Hybrid convolutional neural network construction method for pipeline damage calculation

Abstract

The invention discloses a hybrid convolutional neural network construction method for pipeline damage calculation, belongs to the technical field of neural network construction, and is used for pipeline damage calculation, when pipeline damage is calculated, strain analysis is performed on a pipeline to obtain a strain data set, and the strain data set is divided into a training set and a verification set; inputting the training set into the hybrid convolutional neural network for training, and after training meets a precision requirement, inputting the verification set into the trained hybrid convolutional neural network for pipeline damage calculation to obtain a pipeline damage calculation result; the hybrid convolutional neural network sequentially comprises an input layer, N hybrid convolutional layers, a bidirectional gating circulation layer, an attention mechanism layer, a full connection layer and an output layer, and one hybrid convolutional layer comprises a deep convolutional layer and a pointwise convolutional layer. The method can effectively overcome the problem that a large amount of time is consumed when pipeline damage calculation is carried out on the ocean platform, and improves the health monitoring accuracy of the existing ocean platform.

Description

Technical Field

[0001] The invention discloses a hybrid convolutional neural network construction method for pipeline damage calculation, belonging to the technical field of neural network construction. Background Art

[0002] Composite material pipelines are inevitably subject to defects and damage during the manufacturing, assembly and use process. These hidden dangers will weaken the bearing capacity and reliability of the pipeline, resulting in reduced structural strength, shortening its service life, and may even cause major safety hazards. The complexity of the properties of composite materials and the diversity of the loads they bear lead to complex and varied failure modes, and internal damage such as delamination damage is often difficult to detect intuitively. In order to ensure the safe operation of composite material pipelines and extend their service life, pipeline damage calculation becomes the key. Traditional pipeline damage calculation technologies usually include thermal imaging, laser ultrasonic detection, acoustic emission detection, ray detection, electroluminescence testing, etc. However, the traditional pipeline damage calculation method cannot quickly process the pipeline status, and often requires actual measurement by instruments and processing of actual instrument measurement data, which is expensive and inefficient. Summary of the invention

[0003] The purpose of the present invention is to provide a hybrid convolutional neural network construction method for pipeline damage calculation to solve the problem that pipeline damage calculation is expensive and inefficient in the prior art.

[0004] A hybrid convolutional neural network construction method for pipeline damage calculation, the hybrid convolutional neural network includes an input layer, N hybrid convolutional layers, a bidirectional gated recurrent layer, an attention mechanism layer, a fully connected layer and an output layer in sequence, and a hybrid convolutional layer includes a depth convolutional layer and a point-by-point convolutional layer; When calculating pipeline damage, strain analysis is performed on the pipeline to obtain a strain data set, which is divided into a training set and a validation set. The training set is input into the hybrid convolutional neural network for training. After the training meets the accuracy requirements, the validation set is input into the trained hybrid convolutional neural network for pipeline damage calculation to obtain the pipeline damage calculation results.

[0005] The pipeline vibration signal input by the deep convolution layer is convolved: ; In the formula, represents the convolution operation, is the weight matrix, is the strain data set, is the coordinate of the feature output, is the size of the maximum convolution kernel, is the size of the maximum output channel, is the convolution kernel size, is the size of the output channel; After convolution, use Activation function, followed by batch normalization: ; In the formula, is the feature after batch normalization, is batch normalization, Indicates the use Activation function; Perform a depthwise convolution operation on each feature on each channel: ; ; In the formula, is a depth-wise separable convolution, It is the output feature after deep convolution operation; The output of the depthwise convolution is added to the output of the batch normalization process using the residual principle: ; In the formula, is the output feature of the deep convolutional layer.

[0006] No. The batch normalization layer is , batch normalization is: ; ; ; ; In the formula, yes The elements, yes The elements, Total elements, is the stretching parameter, is the offset parameter, It means taking the expected value. is the maximum value of the minimum batch size, is the value of the minimum batch size, Indicates taking the variance.

[0007] The point-by-point convolutional layer consists of: ; ; In the formula, is point-wise convolution, is the output feature of the mixed convolutional layer.

[0008] The bidirectional gated loop layer includes gated loop layers in the forward and reverse directions, which read information in forward and backward time order respectively, and perform forward gated loop calculation and reverse gated loop calculation respectively, and finally output the values ​​of the two gated loop layers jointly. The gated loop layer includes a reset gate and an update gate.

