CN118130608B - An online detection method for underwater structure welds based on improved deep convolution - Google Patents

Abstract

The invention relates to the technical field of underwater welding detection, in particular to an online detection method of an underwater structure welding seam based on improved deep convolution, which comprises the steps of S1, using a hydrophone sensor to receive an acoustic signal of marine underwater structure welding, S2, using a VMD algorithm to denoise the acoustic signal and extract a weak welding acoustic signal from a high-order IMF component, S3, using a recursion graph to perform time-space transformation on an effective pulse, transferring a one-dimensional signal from a time domain to a phase space domain to provide more potential welding seam state characteristics, S4, further generating a gradient graph from the recursion graph, using the gradient graph to perform characteristic extraction, extracting the characteristics of three directions of an image X, Y and an XY, and S5, using a VGG16 deep convolution network to classify an image sample to finish online judgment of welding seam quality. The invention can realize the on-line detection of the quality of the underwater welding seam, adjusts the welding quality in real time and provides a certain thought for the on-line detection of the quality of the underwater welding seam.

Description

Improved depth convolution-based on-line detection method for welding seam of underwater structure

Technical Field

The invention relates to the technical field of underwater welding detection, in particular to an online detection method for an underwater structure welding seam based on improved deep convolution.

Background

The underwater welding technology plays a key role in the fields of offshore wind power, drilling platforms, ships, submarine pipelines and the like. At present, defects in the welding line can seriously affect the reliability of the structural part. In order to repair as soon as possible and reduce the loss, repair by adopting an underwater welding technology is a necessary choice. Therefore, it is important to develop reliable on-line weld quality detection technology.

Therefore, scientists at home and abroad propose various detection modes such as vision, optics, acoustics and the like, but most of them have limitations. Visual detection is easy to be interfered by the outside, such as factors of light, temperature, humidity and the like, the optical detection precision can be influenced under the conditions of complex welding defects and surface defects, and the optical detection cost is higher. In contrast, acoustic-based non-destructive inspection techniques have many advantages, such as a wide inspection range, a large inspection depth, etc., and are more suitable for inspecting internal defects of welded structures. However, the conventional acoustic signal detection technology is also limited in being applied to the on-line detection of the quality of the welded seam of underwater welding. The first problem is that the signal-to-noise ratio of the acoustic signal is low, effective pulse extraction is difficult, the second problem is that the complex acoustic signal features are difficult to extract, the traditional time domain and frequency domain analysis is not applicable, and the third problem is that the quality discrimination accuracy is low under the conditions of poor quality and unbalanced category of the data set. Therefore, the invention provides an online detection method for the welding line of the underwater structure based on improved deep convolution.

Disclosure of Invention

The invention aims to solve the defects in the prior art, and provides an on-line detection method for the welding seam of the underwater structure based on improved deep convolution, which can realize higher quality recognition rate and provide a certain thought for on-line recognition of the quality of the welding seam of the underwater structure.

In order to achieve the above purpose, the present invention adopts the following technical scheme:

An online detection method of an underwater structure welding seam based on improved depth convolution comprises the following steps:

S1, receiving acoustic signals of marine underwater structure welding by using a hydrophone sensor;

s2, denoising the acoustic signals by using a VMD algorithm, and extracting weak welding acoustic signals from high-order IMF components;

s3, performing time-space domain transformation on the effective pulse by using a recursion diagram, and transferring one-dimensional signals from a time domain to a phase space domain to provide more potential weld state characteristics;

S4, further generating a gradient map from the recursion map, and extracting features in three directions of the image X, Y and the XY by using the gradient map to perform feature extraction;

s5, classifying the image samples by using a VGG16 deep convolution network, and completing online judgment of the weld quality.

