A deep learning-based pipeline internal detection data AI analysis method and system

Abstract

The application discloses a kind of pipe internal detection data AI analysis method and system based on deep learning.The method comprises: obtaining pipe internal detection data, including magnetic flux leakage detection signal data (hereinafter referred to as magnetic flux leakage signal) and IMU data;Magnetic flux leakage signal is preprocessed, and axial, circumferential and radial multi-component magnetic flux leakage signal is extracted;The multi-component magnetic flux leakage signal after preprocessing is input into the deep convolution recurrent neural network model comprising convolution layer, cycle layer and dense layer;Space feature and time sequence feature are extracted by convolution layer and cycle layer respectively;Feature is fused and the results of defect identification, classification and quantization are output by dense layer.The application realizes multi-component data fusion and automatic feature extraction by TBCRNN or DRCRNN deep network model, realizes the integration of intelligent analysis of deformation, magnetic flux leakage and IMU detection by combining gyro, accelerometer and odometer fusion algorithm IMU trajectory solution.

Description

Technical Field

[0001] This invention belongs to the field of pipeline internal inspection data analysis technology, specifically relating to an AI analysis method and system for pipeline internal inspection data based on deep learning. Background Technology

[0002] As a primary mode of transportation for energy media such as oil and natural gas, the safe operation of pipelines directly impacts the stability of energy supply and public safety. Pipeline in-situ inspection technology is a core means of ensuring safe pipeline operation. By deploying in-situ detectors inside the pipeline, various onboard sensors collect data on pipeline wall thickness, geometric deformation, and metal loss, enabling comprehensive detection of pipeline defects. Currently, pipeline in-situ inspection technologies mainly include three categories: deformation detection, magnetic flux leakage (MFL) detection, and inertial measurement unit (IMU) detection. Deformation detection is used to detect geometric deformation defects in pipelines, such as dents and ellipticity deformation; MFL detection identifies metal loss defects, such as corrosion and cracks, by detecting magnetic flux leakage signals in the pipe wall; and IMU detection uses inertial navigation technology to obtain the spatial trajectory of the pipeline centerline.

[0003] However, existing methods for analyzing pipeline in-situ inspection data have several shortcomings: First, data analysis efficiency is low. Deformation detection data, magnetic flux leakage detection data, and IMU data are typically collected separately by different internal detectors, requiring multiple inspections at different times, resulting in long on-site operation times and high costs. Different batches of data require manual alignment and matching, and the analysis of each 100 kilometers of data usually requires two months of manual labor. Second, magnetic flux leakage signal data analysis is extremely difficult. Magnetic flux leakage signal data is a typical high-dimensional (>1000 channels), high-frequency (>1kHz), large dynamic range (>70dB), and extremely low signal-to-noise ratio (<3dB) time-series signal. Conventional deep learning models are designed for two-dimensional image data and cannot be directly applied to one-dimensional magnetic flux leakage time-series signals. Third, the quantitative analysis accuracy of existing magnetic flux leakage data is insufficient. Due to the nonlinear and multi-component coupling characteristics of magnetic flux leakage signals, traditional signal inversion methods based on physical models are unable to accurately quantify defect sizes.

[0004] Fourth, IMU trajectory calculation suffers from cumulative error. Accelerometers exhibit large transient errors during vibration or rapid motion; gyroscopes accumulate drift errors during long-term integration. Fifth, there is a lack of effective data verification mechanisms between deformation detection data and magnetic flux leakage detection data, making it impossible to effectively distinguish between single deformation defects, single metal loss defects, and composite defects. Therefore, a solution capable of integrating deformation, magnetic flux leakage, and IMU detection into a single intelligent analysis is urgently needed. Summary of the Invention

