Two-stage ultrasonic non-destructive testing signal high-precision identification method

Abstract

The application provides a two-stage ultrasonic nondestructive testing signal high-precision identification method, adopts a two-stage mode to separate ultrasonic normal signal classification and abnormal signal detection, solves the problem that abnormal signal modes are various and cannot be exhausted while ensuring the accuracy of normal signal classification, thereby improving the overall identification precision of ultrasonic nondestructive testing signals; adopts an envelope line feature extraction and fusion method for ultrasonic normal signals, eliminates the positive and negative oscillation phenomenon of ultrasonic signals while retaining the echo position and intensity features of ultrasonic signals, reduces the sensitivity of the algorithm to single values, and further reduces the influence caused by the time shift of normal signal echo features and noise interference. Finally, the local similarity features of abnormal signals are extracted and enhanced, thereby solving the problem that abnormal signal modes are various and cannot be exhausted.

Description

Technical Field

[0001] This invention belongs to the field of ultrasonic nondestructive testing signal classification technology, specifically relating to a two-stage ultrasonic nondestructive testing signal high-precision identification method. Background Technology

[0002] In recent years, with the rapid development of the national economy and increasingly fierce international competition for energy, oil and natural gas, as important energy sources, play an indispensable role in national development and improving people's lives. Pipeline transportation is the main mode of oil and gas transportation, making the safe operation of pipelines crucial. However, pipeline transportation operates under extremely harsh conditions, and under the influence of various factors, corrosion and cracks are prone to occur, even leading to oil and gas leaks due to pipeline damage. Oil and gas leaks not only cause huge economic losses but also cause serious pollution to the ecological environment. Therefore, regular non-destructive testing of oil and gas pipelines is essential to ensure the safety of oil and gas transportation.

[0003] Ultrasonic testing (UT) is the most successful non-destructive testing (NDT) technique for quality assessment and defect detection of engineering materials. In industrial applications, high-precision identification of ultrasonic NDT data can effectively improve the accuracy of defect detection identification and inversion.

[0004] Currently, high-precision identification methods for ultrasonic nondestructive testing signals mainly fall into two categories: signal-based methods and data-driven methods. Data-driven methods, which can automatically extract deep features of signals, are widely used in signal recognition. However, on the one hand, the complex pipeline environment can cause severe noise interference to ultrasonic testing signals; on the other hand, due to the absorption and divergence of ultrasonic signals caused by special locations such as weld defects within the pipeline, ultrasonic testing data often contains a large number of irregular abnormal signals, whose patterns cannot usually be exhaustively represented using data annotation methods. For these two reasons, the current methods still have relatively low accuracy in identifying ultrasonic nondestructive testing signals. Summary of the Invention

[0005] To address the shortcomings of the existing technology, this invention provides a two-stage ultrasonic nondestructive testing signal high-precision identification method, specifically including the following:

[0006] Step 1: Label the ultrasonic nondestructive testing signals as normal and abnormal signals to construct a dataset of ultrasonic nondestructive testing signals; specifically:

[0007] Construct a secondary echo sample set S based on the secondary echo signal in the normal signal. 2 Construct a three-echo signal sample set S based on the three echo signals in the normal signal. 3 Construct a four-echo signal sample set S based on the four echo signals in the normal signal.4 ;

[0008] Construct an ultrasonic abnormal signal dataset S based on the abnormal signals 1 ;

[0009] Step 2: Construct a normal signal classification model for ultrasonic non-destructive testing, which is used to classify ultrasonic signals into three categories, namely: secondary echo signals, tertiary echo signals, and quaternary echo signals (the first stage). Specifically, it is expressed as:

[0010] Step 2.1: Use the second-order difference method to extract the local maximum point set LA and the local minimum point set LI of the ultrasonic signal D = [d1 d2…d i …d n ;

[0011] Step 2.2: Obtain the upper envelope and the lower envelope of the ultrasonic signal D. Specifically, it is expressed as:

[0012] Step 2.2.1: Use the Akmin data interpolation method to interpolate LA and LI respectively to obtain the basic upper envelope and the basic lower envelope representing the local extreme points on the upper and lower edges;

[0013] Step 2.2.2: Use formula (1) to obtain the upper envelope EU:

[0014]

[0015] where n is the dimension of the ultrasonic signal, and θ, k are envelope correlation control parameters, which are summarized from a large number of engineering experiences. When θ = 0.8 and k = 2, the effect is better. Specifically, when i < k or i > n - k,