Preferably, in step S2, a hydrophone is used to collect acoustic signals of underwater welding, the acoustic signals are noise-reduced by a VMD algorithm, the acoustic signals are decomposed into a plurality of intrinsic mode IMF components, effective acoustic signal pulses are extracted, and the VMD algorithm operates as follows:

Where { u _k}:＝{u₁,...u_k } and { ω _k}:＝{ω₁,...ω_k } are all modal sets and their frequency centers, K represents the number of IMF decomposition levels, Representing a partial derivative operation, phi (t) representing a unit pulse function, j being an imaginary unit in the hilbert transform, t representing time, x representing convolution, f (t) representing a signal to be decomposed, and being a set of data having a time sequence;

introducing penalty factor ρ and Lagrange multiplier γ converts constraint into non-constraint variation problem, and augmenting Lagrange expression as follows:

The problem is solved iteratively by adopting a multiplication operator alternating algorithm, and the optimal solution formula of the VMD is obtained by continuously and iteratively updating u _k、ω_k and gamma as follows:

Where F (ω), γ (ω), and U _k (ω) are fourier transform signals of F (t), γ (t), and U _k (t), respectively, n is the number of iterations, and τ is the noise margin coefficient.

Preferably, in step S3, the key step of the recursion map is to perform phase space reconstruction, and the recursion map is used to convert the time series data of the effective pulse of the underwater welding sound signal into an image form, so as to reveal the internal structure of the time series, and specifically includes the following steps:

S31: for a time sequence signal u _k (k=1, 2,.. The sampling time interval is determined to be deltat, and a proper embedding dimension m and delay time tau are determined through correlation theoretical calculation, so that the time sequence is reconstructed, and a reconstructed power system is represented by x _i＝[u_i,u_i+τ,...,u_i+(m-1)τ, wherein i=1, 2,.. N- (m-1) tau;

S32, calculating the distance S _ij between the i point x _i and the j point x _j in the reconstructed phase space:

S_ij＝||x_i-x_j||

where i=1, 2,..n- (m-1) τ, j=1, 2,..n- (m-1) τ, |·|| represents the norm;

S33, calculating a recursion value:

R(i,j)=θ(ε_i-S_ij)

Wherein, θ (·) represents a Heaviside function, when the value in the bracket is greater than or equal to 0, θ (·) is 1, conversely, when the value in the bracket is less than 0, θ (·) is 0, ε represents a predetermined critical distance, and a fixed value or a variable value can be obtained.

Preferably, in step S4, the recursive graph is further converted into three gradient graphs, namely an X-direction gradient graph, a Y-direction gradient graph and an XY-direction gradient graph, wherein the change of the X axis is that the pixel value on the right side (X+1) of the current pixel is subtracted from the pixel value on the left side (X-1) of the current pixel, the change of the Y axis is that the pixel value below (Y+1) of the current pixel is subtracted from the pixel value above (Y-1) of the current pixel, a two-dimensional vector is formed after the two components are calculated, and the image gradient of the pixel is obtained, and the gradient expression of an image function f (X, Y) is as follows:

Wherein G _x, Represents the gradient component in the X direction, G _y,Representing the Y-direction gradient component, T representing the transpose.

The amplitude function mag (f) is denoted as g (x, y):

Direction angle function φ (x, y):

For digital images, corresponding to a two-dimensional discrete function gradient, the derivative is approximated using a difference:

G_x(x,y)=H(x+1,y)-H(x-1,y)

G_y(x,y)=H(x,y+1)-H(x,y-1)

In the formula, G _x (X, Y) represents a gradient value of the pixel point in the X direction, G _y (X, Y) represents a gradient value of the pixel point in the Y direction, and H (X, Y) represents a gray value of the pixel point.

Thus, the gradient value G (x, y) and the gradient direction α (x, y) at the pixel point (x, y) are respectively:

Preferably, in step S5, the VGG16 network is trained through a back propagation algorithm, a gradient descent method is used for optimizing network parameters, the VGG16 convolutional neural network structure comprises 13 convolutional layers and 3 full-connection layers, a deeper network structure, smaller convolutional kernels and a pooling sampling domain, abstract features can be better learned, all the convolutional layers use the same convolutional kernels, step sizes and filling, larger receptive fields can be simulated through stacking a plurality of small convolutional kernels, parameters are reduced at the same time, each convolutional layer adopts the convolutional kernels for feature extraction, the largest pooling layer is used for sampling a feature map, the full-connection layers are used for mapping and classifying features, and the prediction probability of each category is output;

the method specifically comprises the following steps:

s51, after the input image is subjected to one-time average value reduction pretreatment, the input image enters two convolution layers, and each convolution layer uses 64 convolution kernels with the size of 3 multiplied by 3 to extract features;

s52, downsampling the feature map through a2×2 maximum pooling layer;

