A multi-channel detection method for end-to-end steel oil and gas pipeline cracks

Abstract

The present application relates to the technical field of testing or analyzing materials by measuring the chemical or physical properties of materials, and specifically to an end-to-end multi-channel detection method for cracks in steel oil and gas pipelines, which is executed by a processor and includes: obtaining a detection signal of a crack in a target pipeline; obtaining a crack defect signal set; performing data enhancement on the signal in the crack defect signal set to obtain an enhanced data set; performing signal rasterization on the enhanced data set to obtain a raster image data set; inputting the data in the raster image data set into SSD model training to obtain a crack defect automatic annotation model; configuring the hyperparameters of a neural network based on a genetic algorithm; inputting the crack defect signal set into DRSN2d model training, and obtaining the best optimization model in combination with the hyperparameters. The present application solves the problems of high dependence on manual labor, low detection accuracy, and low quantization speed in the prior art when detecting cracks inside pipelines.

Description

Technical Field

[0001] The present application relates to the technical field of testing or analyzing materials by means of determining the chemical or physical properties of the materials, and in particular, to a multi-channel detection method for end-to-end cracks in steel oil and gas pipelines. Background Art

[0002] In the detection work inside oil and gas pipelines, cracks are the main failure mode of oil and gas pipelines. The application of steel pipeline crack detection technology can effectively ensure the safety of pipelines. After obtaining the crack detection signal, the existing signal analysis work is generally divided into three steps: manual defect labeling, manual feature extraction, and classifier classification or quantification algorithm quantification. However, since the manual defect labeling and manual feature extraction steps rely on the professional knowledge and working status of practitioners, the amount of information, consistency and efficiency of feature extraction are limited. On the other hand, in the existing technology, the detection and signal quantification of cracks inside the pipeline have high professional requirements for researchers. Traditional signal features cannot fully characterize the actual signal, and the quantification accuracy of tiny cracks is limited. For the above problems, it is urgent to propose a pipeline crack detection signal quantification method with less dependence on manual work, higher accuracy and higher quantification efficiency. Summary of the invention

[0003] The purpose of this application is to provide an end-to-end multi-channel detection method for cracks in steel oil and gas pipelines, which solves the problems of high dependence on manual labor, low detection accuracy and low quantization speed in the prior art for internal crack detection in pipelines.

[0004] The technical solution of this application:

[0005] The present application provides an end-to-end multi-channel detection method for cracks in a steel oil and gas pipeline, which is executed by a processor and is characterized by comprising:

[0006] Acquire detection signals of cracks in the target pipeline;

[0007] Manually annotate the crack defect signals in the detection signal to obtain a crack defect signal set;

[0008] Performing data enhancement on the signals in the crack defect signal set to obtain an enhanced data set;

[0009] Perform signal rasterization on the enhanced data set to obtain a raster image data set;

[0010] The data in the raster image dataset is input into the SSD model training to obtain the automatic crack defect annotation model;

[0011] Configure the hyperparameters of neural networks based on genetic algorithms;

[0012] The crack defect signal set is input into the DRSN2d model training, and the best optimization model is obtained by matching the hyperparameters.

[0013] Furthermore, the signal rasterization is performed on the enhanced data set to obtain a raster image data set, specifically: a raster image and a signal scatter plot are obtained according to the detection signal, the length of the raster image pixel is matched with the interval of the points of the signal scatter plot on the X-axis, and the ratio of the raster image pixel value to the interval of the signal sampling point is recorded to obtain the raster image data.

[0014] Furthermore, the data in the raster image data set is input into the SSD model training to obtain the crack defect automatic annotation model, specifically:

[0015] Divide the raster image dataset into training set, validation set and test set;

[0016] Input the training set into the SSD model for training, and input the validation set into the SSD model for evaluation;

[0017] The test set is input into the evaluated SSD model, and the performance of the model is evaluated to obtain the automatic defect annotation model.

[0018] Furthermore, the training set is input into the SSD model training, specifically:

[0019] Input the training set into the SSD model training to obtain the bounding box position dataset, category dataset, and confidence dataset;

[0020] Calculate the deviation data values ​​of the bounding box position dataset, category dataset, confidence dataset and the training set;

[0021] Based on the SSD model, in conjunction with the deviation data value, loss value calculation, error feedback and parameter optimization are performed in sequence.