[0016] Step 2.2.3: Use formula (2) to obtain the lower envelope ED:

[0017]

[0018] when i < k or i > n - k,

[0019] Step 2.3: For the upper envelope EU, the lower envelope ED, and the input signal D, use n independent 3*3 convolution kernels W i to perform fusion. The fused feature vector is F = [f1 f2 …f i … f n , and the fusion process is expressed as formula (3):

[0020]

[0021] in, W represents an independent convolution kernel i The element in row 0 and column j, especially when i = 0 or n, f i =d i ;

[0022] Step 2.4: Use a one-dimensional convolutional neural network as the fused feature classifier to classify the sample set S. 2 S 3 S 4 Perform steps 2.1 to 2.2, using the output of step 2.2 to train the n independent convolutional kernels of step 2.3 and the one-dimensional convolutional neural network of step 2.4, ultimately generating the ultrasonic nondestructive testing normal signal classification model F. N ;

[0023] Step 2.5: Input the ultrasonic nondestructive testing dataset G to be tested into the ultrasonic nondestructive testing normal signal classification model F generated in step 2.4. N In the output, three classes of the signal to be detected are: the secondary echo dataset GS. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 It also outputs the probability that each sample represents the current result, i.e., GS. 2 The probability that each signal belongs to the second echo signal is ps. 2 GS 2 The probabilities of all signals belonging to the second echo signal together constitute the probability vector PS. 2 GS 3 The probability that each signal belongs to a third-echo signal is ps. 3 GS 3 The probabilities of all signals belonging to the third echo signal together constitute the probability vector PS. 3 GS 4 The probability that each signal belongs to the fourth echo signal is ps. 4 GS 4 The probabilities of all signals belonging to the fourth echo signal together constitute the probability vector PS. 4 ;

[0024] Step 3: Construct an ultrasonic nondestructive testing abnormal signal detection model F A (Second stage) Anomaly detection is performed on each classification dataset generated in step 2; specifically:

[0025] Step 3.1: For the secondary echo dataset GS 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Definition: A dataset GS vThe signal in H is represented as H = [h1 h2 … h n ], v = 2 or 3 or 4, dataset GS v The corresponding probability is PS v Set a window with a width of wide and a starting position of index=0. The window will move along the time axis of the ultrasound signal with a step size of str.

[0026] Step 3.2: Extract a local sub-signal h of the signal H within the window, which can be expressed as formula (4):

[0027] h = [h] index h index+1 …h index+wide (4)

[0028] In the formula, h index+wide This represents the (index+wide)th data point in signal H;

[0029] Step 3.3: Extract the dataset GS v For each local sub-signal, the extraction process is the same as in step 3.2, forming the dataset GS. v Local sub-signal sample set in For local sub-signal sample set A sub-signal sample, where m is the number of samples;

[0030] Step 3.4: Use the Dynamic Time Warping (DTW) algorithm to extract the local sub-signal h and each local sub-signal sample in the local sub-signal sample set. The similarity is calculated, and the average of the similarities of all sub-signal samples is taken as the local feature similarity ls of the input signal H in the current window. index ;

[0031] Step 3.5: Move str time units, that is, update index to index+str, and repeat steps 3.2 to 3.4 to extract local similarity features at different positions, and finally form the local feature similarity vector LS′ of the input signal H. LS′ can be expressed as formula (5):

[0032] LS′=[ls0′ls1′ … ls k ′]k*str+wide<n (5)

[0033] Step 3.6: Utilize probability ps v The local feature similarity of the signal is enhanced, and the enhanced local similarity feature vector is LS = [ls0 ls1 …ls]. i … ls kThe enhancement method is as shown in formula (6):

[0034] ls i =exp(λ*(1-ps) v ))*ls i ′i∈[0,k] (6)

[0035] In the formula, λ is the similarity feature vector enhancement coefficient;

[0036] Step 3.7: Classify the enhanced local feature similarity using a two-layer linear neural network, using the dataset GS. v and abnormal signal dataset S 1 The training process is performed, and the dataset GS is eventually generated. v Ultrasonic nondestructive testing abnormal signal detection model F corresponding to echo signal v A ;

[0037] Step 3.8: Repeat steps 3.1 to 3.7 to construct the second echo dataset GS respectively. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Ultrasonic nondestructive testing abnormal signal detection model F v A v = 2, 3, 4, used to detect abnormal signals in each dataset.