[0005] The technical problem to be solved by this invention is to address the problems of low efficiency, difficulty in analyzing magnetic flux leakage signals, insufficient quantization accuracy, large cumulative error in IMU trajectory calculation, and lack of data verification mechanism in existing pipeline in-detection data analysis methods. This invention provides an AI analysis method and system for pipeline in-detection data based on deep learning, which realizes automated and intelligent analysis of deformation, magnetic flux leakage, and IMU detection data.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides an AI analysis method for pipeline internal inspection data based on deep learning, comprising the following steps: S1: acquiring pipeline internal inspection data, wherein the internal inspection data includes magnetic flux leakage detection signal data and inertial measurement unit (IMU) data, the magnetic flux leakage detection signal data being referred to as magnetic flux leakage signal; S2: preprocessing the magnetic flux leakage signal to extract multi-component magnetic flux leakage signals, wherein the multi-component magnetic flux leakage signals include axial components, circumferential components, and radial components; S3: inputting the preprocessed multi-component magnetic flux leakage signals into a pre-trained deep convolutional recurrent neural network model, wherein the deep convolutional recurrent neural network model includes convolutional layers, recurrent layers, and dense layers; S4: extracting spatial features from the multi-component magnetic flux leakage signals through the convolutional layers, and extracting temporal features from the multi-component magnetic flux leakage signals through the recurrent layers; S5: fusing the extracted spatial features and temporal features through the dense layers to output defect analysis results, wherein the defect analysis results include defect identification results, defect classification results, and defect quantification results.

[0008] This technical solution inputs multi-component magnetic flux leakage signals into a deep convolutional recurrent neural network model containing convolutional, recurrent, and dense layers, enabling automated extraction of spatial and temporal features from high-dimensional magnetic flux leakage time-series signals, as well as automatic identification, classification, and quantification of defects. This reduces the analysis time for every 100 kilometers of pipeline from two months of traditional manual analysis to less than seven days, significantly improving data analysis efficiency and accuracy.

[0009] Furthermore, the leakage magnetic signal is a time-series signal data with high dimension, high frequency, large dynamic range, and extremely low signal-to-noise ratio, wherein the number of channels is greater than 1000, the frequency is greater than 1KHz, the dynamic range is greater than 70dB, and the signal-to-noise ratio is less than 3dB.

[0010] This technical solution targets the characteristics of high-dimensional (>1000 channels), high-frequency (>1KHz), large dynamic range (>70dB), and extremely low signal-to-noise ratio (<3dB) magnetic flux leakage signals. It enables the deep convolutional recurrent neural network model to adapt to such extreme signal conditions, thereby effectively extracting weak defect features that are submerged by noise and improving the applicability and robustness of magnetic flux leakage signal analysis.

[0011] Furthermore, the deep convolutional recurrent neural network model is the TBCRNN model, which includes a tree-structured convolutional network TB-CNN, a convolutional neural network CNN, and a recurrent neural network RNN ​​connected in sequence; the TB-CNN is used to extract tree-structured features from the input multi-component magnetic flux leakage signal, the CNN is used for further convolutional spatial feature extraction, and the RNN is used to extract the temporal dependency features of the magnetic flux leakage signal.

[0012] This technical solution uses the tree-structured convolutional kernel of TB-CNN to extract hierarchical structural features, and then uses CNN and RNN to extract spatial and temporal features respectively. This enables the model to simultaneously capture the multi-scale structural features and long-distance temporal dependencies of the magnetic flux leakage signal, thereby significantly improving the accuracy of identifying complex defect morphologies.

[0013] Furthermore, in step S4, the data fusion of the axial component, circumferential component and radial component in the multi-component leakage magnetic field signal is achieved through the multi-input multi-output cross-correlation convolution calculation in the convolutional neural network.

[0014] This technical solution achieves data fusion of the axial, circumferential, and radial components through multi-input multi-output cross-correlation convolution calculation, making full use of the correlation and complementarity between the multi-component leakage magnetic signals, making the characterization of defect features more comprehensive and accurate, and avoiding the problems of missed detection and misjudgment caused by insufficient information from a single component signal.

[0015] Furthermore, the deep convolutional recurrent neural network model is a DRCRNN model, which is a deep residual convolutional recurrent neural network with more than 40 layers, consisting of convolutional layers, activation layers, pooling layers, recurrent layers, and dense layers stacked together; the recurrent layer includes recurrent units with gate control and recurrent units without gate control, and the DRCRNN model simultaneously realizes qualitative and quantitative analysis of defects.

[0016] This technical solution utilizes a deep residual structure of over 40 layers in DRCRNN, combined with gated and ungated recurrent units, to enable the model to simultaneously output qualitative and quantitative analysis results of defects. This avoids the cascading error accumulation problem of classification followed by quantification in the traditional two-step method, thus obtaining complete defect assessment information in a single inference.

