Ultrasound nondestructive detecting echo signal classificating method based on vague plane characteristic
(1) technical field
The invention belongs to the signal processing technology field, be specifically related to a kind of Ultrasonic NDT signal sorting technique.
(2) background technology
When ultrasound wave is propagated, always carrying the various information that characterize the object physical property in object.Ultrasonic NDT is to utilize various propagation characteristics such as the propagation of ultrasound wave in tested workpiece, reflection, refraction, decay, waveform transformation to analyze and the configuration state of evaluating material inside, therefore the key that the identification and the evaluation of ultrasound echo signal is Ultrasonic Detection.Can judge tested workpiece by the echoed signal of common ultra-sonic defect detector detection and whether have defective, but the shape of difficult accurately estimation defective and size etc., thereby be difficult to defective is carried out qualitative and quantitative analysis.From the commercial Application angle, these problems need to be resolved hurrily.
Flaw echoes in the Ultrasonic Detection can be thought to be formed the non-stationary signal that becomes when being in the centre frequency place of transducer modulation by wideband pulse.Traditional time domain or frequency-region signal disposal route do not make full use of the entrained information of ultrasonic signal, and be limited in one's ability to the raising of accuracy of detection and reliability.And the time frequency analysis technology can be in time-frequency plane the time and the frequecy characteristic of shows signal simultaneously, the valuable information that makes in the ultrasonic signal to be comprised shows more fully, thereby more helps the explanation to testing result.
The basic thought of time frequency analysis is the associating function of design time and frequency, with it energy density or the intensity of signal in different time and frequency is described simultaneously, it is mapped to the time signal of one dimension the time-frequency plane of a two dimension, the time domain, the frequency domain character that have reflected signal, the frequency of signal and the corresponding relation of time have been provided, can either know the frequency component that signal contains, also can know the time that each frequency component occurs.At present, the frequency analysis technology is carried out defects detection and when identification of ultrasonic signal in use, and wavelet transformation is the time frequency method of the most normal use, can directly be used for classification to wavelet conversion coefficient, and this moment, intrinsic dimensionality was very big, and the generalization ability of sorter is bad; Also can carry out threshold process to wavelet coefficient, the statistical information that extracts the different numbers of plies then is as signal characteristic, and the intrinsic dimensionality that obtain this moment is less, but will solve appropriate threshold chooses problem.Except wavelet transformation, other utilizes the defect identification method of time-frequency information just to utilize the raw information of time-frequency plane mostly, and is less to the Study on Feature Extraction of time-frequency plane.In addition, during Ultrasonic Detection, the selection of frequency probe is relevant with factors such as test material, thickness of workpiece, resolution, depth of defect and defective directions, in order to obtain the measurand more information, will use the probe of different frequency to detect usually; The defective of different depth, direction and shape, the time of arrival of its reflection echo is normally different, therefore, ultrasound echo signal is being carried out defect recognition and dividing time-like, be the influence of the centre frequency of avoiding echoed signal and echo time of arrival to echo character, before feature extraction, mostly need echoed signal is carried out frequency inverted and time alignment pre-service.Time alignment normally makes the position of first minimum value of flaw echo identical by waveform displacement, frequency inverted is normally transferred to the signal of different center frequency on the reference frequency by interpolation and extraction, this pretreated method is easy to be subjected to The noise, and there are certain error in actual operating frequency of popping one's head in the practical application and nominal frequency, change probe, this error can be more obvious, is difficult to realize accurate conversion by interpolation and extraction.In a word, present existing ultrasound nondestructive detecting echo signal classificating method is subjected to the time of arrival of different flaw echoes and the influence of center probe frequency when detecting, and can't accurately carry out feature extraction.
(3) summary of the invention
The object of the present invention is to provide a kind ofly can provide sufficient time-frequency information, can utilize character such as its TIME SHIFT INVARIANCE, frequency displacement unchangeability, symmetry to simplify the pre-service of signal again and reduces the method for the ultrasound echo signal classification of calculated amount.
