An online detection method for underwater structure welds based on improved deep convolution

Abstract

The present invention relates to the technical field of underwater welding detection, and in particular to an online detection method for underwater structure welds based on improved deep convolution, comprising: S1, using a hydrophone sensor to receive an acoustic signal of an underwater structure weld in the ocean; S2, using a VMD algorithm to perform denoising on the acoustic signal, and extracting a weak welding acoustic signal from a high-order IMF component; S3, using a recursive graph to perform a time-space domain transformation on an effective pulse, and converting a one-dimensional signal from a time domain to a phase space domain, providing more potential weld state features; S4, further generating a gradient graph from the recursive graph, using the gradient graph to perform feature extraction, and extracting features in three directions of image X, Y and XY; S5, using a VGG16 deep convolutional network to classify image samples, and completing online determination of weld quality. The present invention can realize online detection of underwater welding weld quality, adjust welding quality in real time, and provide a certain idea for online detection of underwater welding weld quality.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater welding detection, and in particular to an online detection method for underwater structure welds based on improved deep convolution. Background Art

[0002] Underwater welding technology plays a key role in offshore wind power, drilling platforms, ships, submarine pipelines and other fields. At present, defects in welds will seriously affect the reliability of structural parts. In order to repair as quickly as possible and reduce losses, the use of underwater welding technology will be an inevitable choice. Therefore, it is very important to develop reliable online weld quality detection technology.

[0003] Therefore, scientists at home and abroad have proposed a variety of detection methods such as vision, optics, and acoustics, but most of them have limitations. Visual detection is easily affected by external interference, such as light, temperature, humidity and other factors. In the case of complex welding defects and surface defects, the accuracy of optical detection will be affected, and the cost of optical detection is relatively high. In contrast, acoustic-based nondestructive testing technology has many advantages, such as a wide detection range and a large detection depth, which is more suitable for internal defect detection of welded structures. However, the application of traditional acoustic signal detection technology to online detection of weld quality in underwater welding also has limitations. The first problem is that the acoustic signal signal-to-noise ratio is low, and it is difficult to extract effective pulses; the second problem is that it is difficult to extract complex acoustic signal features, and traditional time domain and frequency domain analysis is not applicable; the third problem is that when the data set quality is poor and the categories are unbalanced, the quality judgment accuracy is low. To this end, the present invention proposes an online detection method for underwater structure welds based on improved deep convolution. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings existing in the prior art, and proposes an online detection method for underwater structure welds based on improved deep convolution, which can achieve a higher quality recognition rate and provide certain ideas for online identification of underwater welding weld quality.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] An online detection method for underwater structure welds based on improved deep convolution comprises the following steps:

[0007] S1. Using a hydrophone sensor to receive acoustic signals of underwater structure welding in the ocean;

[0008] S2, use VMD algorithm to denoise the acoustic signal and extract weak welding acoustic signals from high-order IMF components;

[0009] 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;

[0010] 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;

[0011] S5. Use the VGG16 deep convolutional network to classify image samples and complete online judgment of weld quality.

[0012] Preferably, 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:

[0013]

[0014]

[0015] where {u k}:={u1,...u k} and {ω k}:={ω1,...ω k} is the set of all modes 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;

[0016] 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:

[0017]

[0018] 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 as follows:

[0019]

[0020]

[0021]

[0022] Among them, F(ω), γ(ω) and U k (ω) are f(t), γ(t) and u k (t), n is the number of iterations, and τ is the noise tolerance coefficient.

[0023] Preferably, in step S3, the key step of the recursive graph is to perform phase space reconstruction, and the time series data of the effective pulses of the underwater welding acoustic signal is converted into an image form by using the recursive graph to reveal the internal structure of the time series, which specifically includes the following steps:

[0024] 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: i =[u i ,u i+τ ,...,u i+(m-1)τ ]where i=1,2,...,n-(m-1)τ;

[0025] S32: Calculate the i point x in the reconstructed phase space i and j point x j The distance S ij :

[0026] S ij =||x i -x j ||

[0027] Where, i=1,2,...,n-(m-1)τ, j=1,2,...,n-(m-1)τ, ||·|| represents the norm;

[0028] S33: Calculate recursive value:

[0029] R(i,j)=θ(ε i -S ij )

[0030] Wherein, θ(·) represents the Heaviside function. When the value in the brackets is greater than or equal to 0, θ(·) is 1. On the contrary, when the value in the brackets is less than 0, θ(·) is 0. ε represents the preset critical distance, which can be a fixed value or a variable value.

