CN101109732B - Classification Method of Ultrasonic Nondestructive Testing Echo Signals Based on Fuzzy Plane Features - Google Patents

Abstract

The invention relates to a classifying method for the echo signals in ultrasonic non-destructive inspection that is based on fuzzy plane, which comprises the following procedures: first the Hibert transformation of ultrasonic signal is worked out to get the analytic signal; then the fuzzy function of the analytic ultrasonic signal is worked out; finally in respect to the sample of training set, the fuzzy function of the ultrasonic signal is mapped on a characteristic space of low dimensions by K-L transformation. The invention uses a classification method of echo signal based on the non variability of the time shift and frequency shift of mode of fuzzy function, can solve any problem in the pre treating method at characteristic extracting, overcome the influence on characteristic extracting from the arrival times of echo signals of different defects and from the central frequency of probe at inspection. The invention is of good noise proofness, good classification effect, and is of a wide application prospect.

Description

Classification Method of Ultrasonic Nondestructive Testing Echo Signals Based on Fuzzy Plane Features

(1) Technical field

The invention belongs to the technical field of signal processing, and in particular relates to a method for classifying ultrasonic nondestructive testing signals.

(2) Background technology

When ultrasonic waves propagate in objects, they always carry various information that characterizes the physical properties of objects. Ultrasonic non-destructive testing is to analyze and evaluate the structural state inside the material by using various propagation characteristics of ultrasonic waves in the inspected workpiece, such as propagation, reflection, refraction, attenuation, and waveform conversion. Therefore, the identification and evaluation of ultrasonic echo signals key. The echo signal detected by an ordinary ultrasonic flaw detector can determine whether there is a defect in the inspected workpiece, but it is difficult to accurately estimate the shape and size of the defect, making it difficult to conduct qualitative and quantitative analysis of the defect. From the perspective of industrial application, these problems need to be solved urgently.

The defect echo signal in ultrasonic testing can be considered as a time-varying non-stationary signal modulated by a broadband pulse at the center frequency of the transducer. Traditional time-domain or frequency-domain signal processing methods do not make full use of the information carried by ultrasonic signals, and have limited ability to improve detection accuracy and reliability. The time-frequency analysis technology can simultaneously display the time and frequency characteristics of the signal in the time-frequency plane, so that the valuable information contained in the ultrasonic signal can be more fully displayed, which is more conducive to the interpretation of the detection results.

The basic idea of time-frequency analysis is to design the joint function of time and frequency, and use it to describe the energy density or intensity of the signal at different times and frequencies at the same time. It maps the one-dimensional time signal to a two-dimensional time-frequency plane, reflecting the The time-domain and frequency-domain characteristics of the signal give the corresponding relationship between the frequency and time of the signal, which can not only know the frequency components contained in the signal, but also know the time when each frequency component appears. At present, when using time-frequency analysis technology to detect and identify defects of ultrasonic signals, wavelet transform is the most commonly used time-frequency method, and wavelet transform coefficients can be directly used for classification. At this time, the feature dimension is large, and the classifier The generalization ability is not good; it is also possible to perform threshold processing on wavelet coefficients, and then extract statistical information of different layers as signal features. At this time, the obtained feature dimensions are less, but the problem of appropriate threshold selection must be solved. In addition to wavelet transform, most other defect identification methods using time-frequency information only use the original information of the time-frequency plane, and there are few studies on the feature extraction of the time-frequency plane. In addition, during ultrasonic testing, the selection of the probe frequency is related to factors such as the test material, workpiece thickness, resolution, defect depth, and defect direction. In order to obtain more information about the measured object, it is usually necessary to use different frequency probes for detection; different For defects of depth, direction and shape, the arrival time of the reflected echo is usually different. Therefore, when identifying and classifying ultrasonic echo signals, in order to avoid the center frequency of the echo signal and the arrival time of the echo. Because of the impact of wave features, frequency conversion and time alignment preprocessing of echo signals are mostly required before feature extraction. Time alignment is usually to make the position of the first minimum value of the defect echo the same through waveform shifting. Frequency conversion is usually to transfer signals of different center frequencies to a reference frequency through interpolation and decimation. This preprocessing method It is easily affected by noise, and there is a certain error between the actual working frequency and the nominal frequency of the probe in practical applications. If the probe is replaced, this error will be more obvious, and it is difficult to achieve accurate transformation through interpolation and decimation. In short, the existing ultrasonic nondestructive testing echo signal classification methods are affected by the arrival time of different defect echo signals and the center frequency of the probe during detection, and cannot accurately extract features.

