Pipeline leakage detection method and system based on multi-sensor feature fusion

Abstract

The application discloses a pipeline leakage detection method and system based on multi-sensor feature fusion, relates to the technical field of pipeline detection, and mainly aims to solve the technical problem of low detection precision of existing pipeline detection technology on pipeline leakage. Mainly includes responding to the pipeline leakage detection request of the target detection pipeline, acquiring the magnetic flux leakage detection signal image in different directions of the target detection pipeline; using the trained multi-sensor feature fusion model, the feature extraction is carried out on the magnetic flux leakage detection signal image in different directions respectively, and the extracted features are subjected to feature fusion processing to obtain multi-sensor fusion features; according to the multi-sensor fusion features and the preset feature threshold, the pipeline leakage detection result is determined, and the pipeline leakage detection result includes the leakage defect grade and the leakage defect position. Mainly used for pipeline detection.

Description

Technical Field

[0001] This invention relates to the field of magnetic flux leakage detection technology, and in particular to a pipeline leakage detection method and system based on multi-sensor feature fusion. Background Technology

[0002] Regular safety inspections of oil and gas pipelines are essential for ensuring their safe operation. Magnetic flux leakage testing (MFL), a non-destructive testing technique, is characterized by high efficiency, robustness, and applicability to oil and gas pipelines, and has been widely used in oil pipeline inspection in recent years. However, the actual operating conditions of oil and gas pipelines generate complex magnetic flux leakage signals, such as instrument wear, measurement jitter, and external environmental factors. These factors can cause small defects to be easily submerged in background defect signals, severely affecting researchers' assessment of the severity of pipeline defects.

[0003] With the rapid development of image processing and deep learning technologies, many scholars have dedicated themselves to applying neural networks to defect detection. However, current neural network-based magnetic flux leakage detection methods only have a high detection rate for severe defects, and cannot accurately estimate or detect the magnetic flux leakage signals of minute defects, nor can they accurately detect the location of these small defects. Therefore, there is an urgent need for a pipeline leakage detection method to accurately identify the degree and location of pipeline defects. Summary of the Invention

[0004] In view of this, the present invention provides a pipeline leakage detection method and system based on multi-sensor feature fusion, the main purpose of which is to solve the technical problem of low detection accuracy of pipeline leakage in existing pipeline detection technologies.

[0005] According to one aspect of the present invention, a pipeline leakage detection method based on multi-sensor feature fusion is provided, comprising: In response to a pipeline leak detection request for the target pipeline, magnetic flux leakage detection signal images of the target pipeline in different directions are acquired. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals, and the magnetic flux leakage detection signals are acquired based on sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the magnetic flux leakage detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. The pipeline leak detection result is determined based on the multi-sensor fusion features and the preset feature threshold. The pipeline leak detection result includes the leak defect level and the leak defect location. The preset feature threshold is determined based on the root mean square error of the sampling fusion features of the partial magnetic flux leakage detection signal images of the target pipeline.

[0006] Furthermore, the trained multi-sensor feature fusion model includes a parallel convolutional neural network and a decoupled fully connected attention mechanism. The trained multi-sensor feature fusion model is used to extract features from the magnetic flux leakage detection signal images in different directions, and the extracted features are then fused to obtain multi-sensor fused features, including: Using a convolutional neural network, feature extraction is performed on the leakage magnetic field detection signal images in different directions to obtain the first signal features in different directions; By utilizing a decoupled fully connected attention mechanism, feature extraction is performed on the magnetic flux leakage detection signal images in different directions to obtain second signal features in different directions; The first signal features in different directions and the second signal features in different directions are fused to obtain multi-sensor fusion features.

[0007] Further, the feature fusion of the first signal features in the different directions and the second signal features in the different directions to obtain multi-sensor fusion features includes: For each signal direction, the first signal feature and the second signal feature are aggregated in the same direction to obtain aggregated features in different directions, as shown in the formula: ;in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; Multi-directional feature aggregation is performed on the aggregated features in different directions to obtain multi-sensor fusion features, as shown in the formula: ;in, This represents the multi-sensor fusion features; a, r, and c represent the weight parameters in different directions. , , This represents the aggregation characteristics of the k-th group of signals in different directions.

[0008] Furthermore, before acquiring magnetic flux leakage detection signal images of the target pipeline in different directions in response to a pipeline leakage detection request, the method further includes: Based on actual detected signal image samples, an improved conditional adversarial generative network is used to generate signal image training samples in different directions. The improved conditional adversarial generative network includes a supervisor, a generator, and a discriminator. An initial multi-sensor feature fusion model is constructed, which includes a parallel convolutional neural network and a decoupled fully connected attention mechanism; Using signal image training samples from each direction, the initial multi-sensor feature fusion model is trained separately to optimize the model weight parameters for the corresponding direction, thus obtaining the trained multi-sensor feature fusion model.

[0009] Furthermore, the step of generating signal image training samples in different directions using an improved conditional adversarial generative network based on actual detected signal image samples includes: Using the generator, samples are generated by generating signal images in different directions based on the expected sample labels and the noise signal image; The discriminator is used to determine the true sample probability of the generated signal image, and outputs the generated signal image sample whose true sample probability meets the true sample probability threshold. The supervisor constructs a sample loss function based on the generated samples from the signal image and the actual detected signal image samples; The generator and the discriminator are optimized based on the true sample probability and the sample loss function, respectively. The generation and discrimination steps of generating and judging signal images are repeated based on the optimized generator and discriminator. Once the number of generated signal image samples meets the required number of generated samples, signal image training samples in different directions are generated based on the generated signal image samples and the actual detected signal image samples.

