Pipeline bending damage degree recognition method and system based on wavelet packet transform

Abstract

The application discloses a pipeline bending part damage degree recognition method and system based on wavelet packet transformation, and relates to the field of pipeline bending part damage degree recognition. The pipeline bending part damage degree recognition method based on wavelet packet transformation comprises the following steps: acquiring strain modal data of a pipeline bending part; reconstructing the strain modal data to obtain reconstructed strain modal data; performing wavelet packet transformation on the reconstructed strain modal data to obtain N-layer wavelet packets; inputting the last layer of wavelet packets in the N-layer wavelet packets into a trained one-dimensional convolutional neural network to obtain the damage state of the pipeline bending part. The application extracts the features of damage itself based on the one-dimensional convolutional neural network and recognizes and positions the damage of the flow pipeline of the offshore platform, effectively overcomes the condition that there is no prior experience when the health of the pipeline bending part of the active offshore platform is detected, and improves the health monitoring accuracy of the active offshore platform.

Description

Technical Field

[0001] This invention relates to the field of damage identification at pipe bends, and in particular to a method and system for identifying the degree of damage at pipe bends based on wavelet packet transform. Background Technology

[0002] Pipelines play a crucial role in marine engineering. Since the points of greatest stress in pipelines are typically located at bends, these bends are highly susceptible to damage, leading to leaks and other problems that can cause serious accidents and irreparable losses to public safety and the economy. Therefore, damage identification at pipeline bends is urgently needed. Currently, a sensor-based approach (installing sensors on the surface of the bend to collect data such as acceleration, stress, and vibration) is used to address this issue. This data is then compared with data from undamaged bends using various methods. However, the biggest drawback of this method is its inability to be applied to operational marine platforms, as data from the intact state is no longer available, thus preventing the accurate identification of the extent of damage at pipeline bends. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for identifying the degree of damage at pipe bends based on wavelet packet transform. This method can effectively overcome the lack of prior experience in health monitoring of pipe bends on active offshore platforms and improve the accuracy of health monitoring of active offshore platforms.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A method for identifying the degree of damage at pipe bends based on wavelet packet transform, comprising:

[0006] Obtain strain modal data at the pipe bend;

[0007] The strain modal data are reconstructed to obtain reconstructed strain modal data;

[0008] Wavelet packet transform is performed on the reconstructed strain modal data to obtain N layers of wavelet packets;

[0009] The last wavelet packet in the N-layer wavelet packet is input into the trained one-dimensional convolutional neural network to obtain the damage state at the bend of the pipeline.

[0010] Optionally, the last wavelet packet in the N wavelet packets is input into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend, which also includes:

[0011] Construct a damage model at pipe bends;

[0012] Modal analysis was performed on the damage model at the bend of the pipeline to obtain simulated strain modal data;

[0013] Noise simulation is performed on the simulated strain mode data to obtain noise-simulated strain mode data;

[0014] The noise-simulated strain mode data is reconstructed to obtain reconstructed simulated strain mode data;

[0015] Wavelet packet transform is performed on the reconstructed simulated strain modal data to obtain N simulated layer wavelet packets;

[0016] The last simulated wavelet packet in the N-layer simulated wavelet packet is used as training data and input into a one-dimensional convolutional neural network to train the one-dimensional convolutional neural network, thus obtaining the trained one-dimensional convolutional neural network.

[0017] Optionally, noise simulation is performed on the simulated strain modal data to obtain noise-simulated strain modal data, specifically including:

[0018] Additive white Gaussian noise is used to simulate noise in actual engineering. The mathematical model is as follows:

[0019]

[0020] The noise-simulated strain mode data:

[0021]

[0022] Where SNR is the signal-to-noise ratio; x is the simulated strain modal data; n is the noise signal; x noise The noise is used to simulate strain mode data; N is the length of the simulated strain mode data; rand(N) is N random decimals in the range [0,1] that conform to a standard normal Gaussian distribution.

[0023] Optionally, the strain modal data is reconstructed to obtain reconstructed strain modal data, specifically including:

[0024] The strain modal data are reconstructed using spline interpolation to obtain the reconstructed strain modal data.

[0025] Optionally, wavelet packet transform is performed on the reconstructed strain modal data to obtain N layers of wavelet packets, specifically including:

[0026] The reconstructed strain modal data are normalized to obtain normalized data;

[0027] The normalized data is subjected to wavelet packet transform to obtain the N-layer wavelet packets.