(3) Contents of the invention

The object of the present invention is to provide a kind of ultrasonic echo signal classification that can provide sufficient time-frequency information, and can use its time shift invariance, frequency shift invariance, symmetry and other properties to simplify the preprocessing of the signal and reduce the amount of calculation Methods.

The present invention is realized through the following technical solutions: firstly, the radio frequency echo signal of ultrasonic nondestructive testing is converted into an analytical signal by Hilbert transform; then the fuzzy function of the analytical signal is obtained to obtain the fuzzy domain representation of the ultrasonic signal; The transformation maps the fuzzy function of the ultrasonic signal to a new low-dimensional feature space, extracts the features of the fuzzy function of the training set samples, and obtains the orthogonal basis vector of the new low-dimensional feature space; The projection of the new low-dimensional feature space is used as a feature, and the classification of ultrasonic signals is realized by using statistical recognition or neural network recognition methods.

Below the present invention is described further, comprises the steps:

The first step is to obtain the Hilbert transform of the ultrasonic signal to obtain the analytical signal.

Since only positive frequencies exist in the spectrum of the analytic signal, cross terms caused by negative frequencies in the time-frequency distribution can be suppressed.

The second step is to find the ambiguity function of the analytical ultrasonic signal.

When the ultrasonic signal is represented by the fuzzy function, the time-shift invariance of the fuzzy function makes the echo signals with different arrival times have the same modulus of the fuzzy function. Therefore, it is not necessary to perform time alignment processing on the echo signal when extracting fuzzy domain features. ; The frequency shift invariance of the ambiguity function shows that the echo signals with different center frequencies have the same modulus of the ambiguity function. Therefore, when expressing the echo signals obtained by probes with different center frequencies in the fuzzy domain, the center frequency of the signal can be directly removed No frequency conversion is required when extracting fuzzy domain features; the symmetry of the fuzzy function allows only half of the fuzzy plane to be considered when extracting fuzzy domain features of the echo signal, which will not lose information and reduce Calculations.

The third step is to use K-L transform to map the ambiguity function of the ultrasonic signal to the low-dimensional feature space for the samples in the training set.

表示第i类的第j个样本，如果信号

的长度为N，则其模糊函数是一个N×N维矩阵，将

的模糊函数矩阵排成一个N²×1的列向量，并用表示，这样

就构成了N²维空间的原始特征向量；确定K-L变换产生矩阵后，对N²维空间的原始特征进行K-L变换，就可得到维数较低的新的低维特征空间。set up

Indicates the j-th sample of the i-th class, if the signal

The length of is N, then its fuzzy function is an N×N dimensional matrix, the

The fuzzy function matrix is arranged into an N ² × 1 column vector, and used said so

The original feature vector of the N ^2- dimensional space is formed; after the KL transformation matrix is determined, the KL transformation is performed on the original features of the N ^2- dimensional space to obtain a new low-dimensional feature space with a lower dimension.

The K-L transform is an optimal orthogonal transform based on the statistical properties of the target. Using the K-L transform method to extract the features of the fuzzy plane of the signal, the features of the fuzzy plane can be mapped to a group of orthogonal bases that are beneficial to signal classification. This method has no special requirements for the characteristics of the probability density function of each point in the fuzzy plane. It can also achieve effective feature dimensionality reduction.

The fourth step is to use matrix transformation to realize the feature extraction of the sample to be recognized.

First calculate the fuzzy function of the sample to be identified, and convert the fuzzy function into an N ² ×1-dimensional vector, and then perform matrix transformation, in which each base vector is N ² ×1-dimensional, and extract the characteristics of the sample to be identified.

The fifth step is to use statistical recognition or neural network recognition methods to realize ultrasonic signal classification.

