CN112014478A - Defect echo blind extraction self-adaptive method submerged in ultrasonic grass-shaped signal - Google Patents

Abstract

In order to solve the problem that the existing defect echo extraction algorithm cannot suppress actual noise so as not to obtain a defect signal, the invention provides a defect echo blind extraction self-adaptive method submerged in an ultrasonic grass-shaped signal. The method comprises the following specific steps: firstly, carrying out similarity analysis on an acquired original signal sample by using an unsupervised machine learning algorithm; then, inputting the similar signal samples into a designed noise reduction self-encoder to train the noise reduction self-encoder, and enabling the noise reduction self-encoder to learn corresponding noise reduction rules; and finally, automatically reducing noise of the signal by using the trained self-encoder, and realizing intelligent extraction of the defect signal. Compared with the existing defect signal extraction technology, the method has stronger adaptability to the inhibition of actual noise, and the strong adaptability is derived from the capability of learning the characteristics of the defect signal and the noise signal respectively by analyzing the rich information provided by the original signal sample.

Description

An adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals

technical field

The invention relates to the technical field of ultrasonic non-destructive testing, in particular to an adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals.

Background technique

Blind extraction of target signals from raw data is a research hotspot in the field of modern signal processing. The so-called blind extraction, that is, without any reference data, nor with any prior information about the characteristics of the target signal, spectrum distribution, etc., only the data collected by the sensor is used as the analysis basis to accurately extract the target signal. Because of this, the blind extraction of signals has become an important technology in many high-tech fields such as speech enhancement, image processing, medical diagnosis, and military reconnaissance.

In the field of ultrasonic non-destructive testing, with the development of automatic/intelligent testing, it has very important application value to solve the blind extraction of defect echoes submerged in ultrasonic grass-like signals. Defect echo is the target signal in ultrasonic nondestructive testing. It is important information for analyzing the size, location, distribution and properties of defects, and also an important basis for judging the quality and safety of workpieces. However, due to the coarse material of the detected material, the uneven distribution of structure and grains, the large difference in the surface roughness of the workpiece, etc., the ultrasonic wave is scattered during the propagation process, resulting in a large number of amplitudes higher than the detection sensitivity (the minimum defect size found). The grass-like noise of the value; if the defect signal is extracted by the time-domain amplitude, it is undoubtedly a blind person touching the image, resulting in misjudgment or missed detection. The more thorny problem is that the grass-like echoes have many and complicated shapes; different grain sizes produce grass-like noises of different amplitude levels, different grain types (equiaxed, columnar), different grains. The distribution also produces grass-like noise with different morphologies. The most typical example is the inspection of coarse-grained austenitic stainless steel welds, which integrates all the above-mentioned characteristics. Various forms of grass noise (Figure 1).

Laser ultrasound is a non-contact, high-sensitivity, non-destructive testing method. It has become an important method for finding micron-scale defects (unfused powder, pores, etc.) inside the workpiece near the surface. It is expected to become the most environmentally friendly and cost-effective in the field of additive manufacturing. Real-time monitoring nondestructive testing technology. However, laser-generated ultrasonic surface waves are highly susceptible to interference from material surface roughness. Unfortunately, additive manufacturing, due to its unique layer-by-layer manufacturing approach, inevitably produces random surface roughnesses that are often comparable in size to the micron-scale defects (unfused powders, pores, etc.) produced during fabrication. Therefore, when LUT monitors it in real time, defect echoes are buried by various grass-like noises caused by roughness (Figure 2). In addition, based on the requirement of real-time monitoring, grinding the surface is not feasible. In summary, it is urgent to develop an effective extraction algorithm for defect echoes in ultrasonic grass-like signals.

In the prior art, there are some defect signal extraction algorithms, such as Fourier transform (FT), short-time Fourier transform (STFT), wavelet transform, split spectrum analysis (SSP), sparse decomposition, full empirical mode decomposition (EEMD), Variational modal decomposition, etc. The noise reduction process of these algorithms relies on a prior model (noise is some mathematically distributed or high frequency component). Therefore, they are widely used in suppressing high-frequency or grass-like noise close to Gaussian distribution (eg Manjula K, Vijayarekha K, Venkatraman B. Quality Enhancement of Ultrasonic TOFD Signals from Carbon Steel Weld Pad with Notches [J]. Ultrasonics, 2017, 84:264-271.).

In the process of implementing the present invention, the inventor of the present application found that the method of the prior art has at least the following technical problems:

The above algorithm partially solves some defect signal extraction in noise background with specific distribution. However, these methods cannot adapt to various noise environments and fail to acquire defect signals.