[0038] For the signal to be detected, the ultrasonic nondestructive testing normal signal classification model generated in step 2 is first executed to classify it as a certain type of normal signal, such as a secondary echo signal, a tertiary echo signal, or a quaternary echo signal, and the probability of the signal being a certain type of normal signal is output. Then, the signal is input into the ultrasonic nondestructive testing abnormal signal detection model generated in step 3 to determine whether it is an abnormal signal. If it is an abnormal signal, an abnormal label is directly output; otherwise, the normal signal label output in step 2 is output.

[0039] The beneficial effects of this invention are as follows: First, by employing a two-stage approach to separate the classification of normal ultrasound signals and the detection of abnormal signals, the accuracy of normal signal classification is ensured while resolving the problem of the inexhaustible diversity of abnormal signal patterns, thereby improving the overall recognition accuracy of ultrasound non-destructive testing signals. Second, by using a method of envelope feature extraction and fusion of normal ultrasound signals, the echo position and intensity characteristics of the ultrasound signals are preserved while eliminating the positive and negative amplitude oscillations of the ultrasound signals, reducing the algorithm's sensitivity to individual values ​​and further mitigating the impact of time shift and noise interference on normal signal echo features. Finally, by extracting and enhancing the local similarity features of abnormal signals, the problem of the inexhaustible diversity of abnormal signal patterns is resolved. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the two-stage ultrasonic nondestructive testing signal high-precision identification method proposed in this invention;

[0041] Figure 2 The ultrasonic nondestructive testing dataset in this invention includes three types of normal signals: (a) is a normal signal with two echoes, (b) is a normal signal with three echoes, and (c) is a normal signal with four echoes.

[0042] Figure 3 These are typical abnormal signals in the ultrasonic nondestructive testing dataset of this invention; where (a) is an abnormal signal caused by external defects, (b) is an abnormal signal caused by internal defects, and (c) is an abnormal signal caused by welds.

[0043] Figure 4 This is a schematic diagram of local similarity feature extraction of abnormal signals in this invention;

[0044] Figure 5 This is a flowchart of the two-stage high-precision identification method for intra-ultrasound detection signals in this invention. Detailed Implementation

[0045] The invention will be further explained below with reference to the accompanying drawings and specific implementation examples.

[0046] like Figure 1 , Figure 5 As shown, a two-stage high-precision identification method for ultrasonic nondestructive testing signals includes:

[0047] Step 1: Label the ultrasonic non-destructive testing (NDT) signals as normal and abnormal signals to construct an NDT signal dataset. Normal signals are those with distinct echo characteristics, allowing for the calculation of the pipe wall thickness. Abnormal signals, caused by the divergence and absorption of ultrasonic waves due to weld defects, lack distinct echo characteristics, making pipe wall thickness calculation impossible. Based on the number of times the ultrasonic signal is reflected by the pipe wall, normal signals can be further categorized into secondary, tertiary, and quaternary echo signals. The normal signals in the NDT dataset are selected, and a normal signal dataset is generated through manual labeling. This dataset consists of three signal types: secondary, tertiary, and quaternary echo signals.

[0048] like Figure 2 and Figure 3 As shown, the ultrasonic nondestructive testing signal dataset S consists of the normal signal sample set S n and abnormal signal sample set S 1Composition. A secondary echo sample set S is constructed based on the secondary echo signal in the normal signal. 2 Construct a three-echo signal sample set S based on the three echo signals in the normal signal. 3 Construct a four-echo signal sample set S based on the four echo signals in the normal signal. 4 Generate a normal signal dataset R n As shown in the formula below:

[0049] S n =S 2 ∪S 3 ∪S 4

[0050] Among them, S 2 S is a sample set of two echo signals. 3 S is a sample set of three echo signals. 4 This is a sample set of three echo signals.