[0022] Furthermore, the step of inputting the test set into the evaluated SSD model, performing performance evaluation on the model, and obtaining the automatic defect annotation model further includes:

[0023] Get verification signal;

[0024] Input the verification signal into the defect automatic annotation model to obtain the verification signal annotation data;

[0025] Based on the verification signal annotation data, the crack defect signal range data of the verification signal is calculated.

[0026] Furthermore, the hyperparameters of the neural network configured based on the genetic algorithm are specifically:

[0027] Determine the hyperparameters that need to be optimized and encode them to form individuals represented by chromosomes;

[0028] Randomly generate the initial population, each individual represents a set of hyperparameter combinations;

[0029] Evaluate the performance of the model corresponding to each individual on the validation set based on the fitness function;

[0030] Based on genetic operations, the next generation population is obtained;

[0031] Based on the crossover operation, part of the chromosomes of two individuals are exchanged to obtain a new individual;

[0032] Perform random changes to some genes of an individual based on the mutation operation.

[0033] Furthermore, the optimal optimization model includes a crack length quantification model, a crack width quantification model, and a crack depth quantification model. Based on the crack length quantification model, the crack width quantification model, and the crack depth quantification model, the length quantification value, the width quantification value, and the depth quantification value are respectively obtained.

[0034] Furthermore, the process of obtaining the depth quantization value and the length quantization value is specifically as follows:

[0035] Input the detection signal into the dimension reduction layer to obtain a one-dimensional signal;

[0036] Input the one-dimensional signal into the pooling layer for filtering;

[0037] The filtered signals are sequentially input into three one-dimensional convolutional layers of the crack depth quantification model to extract the depth features;

[0038] converting the filtered signal into two-dimensional data;

[0039] The two-dimensional data are sequentially input into two two-dimensional convolution layers of the crack length quantification model to extract and obtain the two-dimensional length feature parameters;

[0040] Convert two-dimensional length feature parameters into one-dimensional length data;

[0041] The converted one-dimensional length data is input into the one-dimensional convolution layer of the crack length quantification model to extract the length feature;

[0042] The depth features and length features are input into five layers of RSBU-CW with soft threshold denoising function for feature denoising and feature extraction;

[0043] Set Dropout between the last two layers of the channel number;

[0044] The filtered signal is input into the Dropout and BN layers, and then into the one-dimensional convolution to achieve the final feature extraction;

[0045] Flatten the extracted depth features and length features to obtain one-dimensional depth features and one-dimensional length features;

[0046] The one-dimensional depth feature and the one-dimensional length feature are input into two-layer linear layer calculation to obtain the length quantization value and depth quantization value of the crack.

[0047] Furthermore, the process of obtaining the width quantization value is specifically as follows:

[0048] Get defect-free signals;

[0049] Extracting crack defect width edge data based on detection signals;

[0050] Input the crack defect width edge data and defect-free signal into the DRSN2d model training to obtain the crack width quantification model;

[0051] Based on an upper edge channel signal segment, a lower edge channel signal segment and a signal reference value signal segment;

[0052] The crack defect width edge data is used to obtain upper edge channel signal segment data, lower edge channel signal segment data, and signal reference value signal segment data;

[0053] Input the upper edge channel signal segment data, the lower edge channel signal segment data, and the signal reference value signal segment data into the crack width quantification model to obtain the width quantification value ,in:

[0054]

[0055] Where: is the true width of the crack defect, is the multi-channel sensor spacing.

[0056] Furthermore, it also includes:

[0057] Pruning, quantization and low-rank decomposition are performed on the crack length quantization model, the crack width quantization model and the crack depth quantization model. The technical solution of the present application has at least the following advantages and beneficial effects:

[0058] The present application provides an end-to-end multi-channel detection method for cracks in steel oil and gas pipelines, wherein the end-to-end method refers to directly inputting the original detection signal into the model, and the model output value is the depth quantization value, the length quantization value or the width quantization value, and no other processes such as manual feature extraction and manual filtering and noise reduction are required in the middle. By using the features automatically extracted by model training to quantify defects, the shortcomings of manually extracted features are avoided. Automatic feature extraction makes it possible to more comprehensively extract information from the detection signal, which can not only reduce dependence on manual work, but also improve the detection accuracy and the quantization speed of quantizing the depth, length and width. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of the detection method of this application;