[0031] 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; the change in the X-axis refers to the pixel value on the right side (X+1) of the current pixel minus the pixel value on the left side (X-1) of the current pixel, and the change in 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:

[0032]

[0033] In the formula, G x , represents the gradient component in the X direction, G y , represents the gradient component in the Y direction, and T represents transpose.

[0034] The magnitude function mag(f) is expressed as g(x,y):

[0035]

[0036] Direction angle function φ(x,y):

[0037]

[0038] For digital images, it is equivalent to finding the gradient of a two-dimensional discrete function, and using differences to approximate the derivative:

[0039] G x (x,y)=H(x+1,y)-H(x-1,y)

[0040] G y (x,y)=H(x,y+1)-H(x,y-1)

[0041] In the formula, G x (x,y) represents the gradient value of the pixel in the X direction, G y (x,y) represents the gradient value of the pixel in the Y direction, and H(x,y) represents the grayscale value of the pixel.

[0042] Therefore, the gradient value G(x,y) and gradient direction α(x,y) at the pixel point (x,y) are:

[0043]

[0044]

[0045] Preferably, in step S5, the VGG16 network is trained by a back-propagation algorithm, and the network parameters are optimized by 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, strides, and padding. 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. Finally, the features are mapped and classified through a fully connected layer, and the predicted probability of each category is output.

[0046] The specific steps include:

[0047] 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.

[0048] S52: Downsample the feature map through a 2×2 maximum pooling layer;

[0049] 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;

[0050] 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;

[0051] S55: Features are mapped and classified through three fully connected layers.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. The present invention displays weak welding acoustic signals from the perspective of phase space. First, the acoustic signal is decomposed into multiple IMF components by VMD algorithm; then, the time series is converted into phase space using recursion diagram.

[0054] 2. The present invention improves the detection rate of convolutional neural networks by extracting gradient features from recursive graph images. The VGG16 neural network can automatically learn to identify effective gradient image features. Compared with the classification results of three neural networks, VGG19, RESNET50, and RESNET34, the proposed method can achieve efficient online weld quality classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the present invention;

[0056] Figure 2 A schematic diagram of a welded steel plate experimental sample according to an embodiment of the present invention;

[0057] Figure 3 is an original audio signal diagram of an embodiment of the present invention;

[0058] Figure 4 This is an IMF component diagram of a sound signal after VMD noise reduction according to an embodiment of the present invention;

[0059] Figure 5 is a recursion diagram of an embodiment of the present invention;

[0060] Figure 6It is a gradient map of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the protection scope of the present invention. The embodiments described in the present invention are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present invention.

[0062] Reference Figure 1 , an online detection method for underwater structure welds based on improved deep convolution, comprising the following steps:

[0063] Step 1: Using a hydrophone sensor to receive acoustic signals of underwater structure welding in the ocean;

[0064] Step 2: Use VMD algorithm to denoise the acoustic signal and extract weak welding acoustic signals from high-order IMF components;

[0065] Step 3: Use the recursive graph to transform the effective pulse in the time-space domain, convert the one-dimensional signal from the time domain to the phase space domain, and provide more potential weld state characteristics;

[0066] Step 4: The recursive graph is further converted into a gradient graph, and the gradient graph is used for feature extraction to extract features in the three directions of X, Y and XY of the image;

[0067] Step 5: Use the VGG16 deep convolutional network to classify image samples and complete online judgment of weld quality.