[0051] The abnormal signal dataset is obtained by filtering out normal signals from the ultrasonic nondestructive testing dataset. An ultrasonic abnormal signal dataset S is constructed based on these abnormal signals. 1 ;

[0052] Abnormal signals are all signals in the ultrasonic nondestructive testing dataset that do not meet the conditions for calculating pipe wall thickness, i.e.:

[0053]

[0054] Step 2: Construct a normal signal classification model for ultrasonic nondestructive testing to classify ultrasonic signals into three categories: secondary echo signals, tertiary echo signals, and quaternary echo signals (first stage). This step does not consider the existence of abnormal signals, but only classifies the signals to be detected into three categories of normal signals with different echo counts. First, the envelope features of the signal are extracted. Second, the envelope features are fused with the original signal. Finally, the Softmax function is used as the top-level one-dimensional convolutional neural network to output the signal's category and category probability. The upper and lower envelope features, while preserving the echo position and intensity characteristics of the ultrasonic signal, eliminate the amplitude oscillation of the ultrasonic signal. Specifically, they are described as follows:

[0055] Step 2.1: Extract the ultrasound signal D = [d1 d2 …d] using the quadratic difference method. i … d n The local maxima set LA and local minima set LI are determined by the following operations:

[0056] MaxI=Ind(dif⊙sign⊙dif(X)==-2)+1;

[0057] MaxV = X i, i ∈ MaxI;

[0058] MinI = I(dif ⊙ sign ⊙ dif(X) == 2) + 1, MinV = X i , i ∈ MinI;

[0059] Where MaxI, MinI, MaxI, MinV represent the local maximum sampling time, local maximum, local minimum sampling time, and local minimum of the ultrasonic signal respectively. The I flag takes the abscissa (sampling time), dif represents the data difference operation, ⊙ represents function nesting, and sign is the sign function.

[0060] Compared with the cubic spline interpolation method, the data obtained by the Akima interpolation method is smoother. However, the Akima interpolation method only uses local extreme points to generate envelope line features. In order to let non-extreme points participate in the envelope line calculation and further strengthen the connection between the envelope line and the original signal, the present invention improves the Akima data interpolation method as follows:

[0061] Step 2.2: Obtain the upper envelope line and lower envelope line of the ultrasonic signal D; specifically including:

[0062] Step 2.2.1: Use the Akmin data interpolation method to interpolate LA and LI respectively to obtain the basic upper envelope line of the ultrasonic signal D and the basic lower envelope represent the local extreme points on the upper and lower edges;

[0063] Step 2.2.2: Use formula (7) to obtain the upper envelope line EU:

[0064]

[0065] Where n is the dimension of the ultrasonic signal, θ, k are envelope line correlation control parameters, summarized from a large amount of engineering experience. When θ = 0.8, k = 2, the effect is better. In particular, when i < k or i > n - k, eu i = eu i d ;

[0066] Step 2.2.3: Use formula (8) to obtain the lower envelope line ED:

[0067]

[0068] Step 2.3: For the upper envelope line EU, lower envelope line ED, and input signal D, in order to obtain the optimal fusion rule between the envelope line feature and the original signal, n independent 3*3 convolution kernels W are designed in this embodiment iFor fusion, the size of the convolution kernel represents the three channels of the feature to be fused, namely the ultrasound signal and its envelope data, which span three sampling times simultaneously. The fused feature vector is F = [f1 f2 …f i … f n The fusion process is expressed as formula (3):

[0069]

[0070] in, W represents an independent convolution kernel i The element in a certain row and column j, especially when i = 0 or n, f i =d i The above improvements allow the fused features at a certain sampling time during the feature fusion process to take into account not only the pre-fusion features at the current sampling time but also the pre-fusion features at the left and right adjacent times, making the fused features more comprehensive.

[0071] Step 2.4: Use a one-dimensional convolutional neural network as the fused feature classifier to classify the sample set S. 2 S 3 S 4 Perform steps 2.1 to 2.2, using the output of step 2.2 to train the n independent convolutional kernels of step 2.3 and the one-dimensional convolutional neural network of step 2.6, ultimately generating the ultrasonic nondestructive testing normal signal classification model F. N ;

[0072] Step 2.5: Input the ultrasonic nondestructive testing dataset G to be tested into the ultrasonic nondestructive testing normal signal classification model F generated in step 2.4. N In the output, three classes of the signal to be detected are: the secondary echo dataset GS. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 It also outputs the probability that each sample represents the current result, i.e., GS. 2 The probability that each signal belongs to the second echo signal is ps. 2 GS 2 The probabilities of all signals belonging to the second echo signal together constitute the probability vector PS. 2 GS 3 The probability that each signal belongs to a third-echo signal is ps. 3 GS 3 The probabilities of all signals belonging to the third echo signal together constitute the probability vector PS. 3 GS 4 The probability that each signal belongs to the fourth echo signal is ps. 4 GS 4The probabilities of all signals belonging to the fourth echo signal together constitute the probability vector PS. 4 ;