[0028] Optionally, the reconstructed strain modal data is normalized to obtain normalized data, as shown in the formula:

[0029]

[0030] Where, x i ' represents the i-th value in the normalized data; x i Let x be the i-th value in the reconstructed strain modal data; min(x1,x2,x3,.....x n ) represents the minimum eigenvalue in the reconstructed strain modal data; max(x1,x2,x3,.....x n ) represents the largest eigenvalue in the reconstructed strain modal data; x n This is the last value in the reconstructed strain modal data.

[0031] Optionally, the last wavelet packet in the N wavelet packets is input into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend, specifically including:

[0032] The last wavelet packet in the N-layer wavelet packet is normalized to obtain the normalized last wavelet packet, as shown in the formula:

[0033]

[0034] Among them, y i ' represents the i-th value in the last wavelet packet after normalization; y i The i-th value in the last wavelet packet; min(y1, y2, y3, ..., y n ) represents the minimum eigenvalue in the last wavelet packet; max(y1,y2,y3,.....y n ) represents the largest eigenvalue in the last wavelet packet; y n This is the last value in the last wavelet packet.

[0035] A wavelet packet transform-based system for identifying the degree of damage at pipe bends is described above. This system is applied to the aforementioned method and includes the following components:

[0036] The data acquisition module is used to acquire strain modal data at pipe bends;

[0037] The data reconstruction module is used to reconstruct the strain mode data to obtain reconstructed strain mode data;

[0038] The wavelet packet module is used to perform wavelet packet transformation on the reconstructed strain mode data to obtain N layers of wavelet packets.

[0039] The neural network module is used to input the last wavelet packet in the N-layer wavelet packet into the trained one-dimensional convolutional neural network to obtain the damage state at the bend of the pipeline.

[0040] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for identifying the degree of damage at pipe bends based on wavelet packet transform.

[0041] A computer-readable storage medium storing a computer program, which, when executed, implements the aforementioned method for identifying the degree of damage at pipe bends based on wavelet packet transform.

[0042] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0043] This invention acquires strain modal data at the pipe bend. The strain modal data is then reconstructed to obtain reconstructed strain modal data. Wavelet packet transform is performed on the reconstructed strain modal data to obtain N layers of wavelet packets. The last layer of wavelet packets is input into a trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend. This invention extracts the features of the damage itself based on a one-dimensional convolutional neural network and identifies and locates damage in offshore platform pipelines. This effectively overcomes the lack of prior experience in health monitoring of pipe bends on existing offshore platforms, improving the accuracy of health monitoring for existing offshore platforms. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a method for identifying the degree of damage at pipe bends based on wavelet packet transform in an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of pipe bend damage in the pipe bend damage identification method based on wavelet packet transform in an embodiment of the present invention.

[0047] Figure 3 This is a vibrational shape diagram of the strain mode data of the method for identifying the degree of damage at pipe bends based on wavelet packet transform in this embodiment of the invention.

[0048] Figure 4 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 15 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0049] Figure 5 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 20 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0050] Figure 6 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 30 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0051] Figure 7 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 40 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0052] Figure 8 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 50 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0053] Figure 9 The image shows the mode shape of the simulated strain mode with a signal-to-noise ratio of 60 for the wavelet packet transform-based method for identifying the degree of damage at pipe bends in this embodiment of the invention.

[0054] Figure 10 This is a graph of the node 0 coefficients of the method for identifying the degree of damage at pipe bends based on wavelet packet transform in an embodiment of the present invention.

[0055] Figure 11 This is a graph of the coefficients of node 127 in the wavelet packet transform-based method for identifying the degree of damage at pipe bends in an embodiment of the present invention.

[0056] Figure 12 In this embodiment of the invention, the wavelet packet transform decomposition coefficient of the method for identifying the degree of damage at pipe bends based on wavelet packet transform has 7 levels, and the decomposition object is... Figure 10 and Figure 11 Wavelet packet transform coefficient curves;

[0057] Figure 13 This is a noise simulation strain mode data curve of the method for identifying the degree of damage at pipe bends based on wavelet packet transform in an embodiment of the present invention;

[0058] Figure 14 This is a reconstructed simulated strain modal data curve of the method for identifying the degree of damage at pipe bends based on wavelet packet transform in an embodiment of the present invention;

[0059] Figure 15This is a damage depth loss map of the damage degree identification method at pipe bends based on wavelet packet transform in an embodiment of the present invention.