S53, two convolution blocks, wherein each block comprises two convolution layers, each convolution layer uses 128 convolution kernels of 3 multiplied by 3 to extract features, and a2 multiplied by 2 max pooling layer is arranged behind each convolution block to further reduce the size of the feature map;

S54, three convolution blocks, each block comprising three convolution layers, each convolution layer performing feature extraction using 256, 512 and 512 3 x 3 convolution kernels;

and S55, mapping and classifying the features through three full connection layers.

Compared with the prior art, the invention has the following beneficial effects:

1. The invention displays weak welding acoustic signals from a phase space angle, firstly decomposes the acoustic signals into a plurality of IMF components through a VMD algorithm, and then converts a time sequence into a phase space by using a recurrence chart.

2. According to the invention, the gradient characteristics are extracted from the recursion image, so that the detection rate of the convolutional neural network is improved, the VGG 16 neural network can automatically learn and identify the effective gradient image characteristics, and the method can realize high-efficiency online weld quality classification by comparing the gradient image characteristics with classification results of three kinds of VGG19, RESNET, RESNET neural networks.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a schematic diagram of a welded steel plate experimental sample according to an embodiment of the present invention;

FIG. 3 is a diagram of an original audio signal according to an embodiment of the present invention;

FIG. 4 is a diagram of IMF components of an audio signal after VMD noise reduction in accordance with an embodiment of the present invention;

FIG. 5 is a recursive diagram of an embodiment of the present invention;

Fig. 6 is a gradient map of an embodiment of the present invention.

Detailed Description

The following technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, and thus the protection scope of the present invention is more clearly defined. The described embodiments of the present invention are intended to be only a few, but not all embodiments of the present invention, and all other embodiments that may be made by one of ordinary skill in the art without inventive faculty are intended to be within the scope of the present invention.

Referring to fig. 1, an on-line detection method for an underwater structure weld based on improved deep convolution includes the following steps:

Step 1, receiving acoustic signals of marine underwater structure welding by using a hydrophone sensor;

step 2, denoising the acoustic signals by using a VMD algorithm, and extracting weak welding acoustic signals from high-order IMF components;

Step 3, performing time-space domain transformation on the effective pulse by using a recursion diagram, and transferring one-dimensional signals from a time domain to a phase space domain to provide more potential weld state characteristics;

Step 4, further generating a gradient map from the recursion map, and extracting features in three directions of the image X, Y and the XY by using the gradient map to perform feature extraction;

And 5, classifying the image samples by using a VGG16 deep convolution network to finish the online judgment of the weld quality.

Referring to fig. 1 to 6, the implementation steps of the technical scheme provided by the invention are as follows:

And 2, as shown in fig. 2, after an experimental platform is built, welding the steel plates, and classifying the qualified steel plates and the unqualified steel plates. As shown in fig. 3, after the hydrophone is used to collect acoustic signals of underwater welding, the original audio signals are collected. As shown in fig. 4, the noise reduction process is performed using a VMD algorithm that decomposes the acoustic signal into 3 eigenmode components. Wherein the IMF decomposition layer number K is set to 3 layers, the band limit alpha is set to 3000, the noise holding coefficient tau is set to 0, the direct current component DC is set to 0, and the effective acoustic signal pulse is extracted. The operating principle of the VMD algorithm is as follows:

Where { u _k}:＝{u₁,...u_k } and { ω _k}:＝{ω₁,...ω_k } are all modal sets and their frequency centers, K represents the number of IMF decomposition levels, Representing a partial derivative operation, phi (t) representing a unit pulse function, j being an imaginary unit in the hilbert transform, t representing time, x representing convolution, f (t) representing a signal to be decomposed, and being a set of data having a time sequence;

introducing penalty factor ρ and Lagrange multiplier γ converts constraint into non-constraint variation problem, and augmenting Lagrange expression as follows:

The problem is solved iteratively by adopting a multiplication operator alternating algorithm, and the optimal solution formula of the VMD is obtained by continuously and iteratively updating u _k、ω_k and gamma as follows:

Where F (ω), γ (ω), and U _k (ω) are fourier transform signals of F (t), γ (t), and U _k (t), respectively, n is the number of iterations, and τ is the noise margin coefficient.