SUMMARY OF THE INVENTION

The present invention proposes an adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals, which is used to solve or at least partially solve the technical problem of poor adaptability in the prior art methods.

In order to solve the above-mentioned technical problems, the present invention provides an adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals, including:

S1: Use an unsupervised machine learning algorithm to perform similarity analysis on the collected original signal samples, and classify similar signals into one category, the unsupervised machine learning algorithm is k-means or DBSCAN, and the original signal samples are at the same Under the experimental conditions, the ultrasonic A-scan signal, B-scan signal, and C-scan signal collected for the same test block;

S2: Input the similar signal obtained in S1 into the pre-designed noise reduction auto-encoder for training, so that it can learn the corresponding noise-reduction rules. The pre-designed noise-reduction self-encoder is a A three-layer neural network with input equal to output, in the form: n-h-n, where n represents the number of neurons in the input and output, and h represents the number of neurons in the hidden layer;

S3: Use the trained denoising autoencoder to denoise the signal to be processed.

In one embodiment, S1 specifically includes:

S1.1: Determine the number of clusters for signal clustering;

S1.2: Select signal similarity evaluation criteria;

S1.2: Perform similarity analysis on the collected original signal samples according to the number of clusters of signal clustering and the signal similarity evaluation criterion.

In one embodiment, S1.1 specifically includes:

The number of clusters is determined according to the number of signals or the number of target micro-areas. When the number of signals is used as the basis, one tenth of the number of collected signals is set as the number of clusters; when the number of target micro-areas is used as the number, the target micro-area is set as the number of clusters The number of regions as the number of clusters.

In one embodiment, S1.2 specifically includes:

Taking the Euclidean distance as the evaluation criterion of signal similarity, its expression is:

Among them, S ₁ and S ₂ represent two different signals respectively, and s _1i and s _2i are their corresponding sampling points respectively. The smaller the Euclidean distance between the signals, the higher the similarity between them.

In one embodiment, S2 specifically includes:

S2.1: Preprocess the original data, where the standardized formula for preprocessing is:

Among them, X represents the original data, minX represents the minimum value of X, maxX represents the maximum value of X, x represents a data point in the original data set, and x' represents the standardized data point;

S2.2: Configure the noise reduction autoencoder,

The network structure is an encoding-decoding three-layer network structure n-h-n. The number of neurons in the input layer n and the number of neurons in the output layer are determined according to the number of sampling points of the signal, and the number of neurons in the hidden layer h is determined according to the final training result. The activation function of the encoder is set as a modified linear unit of Relu(x)=max(0,x), and all training variables in the neural network adopt a sliding average;

S2.3: Train the noise reduction autoencoder, and use a regularization loss function that makes the network generalization ability stronger in the process of training the network, MSEreg, defined as:

Among them, w _j is the weight, γ is a hyperparameter set manually, n is the number of weights, and MSE is the mean square error, which is defined as:

为目标值，y为降噪自编码器基于输入的训练数据的预测值，设定初始值为0.01的衰减学习率。where N is the number of training samples,

is the target value, y is the predicted value of the denoising autoencoder based on the input training data, and the decay learning rate is set to an initial value of 0.01.

The above-mentioned one or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:

The present invention no longer uses the traditional prior noise model, but uses the structural characteristics of the signal data itself to summarize the regular characteristics of noise and defect signals, and then acquire the corresponding noise reduction rules. That is, for a group of signals containing unknown noise, the method of the present invention first analyzes and summarizes the data characteristics thereof, and then forms corresponding noise reduction rules according to the analysis results. Therefore, the present invention has stronger adaptability for suppressing the actual unknown noise, which is derived from its ability to learn the features of defect signals and noise signals respectively by analyzing the rich information provided by the original signal samples. The present invention can be used for but not limited to noise suppression caused by coarse grains, rough surfaces, etc., such as detection of austenitic stainless steel welds, laser ultrasonic real-time monitoring of additively manufactured parts, etc., to achieve intelligent detection of defect signals submerged by noise. extract.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are For some embodiments of the present invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Figure 1 shows different grass noises caused by different grain sizes in a specific embodiment.

FIG. 2 shows different grass-like noises caused by different roughnesses in a specific embodiment.

FIG. 3 is a flat weld of austenitic stainless steel with embedded transverse hole defects in a specific embodiment.

FIG. 4 is the result of the noise reduction of the signal at the horizontal hole in the specific embodiment.