[0060] Figure 16 This is a schematic diagram of wavelet packet transform for the method of identifying the degree of damage at pipe bends based on wavelet packet transform in an embodiment of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The purpose of this invention is to provide a method and system for identifying the degree of damage at pipe bends based on wavelet packet transform. It extracts the features of the damage itself based on a one-dimensional convolutional neural network and identifies and locates the damage in the pipelines of offshore platforms. This effectively overcomes the lack of prior experience in health monitoring of pipe bends on existing offshore platforms and improves the accuracy of health monitoring of existing offshore platforms.

[0063] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment of the invention provides a method for identifying the degree of damage at pipe bends based on wavelet packet transform, including:

[0066] Step 101: Obtain strain mode data at the bend in the pipeline.

[0067] like Figure 3 The figure shown is a vibrational diagram of the strain modal data.

[0068] Step 102: Reconstruct the strain modal data to obtain the reconstructed strain modal data.

[0069] Step 103: As Figure 16 As shown, wavelet packet transform is performed on the reconstructed strain mode data to obtain N layers of wavelet packets.

[0070] In practice, the wavelet packet transform is implemented using the Pywt.WaveletPacket function in Python, with the mode set to symmetric. The Ribo3.1 series functions are selected as the wavelet packet basis functions, and the decomposition is performed at 7 levels, resulting in 128 frequency coefficients. The specific command is pywt.WaveletPacket(mode='symmetric',wavelet='rbio3.1',maxlevel=7).

[0071] Step 104: Input the last wavelet packet from the N-layer wavelet packet array into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend. The structural parameters of the trained one-dimensional convolutional neural network are shown in Table 1.

[0072] Table 1. Structure parameters of the trained one-dimensional convolutional neural network

[0073]

[0074] In Table 1, the training set accounts for 90% of the total samples, the test set accounts for 10%, the number of iterations is 1000, the loss function is the cross-entropy function, and the trained one-dimensional convolutional neural network is built on PyTorch, with the following runtime environment: Python 3.8, NumPy 1.23, SciPy 1.9.3, PyTorch 1.12, and CUDA 11.3. The hardware environment is: Intel(R) Xeon(R) Silver4210R CPU@2.40GHz, Nvidia 3090, 128GB DDR4 2400, batch size is 16, and the learning rate is 0.05. The damage depth recognition results are shown in Table 2.

[0075] Table 2 Results of damage depth identification

[0076] Table 6-4 Results of damage depth identification

[0077]

[0078] like Figure 15As shown in Figure 1 and Table 2, for 20 test samples, the trained one-dimensional convolutional neural network achieved a damage depth recognition accuracy of 90%, which is sufficient to complete the damage depth recognition task. For the samples with incorrect recognition, the errors were 8.3% and 11.5%, respectively. The relatively large errors are due to the fact that the present invention has a damage depth recognition accuracy of 0.1 mm, while the damage depth is affected by the wall thickness, ranging from 1 to 3 mm. The smaller denominator in the error calculation results in higher errors for the incorrect cases. However, for the two incorrect cases, the actual errors were only 0.1 mm and 0.3 mm, which is sufficient for practical applications.

[0079] The last wavelet packet from the N-layer wavelet packet array is input into the trained one-dimensional convolutional neural network to obtain the damage state at the bend in the pipeline. This process also includes:

[0080] Construct a damage model at the bend of the pipe.

[0081] Based on the actual engineering situation, establish as follows: Figure 2 The simulation model of the pipe bend shown is illustrated. A groove (Lmm × Dmm × Hmm) is cut into the inner wall of the pipe bend to simulate localized pipe damage. L represents the damage length (5mm), D represents the damage width (1mm), and H represents the damage depth.