[0068] See also Figures 1 to 6 The implementation steps of the technical solution provided by the present invention are as follows:

[0069] Step 2: If Figure 2 As shown in the figure, after the experimental platform is built, the steel plates are welded and the qualified and unqualified steel plates are classified. Figure 3 As shown in Figure 1, the original audio signal is collected after the underwater welding acoustic signal is collected by a hydrophone. Figure 4 As shown, the VMD algorithm is used for noise reduction. The VMD algorithm decomposes the acoustic signal into three intrinsic mode components. The IMF decomposition layer number K is set to 3 layers, the band limit alpha is set to 3000, the noise tolerance coefficient tau is 0, the DC component DC is set to 0, and the effective acoustic signal pulse is extracted. The working principle of the VMD algorithm is as follows:

[0070]

[0071]

[0072] where {u k}:={u1,...u k} and {ω k}:={ω1,...ω k} is the set of all modes 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;

[0073] 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:

[0074]

[0075] 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 as follows:

[0076]

[0077]

[0078]

[0079] Among them, F(ω), γ(ω) and U k (ω) are f(t), γ(t) and u k (t), n is the number of iterations, and τ is the noise tolerance coefficient.

[0080] In this embodiment, the VMD algorithm can decompose non-stationary signals under low signal-to-noise ratio conditions, extract effective pulses, and completely separate pulses and signals. The acoustic signals collected in complex environments have good noise reduction processing effects, and the signal-to-noise ratio of the acoustic signals after noise reduction is high.

[0081] Step 3: The key step of the recurrence graph is to reconstruct the phase space. Figure 5 The recurrence diagram of effective pulse generation is shown below. The recurrence diagram converts the time series data into an image. The specific steps are as follows:

[0082] 3-1) 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: i =[u i ,u i+τ ,...,u i+(m-1)τ ]where i=1,2,...,n-(m-1)τ;

[0083] 3-2) Calculate the i point x in the reconstructed phase space i and j point x j The distance S ij :

[0084] S ij =||x i -x j ||

[0085] Wherein, i=1,2,...,n-(m-1)τ, j=1,2,...,n-(m-1)τ, and ||·|| represents the norm.

[0086] 3-3) Calculate the recursive value:

[0087] R(i,j)=θ(ε i -S ij )

[0088] Wherein, θ(·) represents the Heaviside function. When the value in the bracket is greater than or equal to 0, θ(·) is 1. On the contrary, when the value in the bracket is less than 0, θ(·) is 0. ε represents the preset critical distance, which can be a fixed value or a variable value.

[0089] In this embodiment, the recursive graph is used to convert the time series data into an image form, which can enhance the separability of the data and help understand the inherent structure and change trend of the data, realize data visualization, and make the data structure more intuitive.

[0090] Step 4: Figure 6 As shown, the recursive graph is further used to generate a gradient graph, making the trend of data changes in all directions more obvious. The change of a pixel in the generated image sample in the X and Y directions is calculated by comparing it with adjacent pixels. The change in 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). 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 calculating these two components, a two-dimensional vector is formed, which is the image gradient of the pixel. The gradient expression of the image function f(x,y) is:

[0091]

[0092] In the formula, G x , represents the gradient component in the X direction, G y , represents the gradient component in the Y direction, and T represents transpose.

[0093] The magnitude function mag(f) is expressed as g(x,y):

[0094]

[0095] Direction angle function φ(x,y):

[0096]

[0097] For digital images, it is equivalent to finding the gradient of a two-dimensional discrete function, and using differences to approximate the derivative:

[0098] G x (x,y)=H(x+1,y)-H(x-1,y)

[0099] G y (x,y)=H(x,y+1)-H(x,y-1)

[0100] In the formula, G x (x,y) represents the gradient value of the pixel in the X direction, G y (x,y) represents the gradient value of the pixel in the Y direction, and H(x,y) represents the grayscale value of the pixel.

[0101] Therefore, the gradient value G(x,y) and gradient direction α(x,y) at the pixel point (x,y) are:

[0102]

[0103]

[0104] In this embodiment, the problem of difficult feature extraction can be solved. The gradient map method is used to make the data features more obvious, which is beneficial to subsequent quality judgment and is suitable for the analysis of complex and unstable signals.