[0009] Resetting the gate involves using the reset gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : ; In the formula, is the time step of , is the time step The hidden state of for function; The update gate includes the use of the update gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : .

[0010] Candidate hidden states for: ; Update the memory: ; is the time step The hidden state of Forgot the time step of unimportant information, Yes Further options for information in.

[0011] The reverse gated recurrent layer is: ; ; ; ; In the formula, for The reverse of for The reverse of for The reverse of for The reverse of is the time step The hidden state of .

[0012] and Combined to get the feature vector , as the output result of the bidirectional gated recurrent layer.

[0013] The attention mechanism layer consists of: ; ; ; In the formula, yes The property representation of is the weight matrix of the attention mechanism layer, is the importance weight, is the activation function, Yes The result of the importance-weighted summation is is a context vector randomly generated during training.

[0014] Compared with the prior art, the present invention has the following beneficial effects: it can effectively overcome the situation where a lot of time is spent in calculating pipeline damage on offshore platforms, and improve the accuracy of health monitoring of existing offshore platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the loss value result of convolutional neural network CNN; Figure 2 This is the accuracy result of the convolutional neural network CNN; Figure 3 The loss value result of the convolutional neural network CNN combined with the gated recurrent GRU; Figure 4 This is the accuracy result of the convolutional neural network CNN combined with the gated recurrent GRU; Figure 5 The loss value result of the convolutional neural network CNN combined with the bidirectional gated cycle BIGRU; Figure 6 The accuracy results of the convolutional neural network CNN combined with the bidirectional gated recurrent BIGRU; Figure 7 The loss value results of the convolutional neural network CNN combined with the bidirectional gated recurrent BIGRU and attention mechanism; Figure 8 Accuracy results of the convolutional neural network (CNN) combined with the bidirectional gated recurrent BIGRU and attention mechanism; Fig. 9 is the loss value result of the present invention; Fig.10 This is the accuracy result of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] A hybrid convolutional neural network construction method for pipeline damage calculation, the hybrid convolutional neural network includes an input layer, N hybrid convolutional layers, a bidirectional gated recurrent layer, an attention mechanism layer, a fully connected layer and an output layer in sequence, and a hybrid convolutional layer includes a depth convolutional layer and a point-by-point convolutional layer; When calculating pipeline damage, strain analysis is performed on the pipeline to obtain a strain data set, which is divided into a training set and a validation set. The training set is input into the hybrid convolutional neural network for training. After the training meets the accuracy requirements, the validation set is input into the trained hybrid convolutional neural network for pipeline damage calculation to obtain the pipeline damage calculation results.

[0018] The pipeline vibration signal input by the deep convolution layer is convolved: ; In the formula, represents the convolution operation, is the weight matrix, is the strain data set, is the coordinate of the feature output, is the size of the maximum convolution kernel, is the size of the maximum output channel, is the convolution kernel size, is the size of the output channel; After convolution, use Activation function, followed by batch normalization: ; In the formula, is the feature after batch normalization, is batch normalization, Indicates the use Activation function; Perform a depthwise convolution operation on each feature on each channel: ; ; In the formula, is a depth-wise separable convolution, It is the output feature after deep convolution operation; The output of the depthwise convolution is added to the output of the batch normalization process using the residual principle: ; In the formula, is the output feature of the deep convolutional layer.

[0019] No. The batch normalization layer is , batch normalization is: ; ; ; ; In the formula, yes The elements, yes The elements, Total elements, is the stretching parameter, is the offset parameter, It means taking the expected value. is the maximum value of the minimum batch size, is the value of the minimum batch size, Indicates taking the variance.

[0020] The point-by-point convolutional layer consists of: ; ; In the formula, is point-wise convolution, is the output feature of the mixed convolutional layer.

[0021] The bidirectional gated loop layer includes gated loop layers in the forward and reverse directions, which read information in forward and backward time order respectively, and perform forward gated loop calculation and reverse gated loop calculation respectively, and finally output the values ​​of the two gated loop layers jointly. The gated loop layer includes a reset gate and an update gate.

[0022] Resetting the gate involves using the reset gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : ; In the formula, is the time step of , is the time step The hidden state of for function; The update gate includes the use of the update gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : .

[0023] Candidate hidden states for: ; Update the memory: ; is the time step The hidden state of Forgot the time step of unimportant information, Yes Further options for information in.