[0082] Modal analysis was performed on the damage model at the pipe bend to obtain simulated strain modal data;

[0083] Diagram of damage at pipe bends is shown below. Figure 2 As shown, for Figure 2 Modal analysis was performed on a damaged pipeline model to obtain simulated strain modal data for each order. Considering that the damage occurred at the bend in the inner wall of the pipeline, while the strain modal characteristics were extracted from the outer wall, the damage initiation position S was represented by an angle. The specific location of the damage initiation point was derived using the circumference formula. The specific structural parameters of the pipeline were: bending radius R = 90 mm, bending angle α = 90 mm. Damage location and depth were selected as variables, and different damage conditions were simulated by changing their values. The range of S was [0°, 87.5°]; the range of damage depth H was [1 mm, 3 mm]. Combining these two values, S was taken at 2.5° intervals, and H at 0.5 mm intervals, resulting in 210 damage scenarios. Modal analysis was performed on all damage scenarios to obtain simulated strain modal data at the pipe bend under corresponding conditions. The acquired data was extracted using a method of taking a value every 10 mm to simulate the strain modal values ​​obtained by sensors in actual engineering.

[0084] like Figures 4-9The figure shows the mode shapes of the simulated strain modes at different signal-to-noise ratios. Noise simulation was performed on the simulated strain mode data, as shown below. Figure 13 As shown, noise simulation strain mode data were obtained.

[0085] like Figure 14 As shown, the noise-simulated strain mode data is reconstructed to obtain the reconstructed simulated strain mode data.

[0086] Because this invention strictly references the sensor placement conditions in real-world pipeline inspection projects for point selection, the strain mode data has a relatively short length. Noise simulation does not alter the length of the strain mode data, resulting in an insufficient number of wavelet packet transform layers. This further hinders the one-dimensional convolutional neural network from identifying damage features, severely impacting the final results. Therefore, interpolation is needed to reconstruct the noise-simulated strain modes without destroying data features, extending the data length to meet the needs of subsequent research. Since the noise-simulated strain mode images exhibit obvious irregular high-frequency fluctuations, linear interpolation cannot be used to reconstruct the noise-simulated strain mode data, as this would lead to data feature loss. This invention chooses spline interpolation to reconstruct the noise-simulated strain mode data. Spline interpolation uses a special piecewise polynomial interpolation form, called a spline, with cubic spline interpolation being the most common. The cubic spline function is the most basic and important spline function, and also the most widely used spline function in practice.

[0087] The specific implementation process uses the `splrep` function from the `interpolate` class in the SciPy library of Python. The number of reconstructed points is 3800. This function can find the one-dimensional spline curve representation of the noise-simulated strain modal data. The results are as follows... Figure 14 As shown, after spline interpolation, the noise simulation strain mode data becomes smoother.

[0088] In practice, noise-simulated strain mode data reconstruction can yield relatively smooth noise-simulated strain mode data. However, the value ranges of strain mode data of different orders are different. Therefore, it is necessary to normalize the strain mode data of different orders, mapping them to the range [0,1] while retaining the data features, to achieve a unified dimension and eliminate the influence of different value ranges of strain mode data of different orders on the training of convolutional neural networks.

[0089] Using formula Normalization is performed, ensuring that the maximum and minimum values ​​are not equal to avoid division by zero errors. The normalized data still maintains the same curve characteristics as the input data.

[0090] Where, m i ' represents the i-th value in the normalized simulation data; mi Let m be the i-th value in the reconstructed simulated strain modal data; min(m1,m2,m3,.....m n ) represents the minimum eigenvalue in the reconstructed simulated strain modal data; max(m1,m2,m3,.....m n ) represents the largest eigenvalue in the reconstructed simulated strain modal data; m n This is the last value in the reconstructed simulated strain modal data.

[0091] like Figure 16 As shown, wavelet packet transform is performed on the reconstructed simulated strain modal data to obtain N layers of simulated wavelet packets.

[0092] The data no longer reflects the location and severity of the damage. Therefore, this invention introduces time-frequency analysis to map the damage features from the no longer visible time domain to the frequency domain, and then extracts these features using a one-dimensional convolutional neural network. It is worth noting that, considering practical engineering scenarios, the noise-simulated strain modal data is treated as one-dimensional sequential data, and a series of coefficients with frequency domain characteristics are obtained through wavelet packet transform. The wavelet packet transform uses a set of orthogonal, aperiodic wavelet basis functions to decompose the noise-simulated strain modal data in the frequency domain. For example... Figure 16 As shown, wavelet packet transform performs multi-level and multiple time-frequency decompositions on the input data. Each wavelet packet transform can decompose the high-frequency and low-frequency components of the sequence data. The wavelet packet transform decomposes not only the low-frequency components but also the high-frequency components, and this decomposition is neither redundant nor incomplete. Therefore, it can perform better time-frequency localization analysis on signals containing a large amount of mid- and high-frequency information.