[0060] Figure 2 A flow chart for obtaining the quantified values ​​of the length, width and depth of cracks in a target pipeline using the best optimization model in the detection method of this application;

[0061] Figure 3 A flowchart of filtering a one-dimensional signal input into a pooling layer in the detection method of this application;

[0062] Figure 4 A flow chart of converting a filtered signal into two-dimensional data in the detection method of the present application;

[0063] Figure 5 This is a flow chart for feature denoising and feature extraction in the detection method of this application. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0065] Example

[0066] Please refer to Figure 1-Figure 5 The present application provides an end-to-end multi-channel detection method for cracks in a steel oil and gas pipeline, the method being executed by a processor and comprising:

[0067] Acquire detection signals of cracks in the target pipeline;

[0068] Manually annotate the crack defect signals in the detection signal to obtain a crack defect signal set;

[0069] Performing data enhancement on the signals in the crack defect signal set to obtain an enhanced data set;

[0070] Perform signal rasterization on the enhanced data set to obtain a raster image data set;

[0071] The data in the raster image dataset is input into the SSD model training to obtain the automatic crack defect annotation model;

[0072] Configure the hyperparameters of neural networks based on genetic algorithms;

[0073] The crack defect signal set is input into the DRSN2d model training, and the best optimization model is obtained by matching the hyperparameters.

[0074] It is worth noting that the detection method provided in this embodiment, the end-to-end means directly inputting the original detection signal into the model, and the model output value is the depth quantization value, the length quantization value or the width quantization value. There is no need for manual feature extraction, manual filtering and noise reduction and other processes in the middle. Specifically, defects are quantified by using the features automatically extracted by model training, which avoids the shortcomings of manual feature extraction. Automatic feature extraction makes it possible to more comprehensively extract information from the detection signal, which not only reduces dependence on manual labor, but also improves the detection accuracy and the quantization speed of quantizing depth, length and width, solving the problems of high dependence on manual labor, low detection accuracy and low quantization speed in the prior art for internal crack detection in pipelines.

[0075] It should be noted that a differential eddy current detection technology based on incremental magnetic permeability extraction is used to build a target pipeline detection test platform in the laboratory, and on this basis, cracks in the target pipeline are detected, specifically using probes for detection.

[0076] It should be noted that the single-channel detection used in the prior art uses only one sensor for detection. Due to the small coverage of the sensor, its detection effect is severely limited. The multi-channel detection used in this embodiment has a larger coverage area and can obtain a higher signal-to-noise ratio after correction, which has a better detection effect. On the other hand, for crack detection of steel oil and gas pipelines, in terms of defect labeling, compared with the prior art that requires personnel with higher professional capabilities to complete, this embodiment is based on the SSD model to obtain the best optimization model. The end-to-end model in the field of deep learning can reduce the requirements for people's prior knowledge and has powerful nonlinear characterization capabilities.

[0077] It should be noted that the detection signal of the crack of the target pipeline is obtained by scanning the crack of the target pipeline under the conditions of fixed lift-off value and fixed parameters to obtain the detection signal. In this embodiment, 204 detection signals are obtained. Optionally, other lightweight electromagnetic detection methods can also be used to obtain the detection signal.

[0078] It should be noted that data enhancement is performed on the signals in the crack defect signal set to obtain an enhanced data set. Specifically, data enhancement is performed on the defect signals in the crack defect signal set. The enhancement method is: on the premise of including the defect, the defect signal range is shifted left or right along the sampling point direction / time direction, and the movement cannot exceed the actual defect signal segment, thereby increasing the data volume on the premise of including the defect signal segment. In this embodiment, 7194 defect signals containing 16384 sampling points are obtained after enhancement, which is the enhanced data set.

[0079] In some embodiments, labelme is used to manually annotate defects in the raster images in the first dataset.

[0080] It should be noted that the best in this embodiment is to use the machine learning evaluation index R 2 , RMSE, etc. to calculate the gap between the "model quantization value" and the "true quantization value", measure the model quantization effect through indicators, and select the best optimization model.