[0010] Furthermore, before training the initial multi-sensor feature fusion model using training samples from signal images in each direction, the method further includes: The signal image training samples are encoded using a denoising encoder, and the encoded features are embedded into the feature latent space to obtain the signal feature mapping; Each convolutional separator in a multi-layer convolutional separator extracts noise feature maps from the noise sample signal features, and maps the noise features into equal parts: a first noise feature map and a second noise feature map. The first noise feature map is then fed into the next convolutional separator, and the second noise feature map is reverse-mapped to the current convolutional separator using a recursive function to obtain the denoising guidance features. The signal image training samples are reconstructed using the denoising guidance features and the signal feature mapping to obtain denoised signal image training samples in different directions. The step of training the initial multi-sensor feature fusion model using training samples of signal images from each direction further includes: Using the denoised signal image training samples from each direction, the initial multi-sensor feature fusion model is trained respectively, resulting in a fully trained multi-sensor feature fusion model.

[0011] Further, determining the pipeline leak detection result based on the multi-sensor fusion features and preset feature thresholds includes: The pixel matrix corresponding to the multi-sensor fusion features with feature values ​​greater than the preset feature threshold is determined as the leakage defect matrix; The leakage defect level is determined based on the pixel ratio of the leakage defect matrix, and the location of the leakage defect is obtained by marking the area corresponding to the leakage defect matrix using candidate boxes; The pipeline leak detection results are determined based on the leak defect level and the leak defect location.

[0012] According to another aspect of the present invention, a pipeline leak detection system based on multi-sensor feature fusion is provided, comprising: The acquisition module is used to acquire magnetic flux leakage detection signal images of the target pipeline in different directions in response to the pipeline leakage detection request. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals. The magnetic flux leakage detection signals are acquired based on sensors in the corresponding directions. The feature fusion module is used to extract features from the magnetic flux leakage detection signal images in different directions using a trained multi-sensor feature fusion model, and then perform feature fusion processing on the extracted features to obtain multi-sensor fused features. The determination module is used to determine the pipeline leak detection result based on the multi-sensor fusion features and the preset feature threshold. The pipeline leak detection result includes the leak defect and the location of the leak defect. The preset feature threshold is determined based on the sampling fusion features of a portion of the magnetic flux leakage detection signal image of the target pipeline.

[0013] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the above-described pipeline leak detection method based on multi-sensor feature fusion.

[0014] According to another aspect of the present invention, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the pipeline leak detection method based on multi-sensor feature fusion described above.

[0015] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a pipeline leak detection method and system based on multi-sensor feature fusion. In this embodiment, in response to a pipeline leak detection request from a target pipeline, magnetic flux leakage (MF) detection signal images of the target pipeline in different directions are acquired. These MF detection signals include axial, radial, and circumferential signals, acquired by sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the MF detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. Based on the multi-sensor fused features and a preset feature threshold, a pipeline leak detection result is determined. The pipeline leak detection result includes the leak defect level and leak defect location. The preset feature threshold is determined based on the root mean square error of the sampled fusion features of a portion of the MF detection signal images of the target pipeline. Compared with the prior art, the embodiments of the present invention establish the correlation between signals in different directions by performing feature fusion on axial, radial and circumferential signals and leakage magnetic field detection signals in multiple directions, thereby enhancing the signal image features and extracting richer features. At the same time, a preset feature threshold is set based on the image fusion features of the target pipeline, which improves the matching degree between the preset feature threshold and the target pipeline, thereby greatly improving the detection accuracy of pipeline leakage. The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a pipeline leak detection method based on multi-sensor feature fusion provided by an embodiment of the present invention is shown; Figure 2 The flowchart of another pipeline leak detection method based on multi-sensor feature fusion provided by an embodiment of the present invention is shown; Figure 3 A schematic diagram of the structure of an improved adversarial generative network model provided by an embodiment of the present invention is shown; Figure 4The diagram shows a block diagram of a pipeline leak detection system based on multi-sensor feature fusion provided by an embodiment of the present invention. Figure 5 A schematic diagram of the structure of a terminal provided in an embodiment of the present invention is shown. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0018] To address the problem of low detection accuracy in existing pipeline inspection technologies, this invention provides a pipeline leak detection method based on multi-sensor feature fusion, such as... Figure 1 As shown, the method includes: 101. In response to a pipeline leakage detection request for the target pipeline, acquire magnetic flux leakage detection signal images of the target pipeline in different directions.

[0019] In this embodiment of the invention, the target pipeline for detection is an oil and gas pipeline. The magnetic flux leakage (MFL) detection signal image is a magnetic flux signal image collected by a Hall sensor configured on the MFL detection robot. During MFL detection of the target pipeline, the MFL detection robot is pushed from the sending end of the pipeline at a constant speed of 0.5 m / s to the receiving end. During this process, a permanent magnet magnetizes the pipe wall, forming a saturated magnetic field. When the pipe wall is intact, the magnetic field distribution is uniform. However, when there are defects or other components on the pipe wall, the wall thickness decreases, making it impossible to maintain the original magnetic flux density. When a defect occurs, the magnetic field leaks from the pipe wall into the surrounding air. The Hall sensor installed on the magnetic cross-section can detect the magnetic field leakage and emit a magnetic flux signal. The MFL detection signal images in different directions of the target pipeline are images of the MFL detection signals at any location on the pipe wall in different directions. These different directions specifically include axial, radial, and circumferential signals. The MFL detection signals in different directions are collected by Hall sensors in the corresponding directions; therefore, the multi-directional MFL detection signals are multi-sensor MFL detection signals.

[0020] 102. Using the trained multi-sensor feature fusion model, feature extraction is performed on the magnetic flux leakage detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features.