[0093] The noise-simulated strain mode data, after additive white Gaussian noise-spline interpolation data reconstruction-normalization data processing, is used as the time-domain signal in wavelet packet transform. Through layer-by-layer decomposition using wavelet packet transform, all frequency coefficients of the last layer are used as training data for a one-dimensional convolutional neural network. The number of decomposition layers in the wavelet packet transform is determined by the specific implementation method, and the number of results from the wavelet packet transform is determined by the specific number of decomposition layers.

[0094] At this point, the line graphs of some results show excessively prominent coefficients, so it's necessary to normalize the range of coefficients for the total nodes. Using the formula...

[0095] Among them, Z i 'This represents the last simulated wavelet packet after normalization; y i For the i-th value in the last simulated wavelet packet; min(z1,z2,z3,.....z n) represents the minimum eigenvalue in the last simulated wavelet packet; max(z1,z2,z3,.....z n ) represents the largest eigenvalue in the last simulated wavelet packet; z n This is the last value in the last simulated wavelet packet.

[0096] The last simulated wavelet packet in the N-layer simulated wavelet packet is used as training data and input into a one-dimensional convolutional neural network to train the one-dimensional convolutional neural network, thus obtaining the trained one-dimensional convolutional neural network.

[0097] A one-dimensional convolutional neural network (CNN) structure is established based on the one-dimensional characteristics of the preprocessed data. The CNN model of this invention employs overlapped max pooling during pooling; Dropout regularization is used in the final fully connected layer to avoid overfitting; and the RMSprop optimization algorithm is used for backpropagation parameter updates. Compared to traditional two-dimensional CNNs, the one-dimensional CNN can preserve the one-dimensional spatial characteristics of the damage strain modal data to the greatest extent. The computational flow of the one-dimensional CNN is as follows: The preprocessed pipe damage strain modal data is input into this network. After passing through convolutional layer 1, the data goes through max pooling layer 1, is activated by the ReLU function, and then enters convolutional layer 2 and max pooling layer 2, is activated again by the ReLU function, and enters three fully connected layers. The output channel of the fully connected layers is the pipe inner diameter thickness. The category with the highest probability is the damage depth. After multiple iterations, the model with the highest accuracy on the test set is selected as the optimal model, resulting in the trained one-dimensional CNN.

[0098] Noise simulation is performed on the simulated strain modal data to obtain noise-simulated strain modal data, specifically including:

[0099] In practical pipeline inspection projects, besides the inability to obtain continuous strain values ​​at pipe bends, the results are always affected by varying levels of noise, including systematic errors in measuring instruments and random errors caused by human intervention. Therefore, it is necessary to incorporate white noise into the strain modes to simulate the effects of noise variations in actual operating conditions. This invention uses additive white Gaussian noise (AWGN) to simulate noise in practical pipeline inspection projects. By superimposing white noise signals onto the simulated strain mode data obtained from numerical simulation software using the following white noise mathematical model, the simulation of noisy strain mode data in the actual engineering environment can be achieved.

[0100] Additive white Gaussian noise is used to simulate noise in actual engineering. The mathematical model is as follows:

[0101]

[0102] Where SNR is the signal-to-noise ratio; x is the simulated strain mode data; and n is the noise signal.

[0103] Noise simulation strain modal data:

[0104]

[0105] Where, x noise The noise is used to simulate strain mode data; N is the length of the simulated strain mode data; rand(N) is N random decimals in the range [0,1] that conform to a standard normal Gaussian distribution.

[0106] like Figures 4-9 The image shows the simulated strain mode images with SNR values ​​of 15, 20, 30, 40, 50, and 60. It can be seen that when the SNR is 20, a significant discrepancy in the simulated strain mode data at the damage location can be observed. When the SNR is greater than 50, it becomes difficult to distinguish the simulated strain mode image from the original strain mode image.

[0107] In practical applications, with SNR set to 10, the simulated strain modal data curves are no longer visible to the naked eye.

[0108] Optionally, the strain modal data can be reconstructed to obtain reconstructed strain modal data, specifically including:

[0109] The strain modal data are reconstructed using spline interpolation to obtain the reconstructed strain modal data.