[0081] Furthermore, the signal rasterization is performed on the enhanced data set to obtain a raster image data set, specifically: a raster image and a signal scatter plot are obtained according to the detection signal, the length of the raster image pixel is matched with the interval of the points of the signal scatter plot on the X-axis, and the ratio of the raster image pixel value to the interval of the signal sampling point is recorded to obtain the raster image data.

[0082] In detail, by using the matplotlib library in python to draw the signal grid image, irrelevant information such as the coordinate axis is removed, so that the signal shape itself starts from the left side of the image, and the cutoff length of the right side of all signals is fixed, that is, the side length of the grid image and the signal length of the signal scatter plot can correspond. Then, the signal reference value of each channel is made to correspond to the other side length of the grid image. Then, the image is saved at a higher resolution. Next, the image is scaled by a specified multiple, and the correspondence between the side length of the grid image and the signal length and the signal reference value of each channel is simplified while retaining the clarity of the image, and this correspondence is recorded.

[0083] Furthermore, the data in the raster image data set is input into the SSD model training to obtain the crack defect automatic annotation model, specifically:

[0084] Divide the raster image dataset into training set, validation set and test set;

[0085] Input the training set into the SSD model for training, and input the validation set into the SSD model for evaluation;

[0086] The test set is input into the evaluated SSD model, and the performance of the model is evaluated to obtain the automatic defect annotation model.

[0087] In detail, the raster image data set is randomly divided into a training set, a validation set, and a test set, preferably in a ratio of 8:1:1, that is, the 7194 raster images obtained in this embodiment are divided into 5755:719:720. Then, the training set is input into the SSD model for training. Then, after the SSD model converges, the performance of the model is evaluated using the test set, and finally a defect automatic annotation model is obtained, which is used to realize the automatic annotation of multi-channel crack defect signals.

[0088] Furthermore, the training set is input into the SSD model training, specifically:

[0089] Input the training set into the SSD model training to obtain the bounding box position dataset, category dataset, and confidence dataset;

[0090] Calculate the deviation data values ​​of the bounding box position dataset, category dataset, confidence dataset and the training set;

[0091] Based on the SSD model, in conjunction with the deviation data value, loss value calculation, error feedback and parameter optimization are performed in sequence.

[0092] In detail, after the training set is input into the SSD model for training, the training output data includes: (1) bounding box position data set, which refers to the bounding box position data of the image normalized relative to the size of the input raster image; (2) category data; (3) confidence data. According to the deviation between the output data of the SSD model during the training process and the true value of the training set, that is, the deviation data value between the bounding box position data set and the training set, the deviation data value between the category data set and the training set, and the deviation data value between the confidence data set and the training set, a deviation data set is formed. Then, based on the SSD default settings and the deviation data set, loss value calculation, error feedback and parameter optimization are performed, and the model is evaluated using the validation set during the model training process.

[0093] Furthermore, the step of inputting the test set into the evaluated SSD model, performing performance evaluation on the model, and obtaining the automatic defect annotation model further includes:

[0094] Get verification signal;

[0095] Input the verification signal into the defect automatic annotation model to obtain the verification signal annotation data;

[0096] Based on the verification signal annotation data, the crack defect signal range data of the verification signal is calculated.

[0097] In detail, the crack defect signal range data, the range of which refers to the defect start and end points in time units, and the data also includes the channel number of the defect.

[0098] Furthermore, the hyperparameters of the neural network configured based on the genetic algorithm are specifically:

[0099] Determine the hyperparameters that need to be optimized and encode them to form individuals represented by chromosomes, where the hyperparameters that need to be optimized are at least the number of network layers, the number of neurons in each layer, the learning rate, and the activation function;

[0100] Randomly generate the initial population, each individual represents a set of hyperparameter combinations;

[0101] Evaluate the performance of the model corresponding to each individual on the validation set based on the fitness function, where mean square error, accuracy, etc. are used as fitness values;

[0102] Based on genetic operations, a next generation population is obtained, wherein the genetic operations at least include selection, crossover and mutation;

[0103] Based on the crossover operation, part of the chromosomes of two individuals are exchanged to obtain a new individual;

[0104] Based on the mutation operation, random changes are performed on some genes of individuals to increase the diversity of the population.

[0105] It should be noted that the above-mentioned evaluation and genetic operation process is repeated until the preset number of iterations is reached or the termination condition is met, and the hyperparameter combination corresponding to the optimal individual finally obtained is the optimized result.