In this embodiment, the VMD algorithm may decompose the nonstationary signal under the condition of low signal-to-noise ratio, extract the effective pulse, completely separate the pulse from the signal, collect the obtained acoustic signal in the complex environment, and has good noise reduction processing effect and high signal-to-noise ratio of the acoustic signal after noise reduction.

Step 3, the key step of the recursion diagram is to reconstruct the phase space. A recursive diagram of the generation of the active pulses is shown in fig. 5, which converts the time series of data into image form. The method comprises the following specific steps:

3-1), for a time sequence (signal) u _k (k=1, 2..once., n), determining that its sampling time interval is Δt, determining a suitable embedding dimension m and delay time τ through correlation theoretical calculation, and then reconstructing the time sequence, where the power system after reconstruction is: x _i＝[u_i,u_i+τ,...,u_i+(m-1)τ ] wherein i=1, 2, n- (m-1) τ;

3-2), calculating the distance S _ij between the i point x _i and the j point x _j in the reconstructed phase space:

S_ij＝||x_i-x_j||

where i=1, 2,..n- (m-1) τ, j=1, 2,..n- (m-1) τ, |·|| represents the norm.

3-3), Calculating a recursion value:

R(i,j)=θ(ε_i-S_ij)

Where θ (·) represents a Heaviside function, when the value in parentheses is 0 or more, θ (·) is 1, and conversely, when the value in parentheses is less than 0, θ (·) is 0. Epsilon represents a preset critical distance, and can take a fixed value or a variable value.

In this embodiment, the recursive graph is used to convert the time-series data into the image form, so that the separability of the data can be enhanced, the internal structure and the change trend of the data can be understood, the data visualization can be realized, and the data structure is more visual.

And 4, further generating a gradient map from the recursion map as shown in fig. 6, so that the change trend of the data in each direction is more obvious. The variation of a certain pixel in the generated image sample in both the X and Y directions is calculated by comparing with neighboring pixels. The change in the X-axis refers to the pixel value to the right of the current pixel (X+1) minus the pixel value to the left of the current pixel (X-1). The change in the Y-axis is the pixel value below the current pixel (Y+1) minus the pixel value above the current pixel (Y-1). After the two components are calculated, a two-dimensional vector is formed, and the image gradient of the pixel is obtained. The image function f (x, y) gradient expression is:

Wherein G _x, Represents the gradient component in the X direction, G _y,Representing the Y-direction gradient component, T representing the transpose.

The amplitude function mag (f) is denoted as g (x, y):

Direction angle function φ (x, y):

For digital images, corresponding to a two-dimensional discrete function gradient, the derivative is approximated using a difference:

G_x(x,y)=H(x+1,y)-H(x-1,y)

G_y(x,y)=H(x,y+1)-H(x,y-1)

In the formula, G _x (X, Y) represents a gradient value of the pixel point in the X direction, G _y (X, Y) represents a gradient value of the pixel point in the Y direction, and H (X, Y) represents a gray value of the pixel point.

Thus, the gradient value G (x, y) and the gradient direction α (x, y) at the pixel point (x, y) are respectively:

in the embodiment, the problem of difficult feature extraction can be solved, the data features are more obvious by using a gradient map method, the subsequent quality judgment is facilitated, and the method is suitable for analyzing complex unstable signals.

Step 5, as shown in fig. 2, the process of identifying and classifying the data by adopting the VGG16 neural network is as follows:

5-1), and after one-time average reduction pretreatment, the input image enters two convolution layers (conv 1-1 and conv 1-2), and each convolution layer uses 64 convolution kernels with the 3 multiplied by 3 to extract features, and the step length is 1.