The present invention is achieved by the following technical solutions: at first utilize Hilbert transform that the rf echo signal of Ultrasonic NDT is converted to analytic signal; Obtain the ambiguity function of analytic signal then, the fuzzy field that obtains ultrasonic signal is represented; With Karhunen-Loeve transformation the ambiguity function of ultrasonic signal is mapped to new low dimensional feature space, the ambiguity function of training set sample is carried out feature extraction, obtain the orthogonal basis vector of new low dimensional feature space; At last with matrixing obtain sample to be identified in the projection of new low dimensional feature space as feature, realize the ultrasonic signal classification with statistical recognition or neural network recognition method.
Below the present invention is further illustrated, comprise the steps:
The first step is asked the Hilbert transform of ultrasonic signal, obtains analytic signal.
Because only there is positive frequency in the frequency spectrum of analytic signal, can suppress the cross term that is caused by negative frequency in the time-frequency distributions.
In second step, find the solution the ambiguity function of analysing ultrasonic signal.
When representing ultrasonic signal with ambiguity function, the TIME SHIFT INVARIANCE of ambiguity function makes that time of arrival, different echoed signals had the mould of identical ambiguity function, therefore, does not need again echoed signal to be done time alignment when extracting the fuzzy field feature and handles; The frequency displacement unchangeability of ambiguity function shows, the echoed signal that centre frequency is different has the mould of identical ambiguity function, therefore, carry out fuzzy field when representing in the echoed signal that different center frequency probe is obtained, can directly remove the influence of signal center frequency, not need to carry out again frequency inverted when extracting the fuzzy field feature; The symmetry of ambiguity function allows can only consider half vague plane when extracting the fuzzy field feature of echoed signal, so neither can drop-out, can reduce calculated amount again.
In the 3rd step,, the ambiguity function of ultrasonic signal is mapped to low dimensional feature space with Karhunen-Loeve transformation at the training set sample.
If
If j sample representing the i class is signal
Length be N, then its ambiguity function is that a N * N ties up matrix, will
The ambiguity function matrix line up a N
2* 1 column vector, and use
Expression, like this
Just constituted N
2The original feature vector of dimension space; After determining that Karhunen-Loeve transformation produces matrix, to N
2The primitive character of dimension space carries out Karhunen-Loeve transformation, just can obtain the lower new low dimensional feature space of dimension.
Karhunen-Loeve transformation is a kind of best orthogonal transformation of based target statistical property.Utilize the Karhunen-Loeve transformation method to extract the feature of the vague plane of signal, vague plane characteristic can be mapped on the one group of orthogonal basis that helps the signal classification, this method does not have specific (special) requirements to the characteristic of the probability density function of each point in the vague plane, can realize effective feature dimensionality reduction again.
In the 4th step, realize the feature extraction of sample to be identified with matrixing.
At first calculate the ambiguity function of sample to be identified, and ambiguity function is converted to N
2* 1 dimensional vector carries out matrixing then, and wherein each base vector is N
2* 1 ties up, and extracts the feature of sample to be identified.
In the 5th step, realize the ultrasonic signal classification with statistical recognition or neural network recognition method.
Beneficial effect of the present invention has:
1, utilize the mould of ambiguity function that pair time shift and the insensitive characteristic of frequency displacement are arranged, represent ultrasonic signal with it, the time-frequency information of echoed signal both can be provided, can avoid again before feature extraction, will carrying out frequency inverted and the pretreated problem of time alignment echoed signal;
2, extract vague plane characteristic by the Karhunen-Loeve transformation method, can reduce intrinsic dimensionality greatly, and have very strong noiseproof feature, for ultrasonic signal identification provides effective feature.
The present invention is directed to the deficiency that existing ultrasonic signal sorting technique exists, adopt ambiguity function to represent ultrasonic signal, and utilize Karhunen-Loeve transformation to realize that vague plane characteristic extracts, and can be mapped to ambiguity function in the lower dimensional space that helps the signal classification, and have good noise proofness.