[0105] Step 5: Figure 2 As shown in the figure, the process of using VGG16 neural network to identify and classify data is as follows:

[0106] 5-1) After the input image is preprocessed by subtracting the mean, it enters two convolutional layers (conv1-1 and conv1-2). Each convolutional layer uses 64 3×3 convolution kernels for feature extraction with a step size of 1.

[0107] 5-2) Downsample the feature map through a 2×2 maximum pooling layer (pool1) to reduce the amount of computation and overfitting.

[0108] 5-3), two convolution blocks, each block contains two convolution layers (conv2-1 and conv2-2, conv3-1 and conv3-2), each convolution layer uses 128 3×3 convolution kernels for feature extraction. Each convolution block is followed by a 2×2 maximum pooling layer (pool2 and pool3) to further reduce the size of the feature map.

[0109] 5-4), three convolution blocks (conv4-1 to conv4-3, conv5-1 to conv5-3), each block contains three convolution layers, each convolution layer uses 256, 512 and 512 3×3 convolution kernels for feature extraction. Similarly, each convolution block is followed by a 2×2 maximum pooling layer (pool4 and pool5).

[0110] 5-5) Map and classify the features through three fully connected layers (FC-4096, FC-4096, and FC-1000). The first two fully connected layers have 4096 hidden units each, and the last fully connected layer outputs the model's prediction results.

[0111] In this embodiment, the overall architecture of the VGG16 model constructs a deep network by stacking multiple smaller convolutional layers and pooling layers, which enhances the expressive power of the model. At the same time, through reasonable network design and parameter sharing, the VGG16 model has high efficiency and training speed while maintaining high performance. The VGG16 deep neural network is used to train and classify the gradient map. The input is set to a gradient image of size 224×224. There are 760 training samples in the data set, of which 380 are qualified samples and 380 are unqualified samples. There are 40 test samples, of which 20 are qualified samples and 20 are unqualified samples. The batch processing size of the neural network training samples is set to 16, the batch processing size of the test samples is 2, the number of model iterations is set to 20, the learning rate is set to 0.001, and the test accuracy is accurate to 2 decimal places. It has been experimentally verified that the discrimination accuracy of VGG16 for gradient maps has reached 97.5%, realizing online weld quality recognition.

[0112] The description and practice disclosed in the present invention are easy to think and understand for ordinary technicians in the technical field, and several improvements and modifications can be made without departing from the principles of the present invention. Therefore, modifications or improvements made without departing from the spirit of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. An online detection method for underwater structure welds based on improved deep convolution, characterized in that: The following steps are involved: 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: i =[u i ,u i+τ ,...,u i+(m-1)τ ]where i=1,2,...,n-(m-1)τ; S32: Calculate the i point x in the reconstructed phase space i and j point x j The distance S ij : 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: 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, an X-direction gradient graph, a Y-direction gradient graph, and an 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 gradient component in the X direction, 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: 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 the gradient value of the pixel in the X direction, G y (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. The method for online detection of underwater structure welds based on improved deep convolution according to claim 1 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 the 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 {u k }:={u1,...u k } and {ω k }:={ω1,...ω k } is the set of all modes 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 as follows: Among them, F(ω), γ(ω) and U k (ω) are f(t), γ(t) and u k (t), n is the number of iterations, and τ is the noise tolerance coefficient.

3. The method for online detection of underwater structure welds based on improved deep convolution according to claim 1 is characterized in that: In step S5, the VGG16 network is trained by the back-propagation algorithm, and the network parameters are optimized using the 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 kernel and pooling sampling domain can better learn abstract features. All convolutional layers use the same size of convolution kernel, stride and padding. By stacking multiple small convolution kernels, a larger receptive field can be simulated while reducing the number of parameters. Each convolutional layer uses the convolution kernel for feature extraction, then samples the feature map through the maximum pooling layer, and finally maps and classifies the features through the 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.

Images