The beneficial effects of the present invention have:

1. The mold using the fuzzy function is not sensitive to time shift and frequency shift. Using it to represent the ultrasonic signal can not only provide the time-frequency information of the echo signal, but also avoid the frequency of the echo signal before feature extraction. Transformation and time alignment preprocessing issues;

2. Extracting fuzzy plane features by K-L transformation method can greatly reduce the feature dimension, and has strong anti-noise performance, providing effective features for ultrasonic signal recognition.

The present invention aims at the deficiencies in the existing ultrasonic signal classification methods, uses fuzzy functions to represent ultrasonic signals, and utilizes K-L transformation to realize fuzzy plane feature extraction, can map fuzzy functions to low-dimensional spaces that are beneficial to signal classification, and has excellent performance noise immunity.

(4) Description of drawings

Fig. 1-Fig. 4 are schematic diagrams of fuzzy domain representation of defect echo signal and noise; among them, Fig. 1 is a schematic diagram of fuzzy domain representation of flat-bottomed hole echo signal, Fig. 2 is a schematic diagram of fuzzy domain representation of bypass hole echo signal, Fig. 3 is a schematic representation of grain noise in the fuzzy domain, and FIG. 4 is a schematic representation of Gaussian white noise in the fuzzy domain;

Fig. 5 is a block diagram of fuzzy plane feature extraction based on K-L transformation method.

(5) Specific implementation methods

The present invention will be further described below in conjunction with accompanying drawing and specific embodiment:

The first step is to obtain the Hilbert transform of the ultrasonic signal to obtain the analytical signal.

The ultrasonic radio frequency signal is used as the real part, and the Hilbert transform of the radio frequency signal is used as the imaginary part to obtain the complex analytical signal.

The Hilbert transform is defined as:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ]</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; \&delta;</span>}<span\; class="notranslate">\; \&Right\; Arrow;</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; lim</span>}}{<span\; class="notranslate">\; \&Integral;</span>}_{<span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&infin;</span>}^{\mathrm{<span\; class="notranslate">\; \&delta;</span>}}\frac{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&pi;\&tau;</span>}}\mathrm{<span\; class="notranslate">\; d\&tau;</span>}<span\; class="notranslate">\; +</span>{<span\; class="notranslate">\; \&Integral;</span>}_{\mathrm{<span\; class="notranslate">\; \&delta;</span>}}^{<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&infin;</span>}\frac{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&pi;\&tau;</span>}}\mathrm{<span\; class="notranslate">\; d\&tau;</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

The second step is to find the ambiguity function of the analytical ultrasonic signal.

The fuzzy function is also a time-frequency distribution function, which is defined as the inverse Fourier transform of the instantaneous autocorrelation function with respect to time t:

$\mathrm{<span\; class="notranslate">\; AF</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&nu;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; \&Integral;</span>}_{<span\; class="notranslate">\; -</span><span\; class="notranslate">\; \&infin;</span>}^{<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&infin;</span>}{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; \&tau;</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; <figure-callout\; id="2"\; label="e\; j"\; filenames="S07172628220070904E000011.png,S07172628220070904E000013.png"\; state="\{\{state\}\}">e</figure-callout></span><figure-callout\; id="2"\; label="e\; j"\; filenames="S07172628220070904E000011.png,S07172628220070904E000013.png"\; state="\{\{state\}\}">\; </figure-callout>}}^{\mathrm{<span\; class="notranslate">\; </span>}}\mathrm{<span\; class="notranslate">\; dt</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

in, $R_t(t,\mathrm{\&tau;})=z(t+\frac{\mathrm{\&tau;}}{2})z^{*}(t-\frac{\mathrm{\&tau;}}{2}),$ is the instantaneous autocorrelation function.

