as the HHT unique feature vector, and the residual vector between the wavelet packet energy feature vector and the wavelet packet projected feature vector is calculated as the wavelet packet unique feature vector. The essence of this residual calculation is to eliminate the known common information from the total information, and the remaining part is the unique characteristic information. For example, the HHT unique feature vector will highlight the transient details such as crack tip nonlinear effect, and the wavelet packet unique feature vector will amplify the resonance effect of air holes on the energy of a specific frequency band. These purified unique information is crucial for distinguishing similar defects.

[0063] In general, such a structured representation enables the downstream classifier to learn more fine-grained and robust decision rules. For example, it can learn to give higher weight to the HHT unique feature vector part when judging cracks, and pay more attention to the pattern of the wavelet packet unique feature vector when judging air holes, thereby greatly improving the accuracy and interpretability of classification.

[0064] Specifically, in step S320, the shared feature vector, the HHT-only feature vector and the wavelet packet-only feature vector are structured feature assembled to obtain the fusion feature vector. It should be understood that, since the previous feature preprocessing step has successfully decomposed the original feature into three independent components that are orthogonal to each other in a physical sense and have clear semantics: the shared feature vector, the HHT-only feature vector and the wavelet packet-only feature vector. Therefore, in the technical solution of the present application, the shared feature vector, the HHT-only feature vector and the wavelet packet-only feature vector are further structured feature assembled to obtain the fusion feature vector, so as to integrate these purified and decoupled, complementary information components into a unified, high-dimensional feature representation for subsequent use by the classification model. In this way, it can be ensured that the feature vector finally input into the classifier not only retains common information that can represent the universal attributes of defects, but also highlights key specific information that distinguishes different defect categories, thereby providing the most complete and optimized information basis for achieving high-precision classification.

[0065] More specifically, in one specific example of the present application, the structured feature assembly process is a deterministic vector splicing operation. First, the three independent feature vectors generated in the previous step, i.e., the shared feature vector, the HHT-only feature vector and the wavelet packet-only feature vector, are obtained, and these three vectors are concatenated in a pre-set fixed order, e.g., the shared feature vector first, then the HHT-only feature vector, and finally the wavelet packet-only feature vector, and the resulting vector is the fusion feature vector. This fusion feature vector completely retains all the decoupled information and presents it in a structured manner, which is then directly used as input to the pre-trained support vector machine model to perform the final defect classification decision.

[0066] Specifically, in step S400, the fusion feature vector is input into the pre-trained SVM model to obtain the defect class label. It should be understood that, since the fusion feature vector generated in the previous step is a high-dimensional numerical representation that does not directly provide defect class information, and the distribution boundaries of different class defects (such as cracks and pores) in the feature space are complex and nonlinear. Therefore, in the technical solution of the present application, the fusion feature vector is further input into the pre-trained SVM model to obtain the defect class label, so as to automatically learn and construct a decision boundary that can optimally distinguish the features of different defect classes using the support vector machine (SVM), which is a machine learning model with superior performance in handling high-dimensional, nonlinear classification problems. It is worth mentioning that the defect class label includes a crack label, a pore label and a no defect label. In this way, the complex numerical feature vector can be mapped to an explicit, physically meaningful defect class judgment, thereby achieving automated, intelligent and high-accuracy classification and recognition of internal defects in praseodymium-neodymium alloys.

[0067] More specifically, in one specific example of the present application, the classification process relies on a pre-trained classification model. First, in the offline stage, a large number of known type of Pr-Nd alloy defect samples (including crack, porosity and defect-free samples) are collected, and the complete signal processing and feature fusion process is performed on the signal of each sample to obtain its corresponding fusion feature vector, and each vector is labeled with its known true class label. Then, use this labeled fusion feature vector dataset to train a support vector machine classifier, the training process is to select a suitable kernel function (such as a radial basis function kernel) and optimize its parameters, finally learn the hyperplane that can maximize the separation of different class samples, and solidify and save this trained model. In the subsequent online detection stage, the fusion feature vector generated after the current sample to be tested undergoes all the above steps is directly provided as input to the pre-trained SVM model loaded. The model then applies its internally determined decision function to calculate the input vector, determines its position in the feature space, and finally outputs a unique classification result, which is the defect class label, specifically one of the crack label, porosity label or defect-free label.

[0068] In summary, the Pr-Nd alloy non-destructive testing method based on acoustic feature analysis according to the embodiments of the present application is illustrated, which comprehensively captures the defect information contained in the original probe signal from two complementary physical perspectives of instantaneous dynamic characteristics and frequency band energy distribution by parallel use of Hilbert-Huang transform and wavelet packet transform. Further, this scheme discards simple feature splicing and instead uses canonical correlation analysis as an information decoupling tool to online decompose the two original feature vectors into a shared part describing the commonality of defects, and unique information parts representing the unique resolution capabilities of HHT and wavelet packet respectively. Finally, these three decoupled components are structurally reorganized to form a fusion feature vector that can effectively eliminate redundancy, amplify differences, and has higher information density. A clear and highly refined input is provided for the subsequent classification model, and then the fusion feature is used for defect class recognition, thereby fundamentally improving the recognition accuracy and system stability of similar defects such as cracks and porosities.