FIG. 5 is a stainless steel sheet manufactured by selective laser melting (SLM) and its internal defect distribution in a specific embodiment.

FIG. 6 is a laser ultrasonic scanning form in a specific embodiment.

FIG. 7 is the noise reduction result of the laser ultrasonic A-scan signal in the specific embodiment.

FIG. 8 is a noise reduction result of laser ultrasonic B-scan in a specific embodiment.

FIG. 9 is a working flow chart of the blind extraction adaptive method provided by the present invention.

Detailed ways

The inventors of the present application have found through extensive research and practice that the methods in the prior art partially solve some defect signal extraction in the background of noise with a specific distribution. However, as described in the background art, the actual noise is complex and varied, and the signal extraction algorithm we need more is blind extraction and adaptive, that is, without using any prior knowledge, only by The data law or characteristic analysis of ultrasonic signals can adapt to any grass-like noise environment of different types and intensities, and realize intelligent extraction of defect signals. Therefore, based on this requirement, the present invention proposes an adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals. The algorithm can realize intelligent extraction of defect signals under the background of grass-like noise caused by different grain sizes, different grain types, different grain distributions, and different roughnesses.

Specifically, in order to solve the problem that the existing extraction algorithms in ultrasonic non-destructive testing cannot adapt to various noise environments and cannot obtain defect signals, the present invention proposes a signal reconstruction method based on statistical reasoning and machine learning. The method can automatically learn the characteristics of defect signals and noise signals by analyzing the rich information provided by the collected original signal samples, thereby realizing effective suppression of noise and intelligent extraction of defect signals.

In order to achieve above-mentioned technical effect, the main idea of the present invention is as follows:

First, the unsupervised machine learning algorithm is used to analyze the similarity of the collected original signal samples; then, the similar signal samples are input into the designed noise reduction autoencoder for training, so that it can learn the corresponding noise reduction. Finally, the trained autoencoder is used to automatically denoise the signal to achieve intelligent extraction of defect signals. Compared with the existing defect signal extraction technology, the present invention has stronger adaptability to the suppression of actual noise, and this strong adaptability stems from the fact that it can learn the defect signal and Ability to characterize noisy signals.

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The embodiment of the present invention provides an adaptive method for blind extraction of defect echoes submerged in ultrasonic grass-like signals, including:

S1: Use an unsupervised machine learning algorithm to perform similarity analysis on the collected original signal samples, and classify similar signals into one category, the unsupervised machine learning algorithm is k-means or DBSCAN, and the original signal samples are at the same Under the experimental conditions, the ultrasonic A-scan signal, B-scan signal, and C-scan signal collected for the same test block;

S2: Input the similar signal obtained in S1 into the pre-designed noise reduction auto-encoder for training, so that it can learn the corresponding noise-reduction rules. The pre-designed noise-reduction self-encoder is a A three-layer neural network with input equal to output, in the form: n-h-n, where n represents the number of neurons in the input and output, and h represents the number of neurons in the hidden layer;

S3: Use the trained denoising autoencoder to denoise the signal to be processed.

Specifically, similar signals can be classified into one category through step S1, which can be used as training data. Step S2 is to use the obtained similar signals to learn and learn the corresponding noise reduction rules. The trained noise reduction autoencoder can be obtained through training, so that the noise reduction processing of the signal can be realized.

In one embodiment, S1 specifically includes:

S1.1: Determine the number of clusters for signal clustering;

S1.2: Select signal similarity evaluation criteria;

S1.2: Perform similarity analysis on the collected original signal samples according to the number of clusters of signal clustering and the signal similarity evaluation criterion.

In one embodiment, S1.1 specifically includes:

The number of clusters is determined according to the number of signals or the number of target micro-areas. When the number of signals is used as the basis, one tenth of the number of collected signals is set as the number of clusters; when the number of target micro-areas is used as the number, the target micro-area is set as the number of clusters The number of regions as the number of clusters.

In one embodiment, S1.2 specifically includes:

Taking the Euclidean distance as the evaluation criterion of signal similarity, its expression is:

Among them, S ₁ and S ₂ represent two different signals respectively, and s _1i and s _2i are their corresponding sampling points respectively. The smaller the Euclidean distance between the signals, the higher the similarity between them.