5-2), Downsampling the feature map by a2 x 2 max pooling layer (pool 1) to reduce computation and overfitting.

5-3), Two convolution blocks, each block containing two convolution layers (conv 2-1 and conv2-2, conv3-1 and conv 3-2), each using 128 3 x 3 convolution kernels for feature extraction. Each convolution block is followed by a2 x 2 max pooling layer (pool 2 and pool 3) for further reducing the size of the feature map.

5-4), Three convolution blocks (conv 4-1 through conv4-3, conv5-1 through conv 5-3), each block containing three convolution layers, each using 256, 512, and 512 3 x 3 convolution kernels for feature extraction. Likewise, each convolution block is followed by a2×2 max pooling layer (pool 4 and pool 5).

5-5), The features are mapped and classified by three fully connected layers (FC-4096, FC-4096 and FC-1000). The first two full connection layers have 4096 hidden units, and the last full connection layer outputs the prediction result of the model.

In this embodiment, the overall architecture of the VGG16 model enhances the expressive power of the model by stacking multiple smaller convolutional layers and pooling layers to build a deep network. Meanwhile, through reasonable network design and parameter sharing, the VGG16 model has higher efficiency and training speed while maintaining higher performance. Training and classifying the gradient map by adopting a VGG16 deep neural network, inputting gradient images with the size of 224 multiplied by 224, and obtaining 760 training samples in the data set, wherein 380 samples are qualified samples, 380 samples are unqualified samples, 40 test samples are obtained, wherein 20 samples are qualified samples, and 20 samples are unqualified samples. The batch processing size of the training samples of the neural network is set to be 16, the batch processing size of the test samples is set to be 2, the iteration number of the model is set to be 20, the learning rate is set to be 0.001, and the test accuracy is accurate to 2 decimal places. Experiments prove that the accuracy of the VGG16 in distinguishing the gradient map reaches 97.5%, and the online weld quality recognition is realized.

The description and practice of the invention disclosed herein will be readily apparent to those skilled in the art, and may be modified and adapted in several ways without departing from the principles of the invention. Accordingly, modifications or improvements may be made without departing from the spirit of the invention and are also to be considered within the scope of the invention.

Claims (3)