(4) description of drawings
Fig. 1-Fig. 4 represents synoptic diagram for the fuzzy field of flaw echoes and noise; Wherein, Fig. 1 represents synoptic diagram for the fuzzy field of flat-bottom hole echoed signal, and Fig. 2 represents synoptic diagram for the fuzzy field of by-pass port echoed signal, and Fig. 3 represents synoptic diagram for the fuzzy field of grain noise, and Fig. 4 represents synoptic diagram for the fuzzy field of white Gaussian noise;
Fig. 5 is the vague plane characteristic extraction block diagram based on the Karhunen-Loeve transformation method.
(5) embodiment
Below in conjunction with the drawings and specific embodiments the present invention is further described:
The first step is asked the Hilbert transform of ultrasonic signal, obtains analytic signal.
Ultrasonic radiofrequency signal as real part, as imaginary part, is obtained multiple analytic signal with the Hilbert transform of radiofrequency signal.
Hilbert transform is defined as:
In second step, find the solution the ambiguity function of analysing ultrasonic signal.
Ambiguity function also is a kind of time-frequency distributions function, is defined as the inverse fourier transform of instantaneous autocorrelation function about time t:
Wherein,
Be instantaneous autocorrelation function.
Ambiguity function converts the signal into time delay-frequency deviation plane, can be interpreted as the autocorrelation function in time-frequency combination territory.Fig. 1 has provided the fuzzy field of the two kinds of common noises (grain noise and white Gaussian noise) in two kinds of artificial defect (by-pass port and flat-bottom hole) echoed signals and the Ultrasonic Detection and has represented (mould of ambiguity function).In conjunction with Fig. 1, the energy of the ambiguity function of flaw echoes mainly accumulates near the initial point, and the ambiguity function of two kinds of flaw echos also shows different features; The energy dissipation of the ambiguity function of white Gaussian noise correspondence is in whole vague plane, and is very obvious with the energy distribution difference of flaw echo ambiguity function; The energy of the ambiguity function of grain noise also relatively disperses in vague plane, and ambiguity function shows very strong correlativity on the time delay axle, compares with the ambiguity function of flaw echo, also has significant difference.
In the 3rd step, the ambiguity function of ultrasonic signal is mapped to low dimensional feature space with Karhunen-Loeve transformation.
If
If j sample representing the i class is signal
Length be N, then its ambiguity function is that a N * N ties up matrix.Will
The ambiguity function matrix line up a N
2* 1 column vector, and use
Expression, like this
Just constituted N
2The original feature vector of dimension space.After determining that Karhunen-Loeve transformation produces matrix, to N
2The primitive character of dimension space carries out Karhunen-Loeve transformation, just can obtain the lower feature space of dimension.
The dimension of dispersion matrix is N
2* N
2In general dimension, directly asks N
2* N
2The eigenwert of matrix and proper vector be difficulty relatively, at this moment, can utilize svd to obtain the eigenwert and the proper vector of dispersion matrix.
In the 4th step, realize the feature extraction of sample to be identified with matrixing.
Obtain the orthogonal basis of transformation matrix according to training set after, can directly extract the feature of sample to be identified with it.Fig. 2 has provided the feature extraction block diagram of sample to be identified.When extracting sample to be identified, calculate the ambiguity function of sample to be identified earlier, and ambiguity function is converted to N
2* 1 dimensional vector, (each base vector is N just to carry out matrixing then
2* 1 dimension) extracts the feature of sample.
Mainly accumulate near the initial point of vague plane and the symmetry problem of ambiguity function if consider defect information, can the zone that vague plane characteristic extracts further be dwindled, to reduce calculated amount and storage space.
The 5th step, the classification of ultrasonic signal---experiment and analysis.