The ambiguity function transforms the signal into the delay-frequency offset plane, which can be interpreted as an autocorrelation function in the joint time-frequency domain. Figure 1 shows the fuzzy domain representation (the modulus of the fuzzy function) of echo signals of two artificial defects (bypass holes and flat bottom holes) and two common noises (grain noise and Gaussian white noise) in ultrasonic testing. Combined with Figure 1, the energy of the ambiguity function of the defect echo signal is mainly concentrated near the origin, and the ambiguity functions of the two defect echoes also show different characteristics; the energy of the ambiguity function corresponding to Gaussian white noise is distributed in the entire fuzzy plane , and the energy distribution of the defect echo ambiguity function is very different; the energy of the grain noise ambiguity function is also scattered in the fuzzy plane, and the ambiguity function shows a strong correlation on the time delay axis, and the defect echo Compared with the fuzzy function of , there is also a significant difference.

In the third step, K-L transformation is used to map the ambiguity function of the ultrasonic signal to a low-dimensional feature space.

表示第i类的第j个样本，如果信号

的长度为N，则其模糊函数是一个N×N维矩阵。将

的模糊函数矩阵排成一个N²×1的列向量，并用

表示，这样就构成了N²维空间的原始特征向量。确定K-L变换产生矩阵后，对N²维空间的原始特征进行K-L变换，就可得到维数较低的特征空间。set up

Indicates the j-th sample of the i-th class, if the signal

The length of is N, then its fuzzy function is an N×N dimensional matrix. Will

The fuzzy function matrix is arranged into an N ² × 1 column vector, and used

said so It constitutes the original feature vector of the N ^2- dimensional space. After determining the KL transform generation matrix, the KL transform is performed on the original features of the N ^2- dimensional space to obtain a feature space with a lower dimension.

The dimension of the discrete degree matrix is N ² ×N ² dimensions. Generally speaking, it is difficult to directly obtain the eigenvalues and eigenvectors of the N ² ×N ² matrix. At this time, the singular value decomposition can be used to obtain the characteristics of the discrete degree matrix values and eigenvectors.

The fourth step is to use matrix transformation to realize the feature extraction of the sample to be recognized.

After obtaining the orthogonal basis of the transformation matrix according to the training set, it can be directly used to extract the features of the samples to be recognized. Figure 2 shows the feature extraction block diagram of the samples to be identified. When extracting samples to be identified, the fuzzy function of the sample to be identified must be calculated first, and the fuzzy function is converted into an N ² ×1-dimensional vector, and then the matrix transformation (each base vector is N ² ×1-dimensional) can be performed to extract the sample Characteristics.

Considering that the defect information is mainly gathered near the origin of the fuzzy plane and the symmetry of the fuzzy function, the feature extraction area of the fuzzy plane can be further reduced to reduce the amount of calculation and storage space.

The fifth step is the classification of ultrasonic signals—experiment and analysis.

In order to verify the superiority of signal fuzzy plane features for ultrasonic signal recognition, the classification results of features extracted by wavelet transform method were compared. Because the wavelet transform method will be affected by the difference of the center frequency of the signal, the frequency conversion preprocessing of the echo signal is required before the wavelet transform feature extraction, but the frequency conversion preprocessing of the echo signal is not performed during the fuzzy plane feature extraction.

Table 1 lists the composition of the training set and test set. The training set and the test set each contain 3 types of signals (bypass hole echo signal, flat bottom hole echo signal, grain noise), and each type has 80 samples of known categories. The object to be tested is a copper block containing electron beam welds. Two kinds of defects are artificially manufactured inside and outside the weld area of the copper block: flat bottom holes and bypass holes. When testing, the center frequencies of the probes are 4.83MHz and 2.26MHz respectively. A section near the artificial defect is extracted from the collected ultrasonic signals, and the arrival time of the echo signals obtained by different acquisitions of the same defect is deliberately made different when extracting. The noise signal is a section of signal intercepted from the actual detection signal without defect echo, mainly composed of grain noise. The length of each signal is 256 points, and the sampling interval is 0.02μs. Most of the signals in the training set are collected by probes with a center frequency of 2.26MHz. Since the information contained in probes with different center frequencies will be slightly different, the training set also includes several signals collected by probes with a center frequency of 4.83MHz. . The echo signals of the two center frequencies in the test set each account for half. All signals are normalized before feature extraction, in order to avoid the difference in echo amplitude and mean value caused by the difference in gain or other detection parameter settings during ultrasonic detection.