1. An online detection method for underwater structure welds based on improved deep convolution, characterized in that it comprises the following steps: S1. Using a hydrophone sensor to receive acoustic signals of underwater structure welding in the ocean; S2, use VMD algorithm to denoise the acoustic signal and extract weak welding acoustic signals from high-order IMF components; S3, using the recursive graph to transform the effective pulse into the time-space domain, converting the one-dimensional signal from the time domain to the phase space domain, and providing more potential weld state characteristics; S4, further generating a gradient map from the recursive map, using the gradient map to perform feature extraction, and extracting features in three directions of the image X, Y and XY; S5. Use VGG16 deep convolutional network to classify image samples and complete online judgment of weld quality; In step S3, the key step of the recursive graph is to reconstruct the phase space. The recursive graph is used to convert the time series data of the effective pulses of the underwater welding acoustic signal into an image form to reveal the internal structure of the time series. Specifically, the following steps are included: S31: For the time series signal u _k (k＝1,2,...,n), determine its sampling time interval as Δt, determine the appropriate embedding dimension m and delay time τ through relevant theoretical calculations, and then reconstruct the time series. The reconstructed dynamic system is: x _i ＝[u _i ,u _i+τ ,...,u _i+(m-1)τ ] where i＝1,2,...,n-(m-1)τ; S32: Calculate the distance S _ij between point i _xi and point j _xj in the reconstructed phase space: S _ij =|| _xi - _xj || Where, i＝1,2,...,n-(m-1)τ, j＝1,2,...,n-(m-1)τ, ||·|| represents the norm; S33: Calculate the recursive value R(i,j): R(i,j)＝θ(ε _i -S _ij ), Where θ(·) represents the Heaviside function. When the number in the brackets is greater than or equal to 0, θ(·) is 1. On the contrary, when the number in the brackets is less than 0, θ(·) is 0. ε represents the pre-set critical distance, which can be a fixed value or a variable value. In step S4, the recursive graph is further converted into three gradient graphs, namely, the X-direction gradient graph, the Y-direction gradient graph, and the XY-direction gradient graph; the change of the X-axis refers to the pixel value on the right side of the current pixel (X+1) minus the pixel value on the left side of the current pixel (X-1), and the change of the Y-axis refers to the pixel value below the current pixel (Y+1) minus the pixel value above the current pixel (Y-1); after calculating these two components, a two-dimensional vector is formed, that is, the image gradient of the pixel is obtained; the gradient expression of the image function f(x, y) is: In the formula, G _x , represents the X-direction gradient component, G _y , represents the gradient component in the Y direction, and T represents transposition; The magnitude function mag(f) is expressed as g(x,y): Direction angle function φ(x,y): For digital images, it is equivalent to finding the gradient of a two-dimensional discrete function, and using differences to approximate the derivative: _Gx (x,y)＝H(x+1,y)-H(x-1,y) G _y (x,y) = H (x,y+1) - H (x,y-1) In the formula, _Gx (x,y) represents the gradient value of the pixel in the X direction, _Gy (x,y) represents the gradient value of the pixel in the Y direction, and H(x,y) represents the grayscale value of the pixel; Therefore, the gradient value G(x,y) and gradient direction α(x,y) at the pixel point (x,y) are: 2. According to claim 1, an underwater structure weld online detection method based on improved deep convolution is characterized in that, in step S2, a hydrophone is used to collect underwater welding acoustic signals, the acoustic signals are subjected to noise reduction processing by a VMD algorithm, the acoustic signals are decomposed into multiple intrinsic mode IMF components, and effective acoustic signal pulses are extracted. The working principle of the VMD algorithm is as follows: where { _uk }:={ _u1 ,... _uk } and { _ωk }:={ _ω1 ,... _ωk } are all mode sets and their frequency centers, K represents the number of IMF decomposition levels, represents partial derivative operation, φ(t) represents unit pulse function, j is the imaginary unit in Hilbert transform, t represents time, * represents convolution, f(t) represents the signal to be decomposed, and is a set of data with time series; The penalty factor ρ and Lagrange multiplier γ are introduced to transform the constraint into an unconstrained variational problem. The augmented Lagrange expression is as follows: The multiplication operator alternating algorithm is used to iteratively solve this problem. By continuously iteratively updating u _k , ω _k and γ, the optimal solution formula of VMD is obtained as follows: Wherein, F(ω), γ(ω) and U _k (ω) are the Fourier transform signals of f(t), γ(t) and _uk (t), respectively, n is the number of iterations, and τ is the noise margin coefficient. 3. According to claim 1, an online detection method for underwater structure welds based on improved deep convolution is characterized in that, in step S5, the VGG16 network is trained by a back propagation algorithm, and the network parameters are optimized using a gradient descent method. The VGG16 convolutional neural network structure includes 13 convolutional layers and 3 fully connected layers. The deeper network structure and smaller convolution kernels and pooling sampling domains can better learn abstract features; all convolutional layers use the same size of convolution kernels, step sizes and padding, and by stacking multiple small convolution kernels, a larger receptive field can be simulated while reducing the number of parameters; each convolutional layer uses a convolution kernel for feature extraction, and then samples the feature map through a maximum pooling layer, and finally maps and classifies the features through a fully connected layer, and outputs the predicted probability of each category; The specific steps include: S51: After the input image is preprocessed by subtracting the mean, it enters two convolutional layers. Each convolutional layer uses 64 3×3 convolution kernels for feature extraction. S52: Downsample the feature map through a 2×2 maximum pooling layer; S53: Two convolution blocks, each block contains two convolution layers, each convolution layer uses 128 3×3 convolution kernels for feature extraction; each convolution block is followed by a 2×2 maximum pooling layer to further reduce the size of the feature map; S54: three convolution blocks, each block contains three convolution layers, and each convolution layer uses 256, 512, and 512 3×3 convolution kernels for feature extraction; S55: Features are mapped and classified through three fully connected layers.