In one embodiment, S2 specifically includes:

S2.1: Preprocess the original data, where the standardized formula for preprocessing is:

Among them, X represents the original data, minX represents the minimum value of X, maxX represents the maximum value of X, x represents a data point in the original data set, and x' represents the standardized data point;

S2.2: Configure the noise reduction autoencoder,

The network structure is an encoding-decoding three-layer network structure n-h-n. The number of neurons in the input layer n and the number of neurons in the output layer are determined according to the number of sampling points of the signal, and the number of neurons in the hidden layer h is determined according to the final training result. The activation function of the encoder is set as a modified linear unit of Relu(x)=max(0,x), and all training variables in the neural network adopt a sliding average;

S2.3: Train the noise reduction autoencoder, and use a regularization loss function that makes the network generalization ability stronger in the process of training the network, MSEreg, defined as:

Among them, w _j is the weight, γ is a hyperparameter set manually, n is the number of weights, and MSE is the mean square error, which is defined as:

为目标值，y为降噪自编码器基于输入的训练数据的预测值，设定初始值为0.01的衰减学习率。where N is the number of training samples,

is the target value, y is the predicted value of the denoising autoencoder based on the input training data, and the decay learning rate is set to an initial value of 0.01.

Specifically, through the normalization process, it can prevent the gradient from exploding or disappearing during the network training process, and the sliding average, that is, averages all the values in the process of variable change, so that the screenshot makes the encoder have good robustness.

Please refer to FIG. 7 to FIG. 9 , wherein FIG. 7 is the noise reduction result of the laser ultrasound A-scan signal in the specific embodiment, wherein, the larger fluctuation is the original signal, and the smaller one is the denoised signal, and FIG. 8 is the specific example. The noise reduction result of the laser ultrasound B-scan in the embodiment, FIG. 9 is a working flow chart of the blind extraction adaptive method provided by the present invention, wherein the establishment of the signal database is to construct the collected original signals as a database.

After training through the original signal samples, the autoencoder learns the corresponding noise reduction rules, and the original signal can automatically filter out the noise after the trained autoencoder. It is worth noting that the noise reduction rules are obviously not determined a priori, but are determined by the experimental data itself, and the entire noise reduction process is completely driven by data. Therefore, compared with the existing noise reduction technology, the present invention has stronger adaptability to the actual noise suppression, and this strong adaptability stems from the fact that it can learn the defect signals separately by analyzing the rich information provided by the original signal samples and the ability to characterize noisy signals.

The method provided by the present invention no longer uses the traditional prior noise model, but uses the structural characteristics of the signal data itself to summarize the regular characteristics of noise and defect signals, and then acquire the corresponding noise reduction rules. That is, for a group of signals containing unknown noise, the present invention first analyzes and summarizes the data characteristics thereof, and then forms corresponding noise reduction rules according to the analysis results. Therefore, the present invention has stronger adaptability for suppressing the actual unknown noise, which is derived from its ability to learn the features of defect signals and noise signals respectively by analyzing the rich information provided by the original signal samples. The present invention can be used for but not limited to noise suppression caused by coarse grains, rough surfaces, etc., such as detection of austenitic stainless steel welds, laser ultrasonic real-time monitoring of additively manufactured parts, etc., to achieve intelligent detection of defect signals submerged by noise. extract.

The embodiments of the present invention will be described in detail below by way of examples in conjunction with specific examples.

Embodiment 1: This embodiment takes the austenitic stainless steel weld as the detection object.

Step 1: Build a signal database. The tested block is a butt weld of austenitic stainless steel flat plate, and the grain distribution along the weld direction is extremely uneven, and the grain size fluctuates within 60-300 μm. A 5MHz straight probe was used to scan along the length of the weld just above the weld (Figure 1). The step size was 1mm, the scanning distance was 130mm, the system gain was set to 60dB, the sampling frequency was 100MHz, and the sampling depth was 1000. Finally, a signal matrix of 130 × 1000 is obtained.

Step 2: Signal similarity analysis. The k-means clustering algorithm is used to analyze the similarity of the 1-610 sampling points of the A-scan signal, the number of clusters is set to 13, and the similarity criterion is set to the Euclidean distance.

Step 3: Network training. The 13 clusters of signals are respectively input into the designed autoencoder for training.

Step 4: Input the original signal into the trained autoencoder for automatic noise filtering. Figure 2 shows the noise reduction signal at the transverse hole. It can be seen from the original signal that the hole signal has been completely submerged by noise, and after noise reduction, the hole signal is well revealed.

Specific embodiment 2: This embodiment takes the stainless steel additive manufacturing test block as the detection object.