[0021] In this embodiment of the invention, the detection objective of the trained multi-sensor feature fusion model is to fuse the features of magnetic flux leakage detection signals from multiple directions at the same location. Specifically, the trained multi-sensor feature fusion model includes a parallel convolutional neural network and a decoupled fully connected attention mechanism (DFC). During the feature extraction stage, the convolutional neural network and the decoupled fully connected attention mechanism are used in parallel to extract features from the magnetic flux leakage detection signals in each direction. Finally, the features extracted by different neural networks from different directions are fused to obtain multi-sensor fused features. By employing a convolutional neural network and a decoupled fully connected attention mechanism to extract features in parallel and then aggregating the extracted features, signal feature enhancement can be achieved, making it easier to capture small defect signal images in the pipeline. Furthermore, by fusing the signal image features from different directions acquired by multiple sensors, the correlation between detection signals in different directions can be established, resulting in richer features and providing a more accurate image feature basis for subsequent defect determination.

[0022] 103. Determine the pipeline leak detection result based on the multi-sensor fusion features and preset feature thresholds.

[0023] In this embodiment of the invention, to more accurately determine whether the magnetic flux leakage detection signal image corresponding to the multi-sensor fusion features is a defect, a preset feature threshold is determined based on the root mean square error of the sampling fusion features of partial magnetic flux leakage detection signal images of the target detection pipeline. In one application example, the calculation process of the preset feature threshold includes: 1) acquiring the target magnetic flux leakage detection signal of the target detection data, and selecting from the target magnetic flux leakage detection signal... Target magnetic flux leakage detection signal , , ;in, This represents the axial signal at the i-th position; This is the radial signal at the i-th position; 1) The circumferential signal at position i. 2) Median filtering is used to smooth the selected multi-sensor signals to reduce the interference of additive noise generated during pipeline acquisition on the MFL signal. 3) The smoothed signal is input into the trained multi-sensor feature fusion model, and the model's output is recorded as follows. 4) Calculation The mean square error T is given by the formula: (1); T is set as the preset characteristic threshold. Besides using mean square error, fuzzy comprehensive evaluation (which transforms qualitative evaluation into quantitative evaluation) and regional flood frequency analysis models based on peak values ​​exceeding the threshold can also be used. Processing is performed to determine the preset feature threshold, but currently the method based on mean square error to determine the threshold has the least computational load and is the most convenient to calculate.

[0024] It should be noted that since the wall thickness and diameter of different target inspection pipelines vary, the criteria for judging leakage defects will also differ. By determining a preset feature threshold based on the inspection data of the current target inspection pipeline, a high degree of matching between the preset feature threshold and the characteristics of the target inspection pipeline is achieved, thereby effectively ensuring the accuracy of leakage defect identification.

[0025] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, step 102 utilizes the trained multi-sensor feature fusion model to extract features from the magnetic flux leakage detection signal images in different directions, and then performs feature fusion processing on the extracted features to obtain multi-sensor fused features, including: 201. Using a convolutional neural network, feature extraction is performed on the magnetic flux leakage detection signal images in different directions to obtain the first signal features in different directions.

[0026] 202. Using a decoupled fully connected attention mechanism, feature extraction is performed on the magnetic flux leakage detection signal images in different directions to obtain second signal features in different directions.

[0027] 203. Perform feature fusion on the first signal features in different directions and the second signal features in different directions to obtain multi-sensor fusion features.

[0028] In this embodiment of the invention, the trained multi-sensor feature fusion model includes parallel convolutional neural networks and a decoupled fully connected attention mechanism. The convolutional neural network is an improved model of CNN (Convolutional Neural Network), which, in order to aggregate information with the decoupled fully connected attention mechanism, has fewer fully connected layers than a typical CNN. The multi-sensor feature fusion model includes convolutional neural networks corresponding to the axial, radial, and circumferential directions, respectively, and a decoupled fully connected attention mechanism. The network structure of the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, and an output layer. Taking the axial convolutional neural network as an example, the first... The input matrix of each convolutional layer is Assume that the first... The size of each convolutional kernel is . No. The output matrix of each convolutional layer The calculation formula is: (2); in, It is a non-linear activation function. It's a convolution operation. It is a fixed step size. This is the input to the convolutional layer. To prevent overfitting, pooling layers are used to reduce the dimensionality of the convolutional signal, thus better extracting feature information from small defects. The pooling layer formula is: (3); in The output features of the pooling layer; The input features for the pooling layer; and The height and width of the pooling core, The step size of the pooling core, This is the max pooling function.

[0029] The network structure decoupled from the fully connected attention mechanism includes an upsampling layer, a fully connected layer, and an upsampling layer. Taking the axial direction as an example, the first... The input matrix of each convolutional layer is To process input images with different resolutions, a 1×1 convolutional kernel is used to preprocess the input feature matrix. This decouples the size of the convolution kernel from the size of the feature map, and performs average pooling on the preprocessed feature matrix to output the matrix. The formula is: (4); in It is a non-linear activation function. It is a 1×1 convolution operation. This is the average pooling function; This represents the output characteristics of the fully connected layer.

[0030] Two fully connected layers are used to decouple the signal image along the horizontal and vertical directions, respectively, and downsampling is performed on the signal images in both directions. The original features input to the fully connected layer are defined as... The width and height of the original feature are defined as follows: and The output features of the fully connected layer capture long-range correlations along two directions respectively. The calculation is as follows: (5); in, and For the corresponding weights, It is multiplication in which each element is operated on independently.