[0110] Optionally, wavelet packet transform is performed on the reconstructed strain modal data to obtain N layers of wavelet packets, specifically including:

[0111] The reconstructed strain mode data is normalized to obtain normalized data; wavelet packet transform is then performed on the normalized data to obtain N layers of wavelet packets.

[0112] Optionally, the reconstructed strain modal data is normalized to obtain normalized data, as shown in the formula:

[0113]

[0114] Where, x i ' represents the i-th value in the normalized data; x i Let x be the i-th value in the reconstructed strain modal data; min(x1,x2,x3,.....x n ) represents the minimum eigenvalue in the reconstructed strain modal data; max(x1,x2,x3,.....x n ) represents the largest eigenvalue in the reconstructed strain modal data; x nThis is the last value in the reconstructed strain modal data.

[0115] The last wavelet packet from the N-layer wavelet packet array is input into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend, specifically including:

[0116] Normalize the last wavelet packet in the N-level wavelet packet array to obtain the normalized last wavelet packet, as shown in the formula:

[0117]

[0118] Among them, y i ' represents the i-th value in the last wavelet packet after normalization; y i The i-th value in the last wavelet packet; min(y1, y2, y3, ..., y n ) represents the minimum eigenvalue in the last wavelet packet; max(y1,y2,y3,.....y n ) represents the largest eigenvalue in the last wavelet packet; y n This is the last value in the last wavelet packet.

[0119] Example 2

[0120] This invention provides a system for identifying the degree of damage at pipe bends based on wavelet packet transform. This system is applied in Embodiment 1 and includes:

[0121] The data acquisition module is used to acquire strain modal data at pipe bends;

[0122] The data reconstruction module is used to reconstruct the strain modal data to obtain the reconstructed strain modal data;

[0123] The wavelet packet module is used to perform wavelet packet transformation on the reconstructed strain mode data to obtain N layers of wavelet packets.

[0124] The neural network module is used to input the last wavelet packet in the N-layer wavelet packet into the trained one-dimensional convolutional neural network to obtain the damage state at the bend of the pipe.

[0125] In one embodiment, the present invention also provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method in embodiment 1.

[0126] In one embodiment, the present invention also provides a computer-readable storage medium storing a computer program that, when executed, implements the method in embodiment 1.

[0127] The beneficial effects of this invention are:

[0128] Beneficial effect 1

[0129] Simulated strain modal data reconstruction can yield relatively smooth noisy simulated strain modal data. However, the value ranges of strain modal data of different orders are different, which can seriously affect the training of subsequent damage recognition networks. If only a specific order of strain modal data is selected for training, although an accurate result can be obtained, it cannot be well applied to real-world environments. A modal recognition process needs to be added, which is quite troublesome and complex. Therefore, this invention proposes to use simulated strain modal data of arbitrary orders for training a convolutional neural network for damage recognition, so as to better apply it in practical engineering.

[0130] Beneficial effect 2

[0131] After data reconstruction and normalization using additive white Gaussian noise, spline interpolation, and other processing methods, the data can no longer reflect information such as the location and severity of the damage. Therefore, this invention introduces time-frequency analysis to map the damage features from the no longer visible time domain to the frequency domain, and then extracts the damage features from the frequency domain using a convolutional neural network.

[0132] It is worth noting that this invention considers practical engineering scenarios, and the data here is not time-series data in the traditional sense. For ease of describing its core concepts, this invention treats the noise-simulated strain modal data as one-dimensional sequence data. The frequency domain transformation does not obtain a strictly defined frequency distribution through Fourier transform, but rather a series of coefficients with frequency domain characteristics through wavelet packet transform. Fourier transform fits a sequence signal using a set of trigonometric functions, but the frequency distribution obtained through Fourier transform only contains the distribution information of the sequence in the frequency domain and cannot obtain the position of the signal corresponding to different frequencies in the time domain. Therefore, if time-frequency analysis is performed using Fourier transform, the damage frequency characteristics will be lost. Wavelet packet transform can realize the correspondence between the time and frequency domains. Compared to Fourier transform, which uses a set of orthogonal trigonometric basis functions, wavelet packet transform uses a set of orthogonal, aperiodic wavelet basis functions to decompose the noise-simulated strain modal data in the frequency domain. It is important to note that the result of wavelet packet transform is to obtain frequency characteristics, not specific frequency values; it can be understood as dimensionless numbers similar to normalization. These wavelet basis functions can obtain the time position of data at different frequencies by using variables in scaling and translation.