[0024] The reverse gated recurrent layer is: ; ; ; ; In the formula, for The reverse of for The reverse of for The reverse of for The reverse of is the time step The hidden state of .

[0025] and Combined to get the feature vector , as the output result of the bidirectional gated recurrent layer.

[0026] The attention mechanism layer consists of: ; ; ; In the formula, yes The property representation of is the weight matrix of the attention mechanism layer, is the importance weight, is the activation function, Yes The result of the importance-weighted summation is is a context vector randomly generated during training.

[0027] In the embodiment of the present invention, firstly, according to the damage state of the composite material pipeline in the actual project, the geometric models of 9 pipelines with different damages and 1 healthy pipeline were established by using SOLIDWORKS modeling software, among which there were 4 types of single damage, including hole damage, scratch damage, crack damage, and interlayer debonding damage, and 5 types of multiple damage, including scratch damage and crack damage, crack damage and hole damage, scratch damage and hole damage, crack damage and interlayer debonding damage, and double crack damage.

[0028] A cylindrical groove of H1mm×R1mm is dug in the center of the pipeline to simulate the damage of pipeline holes, where H1 is the damage depth, the size is 5mm, and R1 is the damage radius, the size is 3mm; a groove of L1mm×D1mm×H2mm is polished in the center of the pipeline to simulate the damage of pipeline scratches, where L1 is the damage length, the size is 50mm, D1 is the damage width, the size is 30mm, and H2 is the damage depth, the size is 0.5mm. A groove of L2mm×D2mm×H3mm is cut in the center of the pipeline to simulate the damage of pipeline cracks, where L2 is the damage length, the size is 30mm, D2 is the damage width, the size is 1mm, and H3 is the damage depth, the size is 0.5mm. A groove of L3mm×D3mm×H4mm is dug in the inner layer of the pipeline to simulate the damage between pipeline layers, where L3 is the damage length, the size is 30mm, D3 is the damage width, the size is 10mm, and H4 is the damage depth, the size is 2mm.

[0029] Using the finite element ANSYS software, a finite element model was established, and the models of the 9 pipes with different damages and 1 healthy pipe just established were imported into the Transient Structural module in the ANSYS software. Because the pipe is a composite pipe, the outer layer is made of glass fiber and the inner layer is made of PE, so glass fiber and polyethylene are found in the engineering data and applied to the pipe, and then the contact form of the connection surface is set and the grid is divided. Finally, a fixed constraint is added to one end of the pipe to form a cantilever beam, and the other end is vertically upward and perpendicular to the axis of the pipe. The maximum and minimum values ​​of the random excitation are 200N and -200N respectively. Strain analysis is performed on the above 10 pipes to obtain strain data sets, and the data are divided into training sets and validation sets. In actual application, one end of the composite flexible elbow is clamped on a fixed bracket, and the other end is connected to the exciter with a fixture. The optical fiber strain sensor is evenly attached to the upper bending part of the elbow, and the exciter is set to random excitation. The strain data of 9 pipes with different damages and 1 healthy pipe are recorded. A total of 10 sets of data are used as the original data set of the neural network as the test set.

[0030] The strain data of 10 types of pipelines were processed and input into the neural network model training of the present invention. The structural parameters are shown in Table 1.

[0031] Table 1 Neural network structure parameters of the present invention ; The strain data of 10 types of pipes were input into the comparison model for ablation experiments for comparison. Figure 1 is the training loss value of the convolutional neural network CNN on the strain data of composite flexible bending pipe, Figure 2It is the training accuracy of the convolutional neural network CNN on the strain data of composite flexible bending pipes. The model converged in about 800 generations, with an accuracy of 69.7% and a loss value of 0.74, and the performance is average. Figure 3 is the loss value of the convolutional neural network CNN combined with the gated cycle GRU, Figure 4 It is the accuracy of the convolutional neural network CNN combined with the gated recurrent GRU. Compared with the convolutional neural network CNN, the gating mechanism is introduced. The model converges in about 800 generations, with an accuracy of 73.5% and a loss value of 0.58, and the performance has been improved. Figure 5 It is the loss value of the training result of the convolutional neural network CNN combined with the bidirectional gated cycle BIGRU type. Figure 6 The accuracy of the training results of the convolutional neural network CNN combined with the bidirectional gated recurrent BIGRU type is shown in Figure 2. The model converged in about 850 generations, with an accuracy of 86.1% and a loss value of 0.43, and the performance was improved by 12.6% compared to CNN-GRU. Figure 7 is the loss value of the convolutional neural network CNN combined with the bidirectional gated cycle BIGRU and the attention mechanism, Figure 8 It is the accuracy of the convolutional neural network CNN combined with the bidirectional gated recurrent BIGRU and the attention mechanism. After adding the attention mechanism, the model converged in about 500 generations, with an accuracy of 96.4%, a loss value of 0.11, and a performance improvement of 10.3%. Fig. 9 What is shown is the loss value result of the present invention, Fig.10 What is shown is the accuracy result of the present invention. The model converges in about 100 generations, and the accuracy reaches 100%, the loss value is 0, and the performance is the best.