[0106] Furthermore, the optimal optimization model includes a crack length quantification model, a crack width quantification model, and a crack depth quantification model. Based on the crack length quantification model, the crack width quantification model, and the crack depth quantification model, the length quantification value, the width quantification value, and the depth quantification value are respectively obtained.

[0107] Furthermore, the process of obtaining the depth quantization value and the length quantization value is specifically as follows:

[0108] Input the detection signal into the dimension reduction layer to obtain a one-dimensional signal;

[0109] Input the one-dimensional signal into the pooling layer for filtering;

[0110] The filtered signals are sequentially input into three one-dimensional convolutional layers of the crack depth quantification model to extract the depth features;

[0111] converting the filtered signal into two-dimensional data;

[0112] The two-dimensional data are sequentially input into two two-dimensional convolution layers of the crack length quantification model to extract and obtain the two-dimensional length feature parameters;

[0113] Convert two-dimensional length feature parameters into one-dimensional length data;

[0114] The converted one-dimensional length data is input into the one-dimensional convolution layer of the crack length quantification model to extract the length feature;

[0115] The depth features and length features are input into five layers of RSBU-CW with soft threshold denoising function for feature denoising and feature extraction;

[0116] Set Dropout between the last two layers of the channel number;

[0117] The filtered signal is input into the Dropout and BN layers, and then into the one-dimensional convolution to achieve the final feature extraction;

[0118] Flatten the extracted depth features and length features to obtain one-dimensional depth features and one-dimensional length features;

[0119] The one-dimensional depth feature and the one-dimensional length feature are input into two-layer linear layer calculation to obtain the length quantization value and depth quantization value of the crack.

[0120] It is worth mentioning that please refer to Figure 2 , Figure 2 The process of obtaining the quantitative values ​​of the length, width and depth of the crack in the target pipe for the best optimization model. Please refer to Figure 3 , Figure 3 The process of filtering a one-dimensional signal input into the pooling layer. Please refer to Figure 4 , Figure 4 It is the process of converting the filtered signal into two-dimensional data. Please refer to Figure 5 , Figure 5 It is the process of feature denoising and feature extraction for depth features, length features, and width features. Dropout is set between the last two layers with a large number of channels to prevent overfitting.

[0121] Furthermore, the process of obtaining the width quantization value is specifically as follows:

[0122] Get defect-free signals;

[0123] Extracting crack defect width edge data based on detection signals;

[0124] Input the crack defect width edge data and defect-free signal into the DRSN2d model training to obtain the crack width quantification model;

[0125] Based on the upper edge channel signal segment, the lower edge channel signal segment and the signal reference value signal segment

[0126] The crack defect width edge data is used to obtain upper edge channel signal segment data, lower edge channel signal segment data, and signal reference value signal segment data;

[0127] Input the upper edge channel signal segment data, the lower edge channel signal segment data, and the signal reference value signal segment data into the crack width quantification model to obtain the width quantification value ,in:

[0128]

[0129] Where: is the true width of the crack defect, is the multi-channel sensor spacing.

[0130] It should be noted that when the number of defect signal channels is less than 2, two edge signal segments and two signal reference value signal segments are input. Calculate the predicted length :

[0131]

[0132] Furthermore, the multi-channel detection method provided in this embodiment also includes: performing pruning, quantization and low-rank decomposition on the crack length quantization model, the crack width quantization model and the crack depth quantization model to achieve model compression and acceleration, and obtain the optimal crack length quantization model, the optimal crack width quantization model and the optimal crack depth quantization model with higher quantization accuracy and faster quantization speed. Among them, pruning is performed based on amplitude, mainly because parameters with smaller absolute values ​​are considered to contribute less to the model and can therefore be pruned. The specific steps are:

[0133] Calculate the absolute value of each weight in the crack length quantification model, the crack width quantification model, and the crack depth quantification model;

[0134] Determine pruning thresholds: Set an array of percentile values ​​of parameter distribution, such as 99%, 98%, 97%, 96%, and 95%, and calculate and determine multiple initial pruning thresholds based on the percentile values ​​in the array;