[0031] Using two fully connected layers to aggregate long-distance information in both directions significantly reduces computational complexity. Downsampling reduces the feature map size, allowing all operations in the decoupled fully connected attention mechanism to be performed on a smaller feature map. By downsampling by half in both the horizontal and vertical directions, the overall floating-point operations per second of the decoupled fully connected attention mechanism's attention module are reduced by 75%. The theoretical complexity of the decoupled fully connected attention mechanism is... Finally, a bilinear interpolation upsampling method is used to restore the output features of the fully connected layers to their original size, thus maintaining consistency with the resolution of the convolutional neural network branches.

[0032] In one embodiment of the present invention, for further explanation and limitation, the step of fusing the first signal features in different directions and the second signal features in different directions to obtain multi-sensor fusion features includes: For each signal direction, the first signal feature and the second signal feature are aggregated in the same direction to obtain aggregated features in different directions, as shown in the formula: ;in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; Multi-directional feature aggregation is performed on the aggregated features in different directions to obtain multi-sensor fusion features, as shown in the formula: ;in, This represents the multi-sensor fusion features; a, r, and c represent the weight parameters in different directions. , , This represents the aggregation characteristics of the k-th group of signals in different directions.

[0033] In this embodiment of the invention, to compensate for the shortcomings of classic CNN modules in modeling spatial dependencies, this paper aggregates information from CNN and a decoupled fully connected attention mechanism module to enhance signal features, and employs... Reduce the dimensionality of the input feature map. The specific formula is: (6); in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; It is multiplication in which each element is operated on independently.

[0034] Using the above formula (6), the aggregation results of the feature extraction results (second signal features) of the decoupled fully connected attention mechanism branch and the feature extraction results (second signal features) of the convolutional neural network branch in each direction are calculated to obtain the axial aggregation features. Radial aggregation characteristics ; circumferential aggregation features Where k represents the k-th group, the... Output of group feature fusion The formula is expressed as: (7); in, As the weight of the axial signal, For the weight of the radial signal, The weights are for the circumferential signal.

[0035] It should be noted that, given that important features extracted from magnetic flux leakage (MFL) signal images in one direction may be overlooked by features from signals in other directions, correlations between axial, radial, and circumferential signals were established based on corresponding weights, and signal features from each direction were fused. Due to the presence of deep networks, feature extraction incorporates a large amount of feature information from the MFL signal and can focus on the differences and correlations between multiple signal features. Furthermore, since the fusion of signal features from each set of sensor samples is a necessary condition for pipeline leakage defect inversion, fusing signal features acquired from multiple sensors can effectively improve the accuracy of pipeline leakage defect identification.

[0036] In one embodiment of the present invention, for further explanation and limitation, before the step of acquiring magnetic flux leakage detection signal images in different directions of the target detection pipeline in response to a magnetic flux leakage detection request, the method further includes: Based on actual detected signal image samples, an improved conditional adversarial generative network is used to generate signal image training samples in different directions. Construct an initial multi-sensor feature fusion model; Using signal image training samples from each direction, the initial multi-sensor feature fusion model is trained separately to optimize the model weight parameters for the corresponding direction, thus obtaining the trained multi-sensor feature fusion model.

[0037] In this embodiment of the invention, the initial multi-sensor feature fusion model includes a parallel convolutional neural network and a decoupled fully connected attention mechanism. The training objective of the model is to approximate a certain proportion of the output between signal features from multiple directions. The inputs to these feature extraction networks are signals from different sensors at the same time and location. A separate training strategy is adopted during network training; the multiple sensors and directions in the model are trained separately and sequentially to prevent the model from reaching local optima. Specifically, after obtaining the signal image training samples, labels are defined for the samples according to three directions, and the training samples from the three directions are used to train the model separately. For the training process of signal image training samples in any direction, the hysteresis loss between the desired feature output and the output features is used as the loss function, with the formula: (8); in, For the desired feature output, This provides the actual feature output. After training in the three directions, to fully utilize the information features of the input signal image and reduce the model's dependence on a single signal, weight parameters for the three directions are set. , and ,definition , and Loss functions are introduced in three directions respectively. The loss function combining the three directions is as follows: (9); By dynamically adjusting the weight parameters , and make Minimize, so that , and By approximating each other, the optimized weight parameters in the three directions are obtained, thus completing the training of the initial multi-sensor feature fusion model.

[0038] In one embodiment of the present invention, for further explanation and limitation, the step of generating signal image training samples in different directions using an improved conditional adversarial generative network based on actual detected signal image samples includes: Using the generator, samples are generated by generating signal images in different directions based on the expected sample labels and the noise signal image; The discriminator is used to determine the true sample probability of the generated signal image, and outputs the generated signal image sample whose true sample probability meets the true sample probability threshold. The supervisor constructs a sample loss function based on the generated samples from the signal image and the actual detected signal image samples; The generator and the discriminator are optimized based on the true sample probability and the sample loss function, respectively. The generation and discrimination steps of generating and judging signal images are repeated based on the optimized generator and discriminator. Once the number of generated signal image samples meets the required number of generated samples, signal image training samples in different directions are generated based on the generated signal image samples and the actual detected signal image samples.