Be the advantage of validation signal vague plane characteristic, its classification results with the feature of small wave converting method extraction is compared ultrasonic signal identification.Because small wave converting method can be subjected to the influence of signal center frequency difference, therefore, will do the frequency inverted pre-service to echoed signal before Feature Extraction of Wavelet Transform, and vague plane characteristic is not done the frequency inverted pre-service to echoed signal when extracting.
Table 1 has been listed the composition situation of training set and test set.Training set and test set respectively comprise 3 class signals (by-pass port echoed signal, flat-bottom hole echoed signal, grain noise), and every class has the sample of 80 known class.Measurand is a copper billet that contains the electron beam weld seam, inside and outside the welded seam area of copper billet, has manually made two kinds of defectives: flat-bottom hole and by-pass port.During detection, the center probe frequency is respectively 4.83MHz and 2.26MHz.The ultrasonic signal that collects is extracted near the artificial defect one section respectively, and when extracting, deliberately make not difference time of arrival of the echoed signal that collects of homogeneous of same defective.Noise signal is a segment signal that intercepts out in not containing the actual detected signal of flaw echo, mainly is made up of grain noise.Each signal length is 256 points, and sampling interval is 0.02 μ s.Most of signal is to be that the probe of 2.26MHz is gathered by centre frequency in the training set, because the information that the probe of different center frequency comprises can be slightly different, therefore, also having comprised several centre frequencies in training set is the signal that the 4.83MHz probe is gathered.The echoed signal of two kinds of centre frequencies respectively accounts for half in the test set.All signals have all carried out normalized before feature extraction, the different caused echo amplitude that is provided with because of gain or other detected parameters when avoiding Ultrasonic Detection and the difference of average.
The composition of table 1 training set and test set
Table 2 is the result with sample classification in the test set shown in the feature his-and-hers watches 1 of the feature of wavelet transformation extraction and the extraction of Karhunen-Loeve transformation method.During Feature Extraction of Wavelet Transform, select the DB4 small echo for use, decomposing the number of plies is 4, the average, amplitude of selecting the details wavelet coefficient of layer 2-4 all side, standard deviation as feature.The generation matrix of Karhunen-Loeve transformation is a dispersion matrix between class, and the characteristic number of extraction is 2.Sorter adopts the probabilistic neural network sorter, and network parameter is chosen by experiment.Two kinds of feature extracting methods are all better to the classification results of by-pass port as can be seen from Table 2.The classification results of the feature of extracting with wavelet transformation is compared, and the classification accuracy rate of the feature that the Karhunen-Loeve transformation method is extracted is higher, and required intrinsic dimensionality also seldom.Also listed the classification accuracy rate situation that centre frequency in the test set is respectively the echoed signal of 4.83MHz and 2.26MHz in the table 2.By classification results as can be seen, although in the training set mostly be the echoed signal of being gathered by the 2.26MHz probe, the echoed signal that the feature that the Karhunen-Loeve transformation method is extracted is gathered the 4.83MHz probe also can well be discerned.The feature that wavelet transformation extracts will obviously be better than classification results to the signal of 4.83MHz centre frequency to the classification results of 2.26MHz center frequency signal, and it is less to the performance gap of two kinds of classifications of defects to extract feature based on Karhunen-Loeve transformation, although the pre-service of frequency unchangeability has been adopted in this explanation, the poor performance of the vague plane that the classification performance that wavelet transformation extracts feature still extracts than Karhunen-Loeve transformation.If sample mainly is the echoed signal that the 4.83MHz probe is gathered in the training set, and test set is constant, also can obtain similar result.
In sum, represent ultrasonic signal with ambiguity function, can utilize character such as its TIME SHIFT INVARIANCE, frequency displacement unchangeability, symmetry to simplify the pre-service of signal and reduce calculated amount, vague plane characteristic by the extraction of Karhunen-Loeve transformation method, can reduce intrinsic dimensionality greatly, for ultrasonic signal identification provides effective feature.
Table 2 classification results of the feature of Karhunen-Loeve transformation and small wave converting method extraction