[0073] Step 3: Construct an ultrasonic nondestructive testing abnormal signal detection model F A (Second Stage) Anomaly detection is performed on the various classification datasets generated in Step 2. After the ultrasonic nondestructive testing dataset is classified into three categories of normal signals, it still contains a large number of anomalous signals. Therefore, this step is used to detect these anomalous signals mixed in with the normal signals. Specifically:

[0074] Step 3.1: For the secondary echo dataset GS 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Definition: A dataset GS v The signal in H is represented as H = [h1 h2 … h n ], v = 2 or 3 or 4, dataset GS v The corresponding probability is PS v , Figure 4 As shown, a window is set with a width of wide and a starting position of index = 0. The window will move along the time axis of the ultrasound signal with a step size of str.

[0075] Step 3.2: Extract a local sub-signal h of the signal H within the window, which can be expressed as formula (4):

[0076] h = [h] index h index+1 …h index+wide (4)

[0077] In the formula, h index+wide This represents the (index+wide)th data point in signal H;

[0078] Step 3.3: Extract the dataset GS v For each local sub-signal, the extraction process is the same as in step 3.2, forming the dataset GS. v Local sub-signal sample set in For local sub-signal sample set A sub-signal sample, where m is the number of samples;

[0079] Step 3.4: Use the Dynamic Time Warping (DTW) algorithm to extract the local sub-signal h and each local sub-signal sample in the local sub-signal sample set. The similarity is calculated, and the average of the similarities of all sub-signal samples is taken as the local feature similarity ls of the input signal H in the current window. index ;

[0080] Step 3.5: Move str time units, that is, update index to index+str, and repeat steps 3.2 to 3.4 to extract local similarity features at different positions, and finally form the local feature similarity vector LS′ of the input signal H. LS′ can be expressed as formula (11):

[0081] LS′=[ls0′ls1′…ls i ′…ls k ′]k*str+wide<n (5)

[0082] Step 3.6: Utilize probability ps v The local feature similarity of the signal is enhanced, and the enhanced local similarity feature vector is LS = [ls0 ls1 …ls]. i … ls k The enhancement method is shown in Equation (12). This step further enhances the separability of local similarity feature vectors in the embedding space.

[0083] ls i =exp(λ*(1-ps) v ))*ls i ′i∈[0,k] (6)

[0084] Step 3.7: Classify the enhanced local feature similarity using a two-layer linear neural network, using the dataset GS. v and abnormal signal dataset S 1 The training process is performed, and the dataset GS is eventually generated. v Ultrasonic nondestructive testing abnormal signal detection model F corresponding to echo signal v A ;

[0085] Step 3.8: Repeat steps 3.1 to 3.7 to construct the second echo dataset GS respectively. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Ultrasonic nondestructive testing abnormal signal detection model F v A v = 2, 3, 4, used to detect abnormal signals in each dataset.

[0086] For the signal to be detected, the ultrasonic nondestructive testing normal signal classification model generated in step 2 is first executed to classify it as a certain type of normal signal, such as a secondary echo signal, a tertiary echo signal, or a quaternary echo signal, and the probability of the signal being a certain type of normal signal is output. Then, the signal is input into the ultrasonic nondestructive testing abnormal signal detection model generated in step 3 to determine whether it is an abnormal signal. If it is an abnormal signal, an abnormal label is directly output; otherwise, the normal signal label output in step 2 is output.

Claims

1. A two-stage ultrasonic nondestructive testing signal high-precision identification method, characterized in that, include: Step 1: Label the ultrasonic nondestructive testing signals as normal signals and abnormal signals, and construct a dataset of ultrasonic nondestructive testing signals; Step 2: Construct a normal signal classification model for ultrasonic nondestructive testing to classify ultrasonic signals into three categories: secondary echo signals, tertiary echo signals, and quaternary echo signals. Step 3: Construct an ultrasonic nondestructive testing abnormal signal detection model for detecting abnormal signals in the various classification datasets generated in Step 2; Step 2 includes: Step 2.1: Extract ultrasound signals using the quadratic difference method The set of local maxima LA and the set of local minima LI; Step 2.2: Obtain the upper and lower envelopes of the ultrasound signal D; Step 2.3: Apply n independent 3*3 convolution kernels to the upper envelope EU, lower envelope ED, and input signal D. To merge; Step 2.4: Use a one-dimensional convolutional neural network as the fused feature classifier to classify the sample set S. 2 S 3 S 4 Perform steps 2.1 to 2.2, using the output of step 2.2 to train the n independent convolutional kernels of step 2.3 and the one-dimensional convolutional neural network of step 2.4, ultimately generating a normal signal classification model for ultrasonic nondestructive testing. ; Step 2.5: Input the ultrasonic nondestructive testing dataset G to be tested into the ultrasonic nondestructive testing normal signal classification model generated in step 2.