[0039] In embodiments of the present invention, such as Figure 4 As shown, the improved conditional adversarial generative network includes a supervisor, a generator, and a discriminator. The magnetic flux leakage detection signal is a signal stream that increases with mileage. To train a more accurate network model, an improved conditional adversarial generative network is proposed to increase the amount of simulation data and samples. Input , and These are leakage magnetic flux detection signals from the same location and time in the axial, radial, and circumferential directions. The improved conditional adversarial generative network consists of two "adversarial" models, one of which is a generative model. One is used to capture data distribution; the other is a discriminant model. This is used to estimate the probability that a sample comes from the real signal and the signal from the generator model. In the generator model, the prior input noise... With conditional information By combining these elements to form a joint hidden layer, the adversarial training framework offers great flexibility in the composition of the hidden layers. The supervision module is used to introduce feature losses from both generated and real samples as a loss function during the sample generation process, enabling better sample generation based on this loss function. Then, and The input signal is used for identification in the discrimination model. If the identified signal meets the criteria for a true signal, it is output; otherwise, it is discarded. The method used is... The loss function is expressed in the following formula: (10); in, For the desired output, This is the actual output. During the sample generation process, and They are all nonlinear mapping functions with fixed discriminant models. ,Adjustment The parameters make Minimize the expectation; fix the generative model ,Adjustment The parameters make To maximize the expected value. In summary, the specific optimization process is as follows: (11); in, For input noise, For real signal samples, For the sample labels that need to be generated, The probability of real data. The probability of it being fake data. The discriminator model determines the signal to be real, generating multiple sets of samples. The outputs of the improved conditional adversarial generative network are denoted as follows: , and Each sample group includes axial signal samples, radial signal samples, and circumferential signal samples. After obtaining training samples based on the generative model, the generated samples and the actual measured samples are used as training samples. Signals collected by different sensors are divided into a group of samples at the same location and in the same compartment.

[0040] It should be noted that before generating signal samples, the multi-sensor signals need to be standardized. The axial signal source is defined as... The radial signal source is The circumferential signal is The three sets of input signals collected are defined as follows: , , , are Gaussian / voltage numerical matrices, respectively. , and These are axial, radial, and circumferential signals from the same location and time. Taking the acquired magnetic flux leakage detection signal as a voltage signal as an example, the voltage signal is converted into a Gaussian signal for processing. The unified conversion formula for the signal value is: (12); in, The acquired voltage signal. The processed Gaussian signal, This is the voltage-to-Gauss conversion ratio. This is the amplification factor of the voltage signal. and is a constant, where , .

[0041] Given the differences in reference values ​​for signals acquired by different sensors, the Gaussian signal is standardized based on a unified benchmark value, using the following formula: (13); in, It is a collected Gaussian signal. It is a signal with a unified reference value. It is a baseline value, which is a constant; Indicates the number of Hall sensors; This indicates the number of sampling points of the Hall sensor.

[0042] Standardized sample signals not only reduce the influence of reference values ​​on feature extraction and provide a basis for the feature fusion module of the network structure, but also reduce the differences between small defect detection samples in multi-directional signals.

[0043] In one embodiment of the present invention, for further explanation and limitation, before the step of training the initial multi-sensor feature fusion model using signal image training samples from each direction, the method further includes: The signal image training samples are encoded using a denoising encoder, and the encoded features are embedded into the feature latent space to obtain the signal feature mapping; Each convolutional separator in a multi-layer convolutional separator extracts noise feature maps from the noise sample signal features, and maps the noise features into equal parts: a first noise feature map and a second noise feature map. The first noise feature map is then fed into the next convolutional separator, and the second noise feature map is reverse-mapped to the current convolutional separator using a recursive function to obtain the denoising guidance features. The signal image training samples are reconstructed using the denoising guidance features and the signal feature mapping to obtain denoised signal image training samples in different directions. The step of training the initial multi-sensor feature fusion model using training samples of signal images from each direction further includes: Using the denoised signal image training samples from each direction, the initial multi-sensor feature fusion model is trained respectively, resulting in a fully trained multi-sensor feature fusion model.

[0044] In this embodiment of the invention, due to the significant noise generated during industrial testing, the measured signal samples may contain random interference, such as from Hall sensors, ambient light, and interference during transmission or storage. To improve the visual quality of the data and make the sample signal images easier to analyze, process, and understand, denoising processing of the signal image training samples is performed based on multi-scale sample features of a neural network. This network mainly includes a denoising encoder, a sample feature extraction module, and a signal reconstruction module. The sample feature extraction module includes a multi-layer convolutional separator and a recursive function. The denoising encoder includes multiple convolutional layers and... Structurally, the denoising encoder and signal reconstruction module form a typical U-shaped network structure in the activation layer. The denoising encoder encodes the signal image training samples and embeds the encoded features into the feature latent space. The encoding and embedding process is represented as follows: (14); in, Indicates the sample signal of the first Features of subconvolution; Represents the convolutional layer function; Weight parameters for feature extraction; This represents an axial signal sample.

[0045] The denoising feature guidance module includes multiple convolutional splitters and recursive functions. The convolutional splitters use 3×3 convolutional layers, and the activation function and segmentation function serve as the input feature maps for subsequent convolutional splitters. This module can induce prior features for noise removal from the reconstructed signal. To extract richer feature maps, the integrated feature maps are divided into two equal parts, and the feature maps are reversed on each channel to obtain the feature maps. These are then fed into the convolutional splitter operation using a recursive function to generate richer feature maps. With the number of channels remaining constant, rich features can be effectively generated, making it easier to induce the prior features required for denoising. At each step, the module divides the channel from which the feature maps are extracted into two equal parts; one part is retained, and the other part is fed into the next convolutional splitter. Then, all feature maps are concatenated together, using the segmentation function and connector shown below: (15); in, Indicates the first Each convolutional separator operation, Indicates the first Connector operation. For the characteristics of noise sample signals, Features preserved for noise signals As input to the next convolutional separator, For the first The output characteristics of each connector The number of segments.

[0046] The signal reconstruction module consists of multiple deconvolution layers and The decoder structure, composed of activation functions, reconstructs high-quality images from the latent feature space using a shared structure. This not only reduces the number of parameters in the entire network but also lowers the cost of network iteration. Each layer of the signal reconstruction module introduces denoising guiding features of the corresponding feature map size to reconstruct clear signal samples from the latent feature space. The network is then trained to learn, and the reconstructed signal can be represented as: (16); in To extract the first feature from the sample feature extraction module Features at each convolutional scale; To guide the denoising feature module Features of the connector output, These are the denoised samples for the reconstructed signal.