[0135] Check the performance of the crack length quantization model, crack width quantization model, and crack depth quantization model after pruning: quantize based on all detection signals and check the performance of the model: if the model performance has dropped significantly, for example, the quantization accuracy has dropped by more than 4%, then find a smaller threshold; if the model performance has dropped slightly, for example, the quantization accuracy has dropped by less than or equal to 4%, then find a larger threshold; after the above search, obtain the threshold interval with the most pruning and the quantization accuracy drop that meets the requirements;

[0136] Based on the binary search, a secondary search is performed on the threshold interval, and the micro-crack test data is quantified, and the model performance is checked twice: if the model performance drops significantly, for example, the quantization accuracy drops by more than 2.5%, a smaller threshold is searched; if the model performance drops slightly, for example, the quantization accuracy drops by less than or equal to 2.5%, a larger threshold is searched; after the above search, the threshold with the most pruning and the drop in quantization accuracy meeting the requirements is obtained; the above processing is due to the difficulty in quantizing micro-cracks. In order to ensure the quantization of micro-cracks, the quantization results of micro-cracks are selected as the measurement standard in turn;

[0137] Fine-tune the crack length quantization model, crack width quantization model, and crack depth quantization model: After pruning, fine-tune the remaining parameters and train the model again using a small learning rate and training data so that the remaining parameters can adapt to the structural changes and restore the performance of the model;

[0138] Based on the crack length quantification model, the crack width quantification model and the crack depth quantification model, the target quantification range and the target quantification accuracy are determined;

[0139] The original range is proportionally mapped to the target quantization range to obtain low-precision parameters: the purpose is to reduce the storage space requirements and computational complexity of the model while minimizing the loss of model performance;

[0140] Based on the singular value decomposition method and model performance, the rank of the low-rank matrix after decomposition of the crack length quantization model, the crack width quantization model, and the crack depth quantization model is determined: the purpose is to reduce the amount of storage and calculation, and the singular value decomposition method is preferred;

[0141] Perform approximation on the factorized matrix.

[0142] It should be noted that the deep residual shrinkage network (DRSN) is an end-to-end denoising and feature extraction neural network applied to one-dimensional time-domain signals containing noise. The deep learning method has stronger representation capabilities due to its complex network structure and can learn richer features, but it also brings a large amount of computation. In practical applications, it may face limitations in computing resources and time. Therefore, to solve this problem, this embodiment prunes, quantizes, and low-rank decomposition the model, thereby simplifying the model, removing redundant connections and parameters, reducing the amount of computation while maintaining a certain performance, and making the deep learning model more efficient and practical.

[0143] Furthermore, in order to ensure the quantitative performance of the model, fine-tuning of the model parameters may be performed again after the above processing.

[0144] It is worth noting that in the prior art, the model composed of a convolutional neural network and a fully connected network for diagnosing cracks in a compressed structure needs to be considered to input a specific signal in order to detect cracks, and the diagnosis of compression cracks in aluminum samples is performed, but for the defects of cracks inside the pipeline, detection and quantification cannot be achieved, and the operation process requires more complex human input. In the prior art, there is also a method of training a convolutional neural network by using finite element simulation to simulate the three-dimensional crack detection data inside the solid interior of a rectangular cuboid steel, which realizes accurate prediction of crack characteristics from experimentally measured ultrasonic non-destructive testing signals, but this process is aimed at the interior of a rectangular cuboid steel and cannot be directly and effectively used in the detection of cracks inside a circular pipeline. The detection method provided in the present application is aimed at a hollow pipeline, and the defect signal segment inside the pipeline is automatically annotated by generating a defect automatic annotation model, and the length, width and depth of the cracks in the pipeline are quantified by the generated optimal optimization model, and the length, width and depth of the cracks are quantified, and the length, width and depth of the cracks are obtained. No human signal input operation is required, and the detection process is more efficient, which can realize the detection of cracks inside the pipeline and the quantification of length, width and depth.