4. In the output, three classes of the signal to be detected are: the secondary echo dataset GS. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 It also outputs the probability of the current result for each sample, GS. 2 The probability that each signal belongs to the second echo signal is ps. 2 GS 2 The probabilities of all signals belonging to the second echo signal together constitute the probability vector PS. 2 GS 3 The probability that each signal belongs to a third-echo signal is ps. 3 GS 3 The probabilities of all signals belonging to the third echo signal together constitute the probability vector PS. 3 GS 4 The probability that each signal belongs to the fourth echo signal is ps. 4 GS 4 The probabilities of all signals belonging to the fourth echo signal together constitute the probability vector PS. 4 ; Step 3 includes: Step 3.1: For the secondary echo dataset GS 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Define a dataset GS v The signal in is represented as v=2 or 3 or 4, dataset GS v The corresponding probability is PS v Set a window with a width of wide and a starting position of index=0. The window will move along the time axis of the ultrasound signal with a step size of str. Step 3.2: Extract a local sub-signal h from the signal H within the window: Step 3.3: Extract the dataset GS v Each local sub-signal in the dataset GS constitutes the dataset GS. v Local sub-signal sample set ,in For local sub-signal sample set A sub-signal sample, where m is the number of samples; Step 3.4: Use the Dynamic Time Warping (DTW) algorithm to extract the local sub-signal h and each local sub-signal sample in the local sub-signal sample set. The similarity is calculated, and the average of the similarities of all sub-signal samples is taken as the local feature similarity ls of the input signal H in the current window. index ; Step 3.5: Shift the time by str units, update the index to index + str, and repeat steps 3.2 to 3.4 to extract local similarity features at different locations, ultimately forming the local feature similarity vector of the input signal H. ; ； Step 3.6: Utilize probability ps v The local feature similarity of the signal is enhanced, and the enhanced local similarity feature vector is: ; Step 3.7: Classify the enhanced local feature similarity using a two-layer linear neural network, using the dataset GS. v and abnormal signal dataset S 1 The training process is performed, and the dataset GS is eventually generated. v Ultrasonic nondestructive testing abnormal signal detection model F corresponding to echo signal v A ; Step 3.8: Repeat steps 3.1 to 3.7 to construct the second echo dataset GS respectively. 2 Three-echo dataset GS 3 Four-echo dataset GS 4 Ultrasonic nondestructive testing abnormal signal detection model v=2,3,4, used to detect abnormal signals in each dataset.

2. The method for high-precision identification of two-stage ultrasonic nondestructive testing signals according to claim 1, characterized in that, Step 1 is specifically described as follows: Construct a secondary echo sample set S based on the secondary echo signal in the normal signal. 2 Construct a three-echo signal sample set S based on the three echo signals in the normal signal. 3 Construct a four-echo signal sample set S based on the four echo signals in the normal signal. 4 ; An ultrasound abnormal signal dataset S is constructed based on the abnormal signals. 1 .

3. The high-precision identification method for two-stage ultrasonic nondestructive testing signals according to claim 1, characterized in that, Step 2.2 includes: Step 2.2.1: Use the Akmin data interpolation method to interpolate LA and LI respectively to obtain the envelope of the ultrasound signal D. and basic envelope ; , Indicates local extreme points on the upper and lower edges; Step 2.2.2: Obtain the upper envelope EU; (1)； Where n is the dimension of the ultrasound signal, This is the envelope correlation control parameter; when i < k or i > nk, ; Step 2.2.3: Obtain the lower envelope ED; (2)； When i < k or i > nk, .

4. The method for high-precision identification of two-stage ultrasonic nondestructive testing signals according to claim 1, characterized in that, The fusion process in step 2.3 is represented as follows: (3)； in, Independent convolution kernels The element in row 0, column j, especially when i = 0 or n. .

5. The high-precision identification method for two-stage ultrasonic nondestructive testing signals according to claim 1, characterized in that, The enhancement method in step 3.6 is expressed as follows: ； In the formula, Similarity feature vector enhancement coefficient.

Images