[0047] In one embodiment of the present invention, for further explanation and limitation, the step of determining the pipeline leak detection result based on the multi-sensor fusion features and preset feature thresholds includes: The pixel matrix corresponding to the multi-sensor fusion features with feature values ​​greater than the preset feature threshold is determined as the leakage defect matrix; The leakage defect level is determined based on the pixel ratio of the leakage defect matrix, and the location of the leakage defect is obtained by marking the area corresponding to the leakage defect matrix using candidate boxes; The pipeline leak detection results are determined based on the leak defect level and the leak defect location.

[0048] In this embodiment of the invention, the multi-sensor fusion features are compared with a preset feature threshold. If the multi-sensor fusion features are greater than the preset feature threshold, it indicates that the current fusion features are defective features. The image matrix corresponding to this defective feature needs to be marked with candidate boxes to pinpoint the defect location. Simultaneously, the area or proportion of the image matrix can be calculated. Based on the comparison between the area or proportion of the image matrix and the corresponding threshold, the leakage defect is classified into levels. For example, if the area of ​​the image matrix is ​​greater than a first matrix area threshold, the leakage defect level is determined to be a Level 1 severe defect; if the area of ​​the image matrix is ​​less than the first matrix area threshold, the leakage defect level is determined to be a Level 2 minor defect. This yields a pipeline leakage detection result including the leakage defect level and location. If the multi-sensor fusion features are less than or equal to the preset feature threshold, it indicates that the signal image matrix is ​​a normal matrix and does not need to be marked with candidate boxes.

[0049] This invention provides a pipeline leak detection method based on multi-sensor feature fusion. In this embodiment, in response to a pipeline leak detection request from a target pipeline, magnetic flux leakage (MF) detection signal images of the target pipeline in different directions are acquired. These MF detection signals include axial, radial, and circumferential signals, acquired by sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the MF detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. Based on the multi-sensor fused features and a preset feature threshold, a pipeline leak detection result is determined. The pipeline leak detection result includes the leak defect level and the leak defect location. The preset feature threshold is determined based on the root mean square error of the sampled fusion features of a portion of the MF detection signal images of the target pipeline. Compared with the prior art, the embodiments of the present invention establish the correlation between signals in different directions by performing feature fusion on axial, radial and circumferential signals and leakage magnetic flux detection signals in multiple directions, thereby enhancing the signal image features and extracting richer features. At the same time, a preset feature threshold is set based on the image fusion features of the target pipeline, which improves the matching degree between the preset feature threshold and the target pipeline, thereby greatly improving the detection accuracy of pipeline leakage. Furthermore, as a response to the above Figure 1 The implementation of the method shown in this invention provides a pipeline leak detection system based on multi-sensor feature fusion, such as... Figure 4 As shown, the system includes: The acquisition module 31 is used to acquire magnetic flux leakage detection signal images of the target detection pipeline in different directions in response to the pipeline leakage detection request. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals. The magnetic flux leakage detection signals are collected based on sensors in the corresponding directions. Feature fusion module 32 is used to extract features from the magnetic flux leakage detection signal images in different directions using a trained multi-sensor feature fusion model, and to perform feature fusion processing on the extracted features to obtain multi-sensor fused features. The determination module 33 is used to determine the pipeline leakage detection result based on the multi-sensor fusion features and the preset feature threshold. The pipeline leakage detection result includes the leakage defect and the location of the leakage defect. The preset feature threshold is determined based on the sampling fusion features of the partial magnetic flux leakage detection signal image of the target detection pipeline.

[0050] Furthermore, the feature fusion module 32 includes: The first extraction unit is used to extract features from the magnetic flux leakage detection signal images in different directions using a convolutional neural network, so as to obtain the first signal features in different directions. The second extraction unit is used to extract features from the magnetic flux leakage detection signal images in different directions using a decoupled fully connected attention mechanism, so as to obtain the second signal features in different directions. The fusion unit is used to fuse the first signal features and the second signal features in different directions to obtain multi-sensor fusion features.

[0051] Furthermore, the fusion unit is specifically used to perform feature aggregation of the first signal feature and the second signal feature in the same direction for each signal direction, to obtain aggregated features in different directions, as shown in the formula: ;in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; Multi-directional feature aggregation is performed on the aggregated features in different directions to obtain multi-sensor fusion features, as shown in the formula: ;in, This represents the multi-sensor fusion features; a, r, and c represent the weight parameters in different directions. , , This represents the aggregation characteristics of the k-th group of signals in different directions.

[0052] Furthermore, the system also includes: The generation module is used to generate signal image training samples in different directions based on actual detected signal image samples and using an improved conditional adversarial generative network. The improved conditional adversarial generative network includes a supervisor, a generator, and a discriminator. A building module is used to build an initial multi-sensor feature fusion model, which includes a parallel convolutional neural network and a decoupled fully connected attention mechanism. The training module is used to train the initial multi-sensor feature fusion model using signal image training samples from each direction, thereby optimizing the model weight parameters for the corresponding direction and obtaining the trained multi-sensor feature fusion model.