[0133] Beneficial effect 3

[0134] like Figure 12 The decomposition coefficient is 7 levels, and the decomposition object is... Figure 10 and Figure 11 The wavelet packet transform coefficient diagram. Figure 10 For the coefficient graph of node 0 and Figure 11 The graph shows the coefficients for node 127. Node 0 is the most frequent node, and node 127 is the least frequent node. The total node graph contains frequency coefficient curves for 128 nodes. The line graph of node 0 shows that the most frequent coefficients retain a relatively clear original curve shape. The line graph of node 127 shows that the least frequent coefficients exhibit significant shape distortion. This shape distortion is considered to represent a certain feature, and the task of the convolutional neural network is to identify the shape distortion features representing damage. This is also the greatest advantage of using wavelet packet transform: it eliminates the limitations of traditional modality recognition, transforming into data-driven damage recognition. Figure 12 It can be seen that the value range of the curve at node 0 is much higher than that of other nodes, while the value range of the coefficients at node 127 is as small as 10⁻². Therefore, it is necessary to perform a normalization operation on the coefficients of each node before training the one-dimensional convolutional neural network. Normalizing the wavelet packet transform results is crucial during model training; without this operation, the neural network can hardly identify the depth of the damage, posing a significant challenge to the model training process.

[0135] Beneficial effect 4

[0136] Because the simulated strain modal images of noise exhibit obvious irregular high-frequency fluctuations, linear interpolation cannot be used to reconstruct the data, as this would lead to the loss of data features. This invention chooses spline interpolation to reconstruct the simulated strain modal data of noise. This is because low-order spline interpolation is a convexity-preserving operation, which is significant for retaining high-frequency data features in the simulated strain modal data of noise. Furthermore, low-order spline interpolation can produce an effect similar to high-order polynomial interpolation and can avoid the Runge phenomenon. After spline interpolation, the simulated strain modal data of noise becomes smoother, and some sharp high-frequency noise is blurred; however, at this point, the abrupt changes in strain modes at the damage site are no longer observable. Here, this invention considers the damage features to transform from an explicit value domain to a implicit feature domain. A single data point cannot reflect a common feature domain; multiple data points must undergo wavelet packet transformation, and then a one-dimensional neural network is used to extract features from the wavelet packet transformation results in order to extract the damage feature domain.

[0137] Beneficial effects 5

[0138] Wavelet transform and wavelet packet transform can decompose the sequence data into high-frequency and low-frequency components. When wavelet transform decomposes an existing signal, it transforms the signal into low-frequency and high-frequency components (details). However, wavelet transform only further decomposes the low-frequency component, leaving the high-frequency component undecomposed. Therefore, wavelet transform is good at representing a large class of signals with low-frequency information as the main component, but it cannot well decompose and represent signals containing a large amount of detailed information (fine edges or textures). Wavelet packet transform decomposes not only the low-frequency component but also the high-frequency component, and this decomposition is neither redundant nor incomplete. Therefore, it can perform better time-frequency localization analysis on signals containing a large amount of mid- and high-frequency information. Figure 16 As shown, when using wavelet transform, only the black connecting lines (low-frequency part) are decomposed, while the dashed connecting lines (high-frequency part) are not decomposed. When using wavelet packet transform, both solid connecting lines (low-frequency part) and dashed connecting lines (high-frequency part) are decomposed.

[0139] Compared to wavelet transform, which decomposes only the low-frequency components obtained from the previous decomposition, wavelet packet transform decomposes both the low-frequency and high-frequency components obtained from the previous decomposition. This is why this invention chooses wavelet packet transform. First, since the distribution of damage features in the frequency domain is uncertain, and wavelet transform only decomposes low-frequency components, it is not suitable. Second, after wavelet packet transform, the noise-simulated strain mode data becomes 2^n frequency domain coefficients, which are naturally suitable for training one-dimensional convolutional neural networks. Finally, since one-dimensional convolutional neural networks extract features from one-dimensional sequence data, feature information at the same location can enter the same convolution kernel. This is the key reason why frequency domain features are mapped to time domain features.