Table 1 Composition of training set and test set

Table 2 shows the results of classifying the samples in the test set shown in Table 1 using the features extracted by wavelet transform and K-L transform method. When wavelet transform features are extracted, DB4 wavelet is selected, the number of decomposition layers is 4, and the mean value, amplitude mean square, and standard deviation of detail wavelet coefficients of layers 2-4 are selected as features. The generation matrix of K-L transformation is the inter-class dispersion matrix, and the number of extracted features is 2. The classifier adopts a probabilistic neural network classifier, and the network parameters are selected by experiments. It can be seen from Table 2 that the two feature extraction methods have better classification results for bypass holes. Compared with the classification results of features extracted by wavelet transform, the classification accuracy of features extracted by K-L transform method is higher, and the required feature dimension is less. Table 2 also lists the classification accuracy of echo signals whose center frequencies are 4.83MHz and 2.26MHz in the test set. It can be seen from the classification results that although most of the training set is the echo signal collected by the 2.26MHz probe, the features extracted by the K-L transform method can also be well recognized for the echo signal collected by the 4.83MHz probe. The classification results of the features extracted by the wavelet transform for the 2.26MHz center frequency signal are significantly better than the classification results for the 4.83MHz center frequency signal, and the performance gap between the two types of defect classification based on the K-L transform extraction features is small, which shows that despite the use of Without frequency invariance preprocessing, the classification performance of features extracted by wavelet transform is still worse than that of fuzzy planes extracted by K-L transform. If the samples in the training set are mainly the echo signals collected by the 4.83MHz probe, and the test set remains unchanged, similar results can also be obtained.

To sum up, using the fuzzy function to represent the ultrasonic signal can use its time shift invariance, frequency shift invariance, symmetry and other properties to simplify the preprocessing of the signal and reduce the amount of calculation. The fuzzy plane features extracted by the K-L transform method, It can greatly reduce the feature dimension and provide effective features for ultrasonic signal recognition.

Table 2 Classification results of features extracted by K-L transform and wavelet transform

Claims (3)

1. A fuzzy plane characteristic-based ultrasonic nondestructive detection echo signal classification method is characterized in that firstly, a radio frequency echo signal of ultrasonic nondestructive detection is converted into an analytic signal by using Hilbert transform; then solving a fuzzy function of the analytic signal to obtain fuzzy domain representation of the ultrasonic signal; mapping the fuzzy function of the ultrasonic signal to a new low-dimensional feature space by using K-L transformation, and performing feature extraction on the fuzzy function of the training set sample to obtain an orthogonal basis vector of the new low-dimensional feature space; and finally, solving the projection of the sample to be identified in a new low-dimensional feature space by using matrix transformation as a feature, and realizing ultrasonic signal classification by using a statistical identification or neural network identification method.

2. The method for classifying ultrasonic nondestructive testing echo signals based on the blurred plane features as claimed in claim 1, wherein said step of mapping the blurring function of the ultrasonic signals to the new low-dimensional feature space by using K-L transformation comprises: let x_j ⁽ⁱ⁾J-th sample, signal x, representing class i_j ⁽ⁱ⁾Is N, the fuzzy function is an NxN dimensional matrix, x is_j ⁽ⁱ⁾Is arranged as an N²X 1 column vector, used in combination

Is shown as suchForm N²Original feature vectors of a dimensional space; after determining the K-L transformation generating matrix, for N²And performing K-L transformation on the original features of the dimensional space to obtain a new low-dimensional feature space with lower dimensions.

3. The ultrasonic nondestructive detection echo signal classification method based on the fuzzy planar characteristic as claimed in claim 1 is characterized by the steps of finally using matrix transformation to obtain the projection of the sample to be identified in the new low-dimensional characteristic space as the characteristic: firstly, calculating a fuzzy function of a sample to be identified, and converting the fuzzy function into N²X 1-dimensional vectors, each base vector being N, and then matrix-transforming²X 1 dimension, extracting the characteristics of the sample to be identified; n is a signal x_j ⁽ⁱ⁾Length of (1), x_j ⁽ⁱ⁾Represents the jth sample of the ith class.

Images