[0053] Furthermore, the generation module includes: The first generation unit is used to generate samples by generating signal images in different directions based on the expected sample label and the noise signal image using the generator. The second generation unit is used to use the discriminator to determine the true sample probability of the generated signal image sample, and output the generated signal image sample whose true sample probability meets the true sample probability threshold. The construction unit is used by the supervisor to construct a sample loss function based on the generated samples of the signal image and the actual detected signal image samples; The parameter optimization unit is used to optimize the parameters of the generator and the discriminator based on the real sample probability and the sample loss function, respectively, and repeat the generation and discrimination steps of generating and judging signal images based on the optimized generator and discriminator; The third generation unit is used to generate signal image training samples in different directions based on the signal image generated samples and the actual detected signal image samples after the number of signal image generated samples meets the number of generated samples.

[0054] Furthermore, the system also includes: The encoding module is used to encode the signal image training samples using a denoising encoder and embed the encoded features into the feature latent space to obtain the signal feature mapping; The feature guidance module is used to extract noise feature maps from the noise sample signal features using each layer of the multi-layer convolutional separator, and map the noise features into equal parts: a first noise feature map and a second noise feature map; the first noise feature map is fed into the next convolutional separator, and the second noise feature map is reverse-mapped to the current convolutional separator using a recursive function to obtain the denoising guidance feature; The signal reconstruction module is used to reconstruct the signal image training samples using the denoising guidance features and the signal feature mapping to obtain denoised signal image training samples in different directions. The training module is further used to train the initial multi-sensor feature fusion model using the denoised signal image training samples in each direction, thereby obtaining the trained multi-sensor feature fusion model.

[0055] Furthermore, the determining module 33 includes: The first determining unit is used to determine the pixel matrix corresponding to the multi-sensor fusion features whose feature values ​​are greater than the preset feature threshold as the leakage defect matrix. The marking unit is used to determine the leakage defect level based on the pixel ratio of the leakage defect matrix, and to use the candidate box to mark the area corresponding to the leakage defect matrix to obtain the leakage defect location; The second determining unit is used to determine the pipeline leak detection result based on the leak defect level and the leak defect location.

[0056] This invention provides a pipeline leak detection system based on multi-sensor feature fusion. In this embodiment, in response to a pipeline leak detection request from a target pipeline, magnetic flux leakage (MF) detection signal images of the target pipeline in different directions are acquired. These MF detection signals include axial, radial, and circumferential signals, acquired by sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the MF detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. Based on the multi-sensor fused features and a preset feature threshold, the pipeline leak detection result is determined. The pipeline leak detection result includes the leak defect level and the leak defect location. The preset feature threshold is determined based on the root mean square error of the sampled fusion features of a portion of the MF detection signal images of the target pipeline. Compared with the prior art, the embodiments of the present invention establish the correlation between signals in different directions by performing feature fusion on axial, radial and circumferential signals and leakage magnetic field detection signals in multiple directions, thereby enhancing the signal image features and extracting richer features. At the same time, a preset feature threshold is set based on the image fusion features of the target pipeline, which improves the matching degree between the preset feature threshold and the target pipeline, thereby greatly improving the detection accuracy of pipeline leakage. According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, the computer-executable instruction being able to execute the pipeline leak detection method based on multi-sensor feature fusion in any of the above method embodiments.

[0057] Figure 5 The diagram shows a structural schematic of a terminal according to an embodiment of the present invention. The specific implementation of the terminal is not limited by the specific embodiments of the present invention.

[0058] like Figure 5 As shown, the terminal may include: a processor 402, a communications interface 404, a memory 406, and a communications bus 408.

[0059] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0060] Communication interface 404 is used for network communication with other devices such as clients or other servers.

[0061] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above embodiment of the pipeline leakage detection method based on multi-sensor feature fusion.

[0062] Specifically, program 410 may include program code that includes computer operation instructions.

[0063] Processor 402 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The terminal may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0064] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0065] Specifically, program 410 can be used to cause processor 402 to perform the following operations: In response to a pipeline leak detection request for the target pipeline, magnetic flux leakage detection signal images of the target pipeline in different directions are acquired. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals, and the magnetic flux leakage detection signals are acquired based on sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the magnetic flux leakage detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. The pipeline leak detection result is determined based on the multi-sensor fusion features and the preset feature threshold. The pipeline leak detection result includes the leak defect level and the leak defect location. The preset feature threshold is determined based on the root mean square error of the sampling fusion features of the partial magnetic flux leakage detection signal images of the target pipeline.

[0066] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A pipeline leak detection method based on multi-sensor feature fusion, characterized in that, include: In response to a pipeline leak detection request for the target pipeline, magnetic flux leakage detection signal images of the target pipeline in different directions are acquired. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals, and the magnetic flux leakage detection signals are acquired based on sensors in the corresponding directions. Using a trained multi-sensor feature fusion model, features are extracted from the magnetic flux leakage detection signal images in different directions, and the extracted features are fused to obtain multi-sensor fused features. The pipeline leak detection result is determined based on the multi-sensor fusion features and the preset feature threshold. The pipeline leak detection result includes the leak defect level and the leak defect location. The preset feature threshold is determined based on the root mean square error of the sampling fusion features of the partial magnetic flux leakage detection signal image of the target pipeline. The trained multi-sensor feature fusion model includes a parallel convolutional neural network and a decoupled fully connected attention mechanism. The trained multi-sensor feature fusion model is used to extract features from the magnetic flux leakage detection signal images in different directions, and the extracted features are then fused to obtain multi-sensor fused features, including: Using a convolutional neural network, feature extraction is performed on the leakage magnetic field detection signal images in different directions to obtain the first signal features in different directions; By utilizing a decoupled fully connected attention mechanism, feature extraction is performed on the magnetic flux leakage detection signal images in different directions to obtain second signal features in different directions; The first signal features in different directions and the second signal features in different directions are fused to obtain multi-sensor fusion features; The feature fusion of the first signal features in different directions and the second signal features in different directions to obtain multi-sensor fusion features includes: For each signal direction, the first signal feature and the second signal feature are aggregated in the same direction to obtain aggregated features in different directions, as shown in the formula: ;in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; Multi-directional feature aggregation is performed on the aggregated features in different directions to obtain multi-sensor fusion features, as shown in the formula: ;in, This represents the multi-sensor fusion features; a, r, and c represent the weight parameters in different directions. , , This represents the aggregation characteristics of the k-th group of signals in different directions.