[0069] Further, a Pr-Nd alloy non-destructive testing system based on acoustic feature analysis is also provided.

[0070] Figure 5 A block diagram of the Pr-Nd alloy non-destructive testing system based on acoustic feature analysis according to the embodiments of the present application. As Figure 5As shown, the Pr-Nd alloy nondestructive detection system 100 based on acoustic feature analysis according to the embodiment of the present application comprises: a band-pass filtering module 110, configured to perform adaptive band-pass filtering on the collected original probe signal to obtain a filtered signal; a parallel feature extraction module 120, configured to perform parallel feature extraction on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector; a feature fusion module 130, configured to perform feature fusion on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a fusion feature vector; and a defect detection module 140, configured to input the fusion feature vector into a pre-trained SVM model to obtain a defect category label.

[0071] As described above, the Pr-Nd alloy nondestructive detection system 100 based on acoustic feature analysis according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a Pr-Nd alloy nondestructive detection algorithm based on acoustic feature analysis, etc. In one possible implementation, the Pr-Nd alloy nondestructive detection system 100 based on acoustic feature analysis according to the embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the Pr-Nd alloy nondestructive detection system 100 based on acoustic feature analysis can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the Pr-Nd alloy nondestructive detection system 100 based on acoustic feature analysis can also be one of the many hardware modules of the wireless terminal.

[0072] The above has described the embodiments of the present disclosure, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles, practical application, or improvement of the technology in the market of the embodiments, or to enable other ordinary skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for non-destructive testing of praseodymium-neodymium alloys based on acoustic feature analysis, characterized in that, The method comprises the following steps: performing adaptive band-pass filtering on the collected original probe signal to obtain a filtered signal; performing parallel feature extraction on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector; performing feature fusion on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a fusion feature vector; inputting the fusion feature vector into a pre-trained SVM model to obtain a defect category label; wherein the feature fusion on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain the fusion feature vector comprises: performing feature preprocessing on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a shared feature vector, an HHT unique feature vector and a wavelet packet unique feature vector, comprising: obtaining a training set HHT feature matrix and a training set WPT feature matrix; performing canonical correlation analysis on the training set HHT feature matrix and the training set WPT feature matrix to obtain a canonical basis matrix; based on the canonical basis matrix, online feature decoupling is performed on the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain the shared feature vector, the HHT unique feature vector and the wavelet packet unique feature vector; performing structured feature assembly on the shared feature vector, the HHT unique feature vector and the wavelet packet unique feature vector to obtain the fusion feature vector; wherein the online feature decoupling on the HHT instantaneous feature vector and the wavelet packet energy feature vector comprises: projecting the HHT instantaneous feature vector and the wavelet packet energy feature vector into a feature space of the canonical basis matrix to obtain an HHT projected feature vector and a wavelet packet projected feature vector; calculating a mean vector of the HHT projected feature vector and the wavelet packet projected feature vector as the shared feature vector; calculating a residual vector between the HHT instantaneous feature vector and the HHT projected feature vector as the HHT unique feature vector; calculating a residual vector between the wavelet packet energy feature vector and the wavelet packet projected feature vector as the wavelet packet unique feature vector.

2. The method for non-destructive testing of Pr-Nd alloys based on analysis of acoustic characteristics according to claim 1, characterized in that, performing adaptive band-pass filtering on the collected original probe signal to obtain a filtered signal comprises: performing defect echo gating processing on the original probe signal to obtain a gated signal; performing center frequency estimation based on fast Fourier transform on the gated signal to obtain a peak frequency of an energy spectrum; based on the peak frequency of the energy spectrum, performing adaptive band-pass filtering on the gated signal to obtain the filtered signal, wherein the center frequency of the adaptive band-pass filtering is the peak frequency of the energy spectrum.

3. The method for non-destructive testing of Pr-Nd alloys based on analysis of acoustic characteristics according to claim 1, characterized in that, performing parallel feature extraction on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector comprises: performing Hilbert-Huang transform instantaneous feature extraction on the filtered signal to obtain the HHT instantaneous feature vector; performing wavelet packet transform energy feature extraction on the filtered signal to obtain the wavelet packet energy feature vector.

4. The method for non-destructive testing of Pr-Nd alloys based on analysis of acoustic characteristics according to claim 1, characterized in that, The defect category label comprises a crack label, a pore label and a no defect label.

5. An acoustic feature analysis based PrNd alloy non-destructive testing system based on the acoustic feature analysis based PrNd alloy non-destructive testing method according to any one of claims 1 to 4, characterized in that, The method comprises the following steps: a band-pass filtering module for performing adaptive band-pass filtering on the collected original probe signal to obtain a filtered signal; a parallel feature extraction module for performing parallel feature extraction on the filtered signal to obtain an HHT instantaneous feature vector and a wavelet packet energy feature vector; The feature fusion module is configured to fuse the HHT instantaneous feature vector and the wavelet packet energy feature vector to obtain a fused feature vector. The defect detection module is configured to input the fused feature vector into a pre-trained SVM model to obtain a defect category label.

Images