[0145] The above description is only a preferred implementation mode of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A multi-channel detection method for end-to-end cracks in steel oil and gas pipelines, the method being executed by a processor, characterized in that: include: Acquire detection signals of cracks in the target pipeline; Manually annotate the crack defect signals in the detection signal to obtain a crack defect signal set; Performing data enhancement on the signals in the crack defect signal set to obtain an enhanced data set; Perform signal rasterization on the enhanced data set to obtain a raster image data set; The data in the raster image dataset is input into the SSD model training to obtain the automatic crack defect annotation model, specifically: Divide the raster image dataset into training set, validation set and test set; Input the training set into the SSD model for training, and input the validation set into the SSD model for evaluation; Input the test set into the evaluated SSD model, perform performance evaluation on the model, and obtain the automatic defect annotation model, which also includes: Get verification signal; Input the verification signal into the defect automatic annotation model to obtain the verification signal annotation data; Based on the verification signal annotation data, calculate the crack defect signal range data of the verification signal; The hyperparameters of the neural network are configured based on the genetic algorithm, specifically: Determine the hyperparameters that need to be optimized and encode them to form individuals represented by chromosomes; Randomly generate the initial population, each individual represents a set of hyperparameter combinations; Evaluate the performance of the model corresponding to each individual on the validation set based on the fitness function; Based on genetic operations, the next generation population is obtained; Based on the crossover operation, part of the chromosomes of two individuals are exchanged to obtain a new individual; Perform random changes on some genes of an individual based on mutation operations; The crack defect signal set is input into the DRSN2d model training, and the optimal optimization model is obtained by matching the hyperparameters. The optimal optimization model includes a crack length quantization model, a crack width quantization model, and a crack depth quantization model. The length quantization value, the width quantization value, and the depth quantization value are obtained based on the crack length quantization model, the crack width quantization model, and the crack depth quantization model, respectively; wherein, the process of obtaining the depth quantization value and the length quantization value is specifically as follows: Input the detection signal into the dimension reduction layer to obtain a one-dimensional signal; Input the one-dimensional signal into the pooling layer for filtering; The filtered signals are sequentially input into three one-dimensional convolutional layers of the crack depth quantification model to extract the depth features; converting the filtered signal into two-dimensional data; The two-dimensional data are sequentially input into two two-dimensional convolution layers of the crack length quantification model to extract and obtain the two-dimensional length feature parameters; Convert two-dimensional length feature parameters into one-dimensional length data; The converted one-dimensional length data is input into the one-dimensional convolution layer of the crack length quantification model to extract the length feature; The depth features and length features are input into five layers of RSBU-CW with soft threshold denoising function for feature denoising and feature extraction; Set Dropout between the last two layers of the channel number; The filtered signal is input into the Dropout and BN layers, and then into the one-dimensional convolution to achieve the final feature extraction; Flatten the extracted depth features and length features to obtain one-dimensional depth features and one-dimensional length features; The one-dimensional depth feature and the one-dimensional length feature are input into two-layer linear layer calculation to obtain the length quantization value and depth quantization value of the crack.

2. The multi-channel detection method according to claim 1, characterized in that: The signal rasterization is performed on the enhanced data set to obtain the raster image data set, specifically: a raster image and a signal scatter plot are obtained according to the detection signal, the length of the raster image pixel is matched with the interval of the points of the signal scatter plot on the X-axis, and the ratio of the raster image pixel value to the interval of the signal sampling point is recorded to obtain the raster image data.

3. The multi-channel detection method according to claim 1, characterized in that: The training set is input into the SSD model training, specifically: Input the training set into the SSD model training to obtain the bounding box position dataset, category dataset, and confidence dataset; Calculate the deviation data values ​​of the bounding box position dataset, category dataset, confidence dataset and the training set; Based on the SSD model, in conjunction with the deviation data value, loss value calculation, error feedback and parameter optimization are performed in sequence.

4. The multi-channel detection method according to claim 1, characterized in that: The process of obtaining the width quantization value is as follows: Get defect-free signals; Extracting crack defect width edge data based on detection signals; Input the crack defect width edge data and defect-free signal into the DRSN2d model training to obtain the crack width quantification model; Based on an upper edge channel signal segment, a lower edge channel signal segment and a signal reference value signal segment; The crack defect width edge data is used to obtain upper edge channel signal segment data, lower edge channel signal segment data, and signal reference value signal segment data; Input the upper edge channel signal segment data, the lower edge channel signal segment data, and the signal reference value signal segment data into the crack width quantification model to obtain the width quantification value ,in: ； Where: is the true width of the crack defect, is the multi-channel sensor spacing.

5. The multi-channel detection method according to claim 4, characterized in that: Also includes: Pruning, quantization, and low-rank decomposition are performed on the crack length quantization model, the crack width quantization model, and the crack depth quantization model.

Images