2. The method according to claim 1, characterized in that, Before acquiring magnetic flux leakage detection signal images in different directions of the target pipeline in response to a pipeline leakage detection request, the method further includes: Based on actual detected signal image samples, an improved conditional adversarial generative network is used to generate signal image training samples in different directions. The improved conditional adversarial generative network includes a supervisor, a generator, and a discriminator. An initial multi-sensor feature fusion model is constructed, which includes a parallel convolutional neural network and a decoupled fully connected attention mechanism; Using signal image training samples from each direction, the initial multi-sensor feature fusion model is trained separately to optimize the model weight parameters for the corresponding direction, thus obtaining the trained multi-sensor feature fusion model.

3. The method according to claim 2, characterized in that, The step of generating signal image training samples in different directions based on actual detected signal image samples using an improved conditional adversarial generative network includes: Using the generator, samples are generated by generating signal images in different directions based on the expected sample labels and the noise signal image; The discriminator is used to determine the true sample probability of the generated signal image, and outputs the generated signal image sample whose true sample probability meets the true sample probability threshold. The supervisor constructs a sample loss function based on the generated samples from the signal image and the actual detected signal image samples; The generator and the discriminator are optimized based on the true sample probability and the sample loss function, respectively. The generation and discrimination steps of generating and judging signal images are repeated based on the optimized generator and discriminator. Once the number of generated signal image samples meets the required number of generated samples, signal image training samples in different directions are generated based on the generated signal image samples and the actual detected signal image samples.

4. The method according to claim 2, characterized in that, Before training the initial multi-sensor feature fusion model using training samples from signal images in each direction, the method further includes: The signal image training samples are encoded using a denoising encoder, and the encoded features are embedded into the feature latent space to obtain the signal feature mapping; Each convolutional separator in a multi-layer convolutional separator extracts noise feature maps from the noise sample signal features, and maps the noise features into equal parts: a first noise feature map and a second noise feature map. The first noise feature map is then fed into the next convolutional separator, and the second noise feature map is reverse-mapped to the current convolutional separator using a recursive function to obtain the denoising guidance features. The signal image training samples are reconstructed using the denoising guidance features and the signal feature mapping to obtain denoised signal image training samples in different directions. The step of training the initial multi-sensor feature fusion model using training samples of signal images from each direction further includes: Using the denoised signal image training samples from each direction, the initial multi-sensor feature fusion model is trained respectively, resulting in a fully trained multi-sensor feature fusion model.

5. The method according to any one of claims 1-4, characterized in that, The step of determining the pipeline leak detection result based on the multi-sensor fusion features and preset feature thresholds includes: The pixel matrix corresponding to the multi-sensor fusion features with feature values ​​greater than the preset feature threshold is determined as the leakage defect matrix; The leakage defect level is determined based on the pixel ratio of the leakage defect matrix, and the location of the leakage defect is obtained by marking the area corresponding to the leakage defect matrix using candidate boxes; The pipeline leak detection results are determined based on the leak defect level and the leak defect location.

6. A pipeline leak detection system based on multi-sensor feature fusion, characterized in that, include: The acquisition module is used to acquire magnetic flux leakage detection signal images of the target pipeline in different directions in response to the pipeline leakage detection request. The magnetic flux leakage detection signals in different directions include axial signals, radial signals and circumferential signals. The magnetic flux leakage detection signals are acquired based on sensors in the corresponding directions. The feature fusion module is used to extract features from the magnetic flux leakage detection signal images in different directions using a trained multi-sensor feature fusion model, and then perform feature fusion processing on the extracted features to obtain multi-sensor fused features. The determination module is used to determine the pipeline leak detection result based on the multi-sensor fusion features and the preset feature threshold. The pipeline leak detection result includes the leak defect and the location of the leak defect. The preset feature threshold is determined based on the sampling fusion features of a portion of the magnetic flux leakage detection signal image of the target pipeline. The feature fusion module includes: The first extraction unit is used to extract features from the magnetic flux leakage detection signal images in different directions using a convolutional neural network, so as to obtain the first signal features in different directions. The second extraction unit is used to extract features from the magnetic flux leakage detection signal images in different directions using a decoupled fully connected attention mechanism, so as to obtain the second signal features in different directions. A fusion unit is used to fuse the first signal features in different directions and the second signal features in different directions to obtain multi-sensor fusion features; The fusion unit is specifically used to perform feature aggregation of the first signal feature and the second signal feature in the same direction for each signal direction, to obtain aggregated features in different directions, as shown in the formula: ;in, This represents a depthwise separable convolution function; Represents the normalization function; This indicates the aggregation feature in any direction; Indicates the first signal characteristic in any direction; This represents the second signal characteristic in any direction; Multi-directional feature aggregation is performed on the aggregated features in different directions to obtain multi-sensor fusion features, as shown in the formula: ;in, This represents the multi-sensor fusion features; a, r, and c represent the weight parameters in different directions. , , This represents the aggregation characteristics of the k-th group of signals in different directions.

7. A storage medium storing at least one executable instruction that causes a processor to perform an operation corresponding to the pipeline leak detection method based on multi-sensor feature fusion as described in any one of claims 1-5.

8. A terminal, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the pipeline leak detection method based on multi-sensor feature fusion as described in any one of claims 1-5.

Images