Step 1: Build a signal database. The tested block is a rectangular stainless steel sheet produced by selective laser melting process (SLM), with a thickness of 5 mm and an average surface roughness of 75 μm. Six groove defects with different burial depths were processed inside the test block, and the specific layout and related parameters are shown in Figure 3. A 2MHz laser ultrasound was used for grid scanning (Fig. 4), the scanning range was 20mm×25mm, covering 6 defects, the step size was 0.1mm, and an A-scan signal was recorded for each grid point, and the sampling depth was 2000. Finally, a three-dimensional signal matrix of 200×250×2000 is obtained.

Step 2: Signal similarity analysis. The k-means clustering algorithm was used to analyze the similarity of 1-500 sampling points of the A-scan signal, the number of clusters was set to 3000, and the similarity criterion was set to Euclidean distance.

Step 3: Network training. The 3000 clusters of signals were input into the designed autoencoder for training.

Step 4: Input the original signal into the trained autoencoder for automatic noise filtering. Figure 5 shows the result of A-scan noise reduction. It can be seen that the surface wave is well displayed after noise reduction. Figure 6 shows the B-scan noise reduction result formed by a certain sequence of A-scan combinations, and it can be seen that the clarity of the image has been greatly improved.

The specific implementation examples described in the present invention are merely illustrative of the methods and steps of the present invention. Those skilled in the technical field of the present invention can make corresponding modifications, additions or variations to the specific implementation steps described (ie, use similar alternatives), but will not deviate from the principle and essence of the present invention or go beyond the appended claims the scope defined by the book. The scope of the present invention is limited only by the appended claims.

Claims (5)

1. a defect echo blind extraction adaptive method submerged in the ultrasonic grass-like signal is characterized in that, comprising: S1: Use an unsupervised machine learning algorithm to perform similarity analysis on the collected original signal samples, and classify similar signals into one category, the unsupervised machine learning algorithm is k-means or DBSCAN, and the original signal samples are at the same Under the experimental conditions, the ultrasonic A-scan signal, B-scan signal, and C-scan signal collected for the same test block; S2: Input the similar signal obtained in S1 into the pre-designed noise reduction auto-encoder for training, so that it can learn the corresponding noise-reduction rules. The pre-designed noise-reduction self-encoder is a A three-layer neural network with input equal to output, in the form: n-h-n, where n represents the number of neurons in the input and output, and h represents the number of neurons in the hidden layer; S3: Use the trained denoising autoencoder to denoise the signal to be processed. 2. The adaptive method for blind extraction of defect echoes according to claim 1, wherein S1 specifically comprises: S1.1: Determine the number of clusters for signal clustering; S1.2: Select signal similarity evaluation criteria; S1.2: Perform similarity analysis on the collected original signal samples according to the number of clusters of signal clustering and the signal similarity evaluation criterion. 3. The adaptive method for blind extraction of defect echoes according to claim 1, wherein S1.1 specifically comprises: The number of clusters is determined according to the number of signals or the number of target micro-areas. When the number of signals is used as the basis, one tenth of the number of collected signals is set as the number of clusters; when the number of target micro-areas is used as the number, the target micro-area is set as the number of clusters The number of regions as the number of clusters. 4. The adaptive method for blind extraction of defect echoes according to claim 1, wherein S1.2 specifically comprises: Taking the Euclidean distance as the evaluation criterion of signal similarity, its expression is:

Among them, S ₁ and S ₂ represent two different signals respectively, and s _1i and s _2i are their corresponding sampling points respectively. The smaller the Euclidean distance between the signals, the higher the similarity between them. 5. The adaptive method for blind extraction of defect echoes according to claim 1, wherein S2 specifically comprises: S2.1: Preprocess the original data, where the standardized formula for preprocessing is:

Among them, X represents the original data, min X represents the minimum value of X, max X represents the maximum value of X, x represents a data point in the original data set, and x' represents the standardized data point; S2.2: Configure the noise reduction autoencoder, The network structure is an encoding-decoding three-layer network structure n-h-n. The number of neurons in the input layer n and the number of neurons in the output layer are determined according to the number of sampling points of the signal, and the number of neurons in the hidden layer h is determined according to the final training result. The activation function of the encoder is set as a modified linear unit of Relu(x)=max(0,x), and all training variables in the neural network adopt a sliding average; S2.3: Train the noise reduction autoencoder, and use a regularization loss function that makes the network generalization ability stronger in the process of training the network, MSEreg, defined as:

Among them, w _j is the weight, γ is a hyperparameter set manually, n is the number of weights, and MSE is the mean square error, which is defined as:

where N is the number of training samples,

is the target value, y is the predicted value of the denoising autoencoder based on the input training data, and the decay learning rate is set to an initial value of 0.01.

Images