[0032] From the results, it can be seen that the gating mechanism, bidirectional structure, attention mechanism and hybrid convolution layer introduced in the present invention can improve the training effect of the neural network model for the strain data of composite material flexible pipes. The introduction of attention mechanism and bidirectional mechanism improves the accuracy by more than 10%, while the introduction of hybrid convolution layer makes the training accuracy reach 100%.

[0033] The damaged pipe was placed on the experimental platform, and one end was fixed and the other end was stimulated to obtain new strain data measured in the experiment under actual working conditions, which was used to test the performance of the trained neural network. The strain data obtained in the experiment was input into the trained model, and the results obtained are shown in Table 2. It can be seen from the table that the accuracy of the test of the present invention is the highest, which is 96.89%, and the accuracy of the CNN model test is the lowest, which is 65.44%.

[0034] Table 2 Test results of five neural networks .

[0035] The hybrid convolutional neural network model used in the present invention can accurately and efficiently identify various damages in the pipeline through its complex network structure and powerful learning ability. Traditional non-destructive testing technologies usually include thermal imaging, laser ultrasonic testing, acoustic emission testing, ray detection, electroluminescence testing, etc. Traditional non-destructive testing systems are generally complex in equipment and are limited by the testing environment. They often require manual operation and verification, which is not conducive to timely detection of the initiation and spread of damage under complex working conditions. Compared with traditional damage identification methods, the hybrid convolutional neural network model can automatically extract useful features from raw data without manual design and selection of features. This automated feature extraction capability not only improves the accuracy of damage identification, but also greatly simplifies the identification process and reduces the errors caused by human factors. In pipeline damage identification, the hybrid convolutional neural network model can process a large amount of signal data and quickly identify the type of damage. This high efficiency enables the invention to achieve real-time monitoring and early warning in practical applications, and timely discover and deal with potential pipeline safety hazards. Pipeline damage signals are often affected by multiple factors such as noise and interference, resulting in a decrease in signal quality, thereby affecting the accuracy of damage identification. However, the hybrid convolutional neural network in the present invention can effectively suppress noise and interference and extract useful damage features through two convolution methods, one-dimensional deep convolution and point-by-point convolution, in its structure. The final experimental results also show that the hybrid convolutional neural network can still maintain a high recognition accuracy when processing noisy pipeline damage signals. This powerful anti-noise ability and robustness enable the hybrid convolutional neural network to have better adaptability and stability in complex and changeable practical application environments. In addition, the hybrid convolutional neural network in the present invention also uses a data enhancement method to enhance the obtained strain data, further improving the generalization ability and robustness of the method, so that it can better cope with various unknown and complex damage conditions of composite pipes. At present, in the field of pipeline engineering, domestic and foreign research mainly focuses on the detection and identification of single-material pipes such as steel pipes, three-type polypropylene (PPR) pipes, polyethylene (PE) pipes, and cement pipes. These single-material pipe identification methods are relatively mature and diverse, thanks to the clarity and stability of their material properties. However, when it comes to composite pipes, the situation becomes much more complicated. Due to the diversity of its material composition and the complexity of its structure, the types of damage to composite pipes are often more numerous and difficult to accurately identify. Although some studies have attempted to identify damage in composite pipes, due to the limitations of technical means and recognition algorithms, the accuracy of these methods is generally low and cannot meet the high-precision requirements in practical applications. In order to fill the gap in this field, this paper proposes an innovative recognition method - using a hybrid convolutional neural network model to identify and classify the types of damage in composite pipes.This method combines the advantages of one-dimensional convolutional neural networks in one-dimensional data processing and identification, and trains and learns various damage data of composite pipes through deep learning algorithms, thereby achieving accurate identification of damage types. Experimental results show that the accuracy of composite pipe damage type identification using the hybrid convolutional neural network model proposed in this invention is as high as more than 96%. This achievement not only fills the gap in the field of composite pipe damage identification, but also provides strong technical support for the accuracy and reliability of composite pipe safety monitoring.