[0140] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0141] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying the degree of damage at pipe bends based on wavelet packet transform, characterized in that, The method for identifying the degree of damage at pipe bends based on wavelet packet transform includes: Obtain strain modal data at the pipe bend; The strain modal data are reconstructed to obtain reconstructed strain modal data; Wavelet packet transform is performed on the reconstructed strain modal data to obtain N layers of wavelet packets; The last wavelet packet in the N-layer wavelet packet is input into the trained one-dimensional convolutional neural network to obtain the damage state at the bend of the pipeline.

2. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 1, characterized in that, The last wavelet packet in the N-layer wavelet packet sequence is input into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend. This process also includes: Construct a damage model at pipe bends; Modal analysis was performed on the damage model at the bend of the pipeline to obtain simulated strain modal data; Noise simulation is performed on the simulated strain mode data to obtain noise-simulated strain mode data; The noise-simulated strain mode data is reconstructed to obtain reconstructed simulated strain mode data; Wavelet packet transform is performed on the reconstructed simulated strain modal data to obtain N layers of simulated wavelet packets; The last simulated wavelet packet in the N-layer simulated wavelet packet is used as training data and input into a one-dimensional convolutional neural network to train the one-dimensional convolutional neural network, thus obtaining the trained one-dimensional convolutional neural network.

3. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 2, characterized in that, The simulated strain modal data is subjected to noise simulation to obtain noise-simulated strain modal data, specifically including: Additive white Gaussian noise is used to simulate noise in actual engineering. The mathematical model is as follows: The noise-simulated strain mode data: Where SNR is the signal-to-noise ratio; x is the simulated strain modal data; n is the noise signal; x noise The noise is used to simulate strain mode data; N is the length of the simulated strain mode data; rand(N) is N random decimals in the range [0,1] that conform to a standard normal Gaussian distribution.

4. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 1, characterized in that, The strain modal data is reconstructed to obtain reconstructed strain modal data, specifically including: The strain modal data are reconstructed using spline interpolation to obtain the reconstructed strain modal data.

5. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 1, characterized in that, The reconstructed strain modal data is subjected to wavelet packet transform to obtain N layers of wavelet packets, specifically including: The reconstructed strain modal data are normalized to obtain normalized data; The normalized data is subjected to wavelet packet transform to obtain the N-layer wavelet packets.

6. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 5, characterized in that, The reconstructed strain modal data are normalized to obtain the normalized data, as shown in the formula: Where, x i ' represents the i-th value in the normalized data; x i Let x be the i-th value in the reconstructed strain modal data; min(x1,x2,x3,.....x n ) represents the minimum eigenvalue in the reconstructed strain modal data; max(x1,x2,x3,.....x n ) represents the largest eigenvalue in the reconstructed strain modal data; x n This is the last value in the reconstructed strain modal data.

7. The method for identifying the degree of damage at pipe bends based on wavelet packet transform according to claim 1, characterized in that, The last wavelet packet in the N-layer wavelet packet sequence is input into the trained one-dimensional convolutional neural network to obtain the damage state at the pipe bend, specifically including: The last wavelet packet in the N-layer wavelet packet is normalized to obtain the normalized last wavelet packet, as shown in the formula: Among them, y i ' represents the i-th value in the last wavelet packet after normalization; y i The i-th value in the last wavelet packet; min(y1, y2, y3, ..., y n ) represents the minimum eigenvalue in the last wavelet packet; max(y1,y2,y3,.....y n ) represents the largest eigenvalue in the last wavelet packet; y n This is the last value in the last wavelet packet.

8. A system for identifying the degree of damage at pipe bends based on wavelet packet transform, characterized in that, The wavelet packet transform-based pipe bend damage identification system is applied to the method described in any one of claims 1-7, wherein the wavelet packet transform-based pipe bend damage identification system comprises: The data acquisition module is used to acquire strain modal data at pipe bends; The data reconstruction module is used to reconstruct the strain mode data to obtain reconstructed strain mode data; The wavelet packet module is used to perform wavelet packet transformation on the reconstructed strain mode data to obtain N layers of wavelet packets. The neural network module is used to input the last wavelet packet in the N-layer wavelet packet into the trained one-dimensional convolutional neural network to obtain the damage state at the bend of the pipeline.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed, implements the method as described in any one of claims 1 to 7.

Images