[0036] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features may be replaced by equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hybrid convolutional neural network construction method for pipeline damage calculation, characterized in that: The hybrid convolutional neural network includes an input layer, N hybrid convolutional layers, a bidirectional gated recurrent layer, an attention mechanism layer, a fully connected layer, and an output layer. A hybrid convolutional layer includes a depth convolutional layer and a point-by-point convolutional layer. When calculating pipeline damage, strain analysis is performed on the pipeline to obtain a strain data set, which is divided into a training set and a validation set. The training set is input into the hybrid convolutional neural network for training. After the training meets the accuracy requirements, the validation set is input into the trained hybrid convolutional neural network for pipeline damage calculation to obtain the pipeline damage calculation results.

2. According to claim 1, a hybrid convolutional neural network construction method for pipeline damage calculation is characterized in that: The pipeline vibration signal input by the deep convolution layer is convolved: ； In the formula, represents the convolution operation, is the weight matrix, is the strain data set, is the coordinate of the feature output, is the size of the maximum convolution kernel, is the size of the maximum output channel, is the convolution kernel size, is the size of the output channel; After convolution, use Activation function, followed by batch normalization: ； In the formula, is the feature after batch normalization, is batch normalization, Indicates the use Activation function; Perform a depthwise convolution operation on each feature on each channel: ； ； In the formula, is a depth-wise separable convolution, It is the output feature after deep convolution operation; The output of the depthwise convolution is added to the output of the batch normalization process using the residual principle: ； In the formula, is the output feature of the deep convolutional layer.

3. The method for constructing a hybrid convolutional neural network for pipeline damage calculation according to claim 2, characterized in that: No. The batch normalization layer is , batch normalization is: ； ； ； ； In the formula, yes The elements, yes The elements, Total elements, is the stretching parameter, is the offset parameter, It means taking the expected value. is the maximum value of the minimum batch size, is the value of the minimum batch size, Indicates taking the variance.

4. The method for constructing a hybrid convolutional neural network for pipeline damage calculation according to claim 3, characterized in that: The point-by-point convolutional layer consists of: ； ； In the formula, is point-wise convolution, is the output feature of the mixed convolutional layer.

5. The method for constructing a hybrid convolutional neural network for pipeline damage calculation according to claim 4, characterized in that: The bidirectional gated loop layer includes gated loop layers in the forward and reverse directions, which read information in forward and backward time order respectively, and perform forward gated loop calculation and reverse gated loop calculation respectively, and finally output the values ​​of the two gated loop layers jointly. The gated loop layer includes a reset gate and an update gate.

6. A hybrid convolutional neural network construction method for pipeline damage calculation according to claim 5, characterized in that: Resetting the gate involves using the reset gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : ； In the formula, is the time step of , is the time step The hidden state of for function; The update gate includes the use of the update gate weight matrix right and The matrix composed of is linearly transformed, and the result of the linear transformation is passed through Function, get the value of the reset gate : 。 7. A hybrid convolutional neural network construction method for pipeline damage calculation according to claim 6, characterized in that: Candidate hidden states for: ； Update the memory: ； is the time step The hidden state of Indicates forgetting the time step of unimportant information, Yes Further options for information in.

8. The method for constructing a hybrid convolutional neural network for pipeline damage calculation according to claim 7, characterized in that: The reverse gated recurrent layer is: ； ； ； ； In the formula, for The reverse of for The reverse of for The reverse of for The reverse of is the time step The hidden state of .

9. A hybrid convolutional neural network construction method for pipeline damage calculation according to claim 8, characterized in that: and Combined to get the feature vector , as the output result of the bidirectional gated recurrent layer.

10. A hybrid convolutional neural network construction method for pipeline damage calculation according to claim 9, characterized in that: The attention mechanism layer consists of: ； ； ； In the formula, yes The property representation of is the weight matrix of the attention mechanism layer, is the importance weight, is the activation function, Yes The result of the importance-weighted summation is is a context vector randomly generated during training.

Images