[0017] Furthermore, the recurrent layer in the DRCRNN model includes LSTM units and / or GRU units, and the dense layer includes a classification output branch and a regression output branch. The classification output branch outputs defect identification results and defect classification results, and the regression output branch outputs defect size fitting results. The defect size fitting results are used as a component of the defect quantification results.

[0018] This technical solution, through the parallel design of classification output branches and regression output branches, enables the DRCRNN model to share the parameters of the underlying feature extraction network. While ensuring classification accuracy, it improves the accuracy of defect size fitting, realizes end-to-end joint optimization of qualitative and quantitative defect analysis, and reduces the computational resource consumption of model inference.

[0019] Furthermore, the training method of the deep convolutional recurrent neural network model includes: constructing a sample dataset, which includes simulated magnetic leakage signals and actual detected magnetic leakage signals; using 80% of the sample dataset as a training dataset and 20% as a validation dataset; using the training dataset to train the model through forward propagation; using the validation dataset to calculate the model recognition error; feeding the model recognition error back to the input layer, adjusting the weight parameters in the network model through an optimization algorithm, iterating repeatedly until convergence, and obtaining the trained deep convolutional recurrent neural network model.

[0020] This technical solution constructs a mixed sample dataset containing simulated data and actual detection data, adopts an 80 / 20 training and validation split strategy, and uses backpropagation of validation errors to iteratively optimize model parameters, enabling the model parameters to exceed the million level. This allows the model to fully learn the complex defect feature patterns contained in the leakage magnetic signal, ensuring that the model has good generalization ability and robustness under different pipeline operating conditions.

[0021] Furthermore, the method also includes an IMU trajectory calculation step: converting the IMU data from hexadecimal format to decimal format; converting the IMU acceleration data from the body coordinate system to the world coordinate system; using a fusion algorithm to calculate the gyroscope data, accelerometer data, and odometer data to obtain the IMU attitude estimate and the pipeline centerline trajectory; and detecting abnormal deformation of the pipeline based on the pipeline centerline trajectory.

[0022] This technical solution effectively eliminates the cumulative drift error of the gyroscope and the transient noise error of the accelerometer by converting the format from hexadecimal to decimal, the coordinate transformation from the body coordinate system to the world coordinate system, and the fusion calculation of the gyroscope, accelerometer and odometer. It obtains a stable and accurate attitude estimation and the three-dimensional spatial trajectory of the pipeline centerline, and realizes the accurate detection of abnormal deformation of the pipeline.

[0023] Furthermore, the method also includes a data verification step: by automatically aligning and mutually verifying the deformation detection data and leakage magnetic signals of the same pipe section, the defect type is determined to be a single deformation defect, a single metal loss defect, or a composite defect of deformation and metal damage.

[0024] This technical solution can effectively distinguish between three types of defects—single deformation defects, single metal loss defects, and composite defects of deformation and metal damage—by automatically aligning and cross-verifying deformation detection data and magnetic flux leakage signals of the same pipe section. This avoids unnecessary on-site excavation verification work caused by misjudgment of defect type, thereby reducing the overall cost of pipeline maintenance.

[0025] Secondly, this invention provides an AI analysis system for pipeline in-situ inspection data based on deep learning, comprising: a data acquisition module, a data preprocessing module, a model inference module, an IMU trajectory calculation module, and a result output module. The data acquisition module acquires pipeline in-situ inspection data; the data preprocessing module preprocesses the magnetic flux leakage signal data and extracts multi-component magnetic flux leakage signals; the model inference module inputs the preprocessed multi-component magnetic flux leakage signals into a pre-trained deep convolutional recurrent neural network model and outputs defect analysis results; the IMU trajectory calculation module performs coordinate transformation and fusion calculation on the IMU data to obtain the pipeline centerline trajectory and deformation detection results; and the result output module outputs defect identification results, defect classification results, defect quantification results, and a pipeline integrity evaluation report.

[0026] This technical solution integrates five modules: data acquisition, preprocessing, model inference, IMU trajectory calculation, and result output. It achieves integrated intelligent analysis of deformation detection, magnetic flux leakage detection, and IMU detection. It can achieve the effect of two to three traditional inspections in one pipeline inspection operation, shortening the on-site operation time by about 30% and reducing the on-site operation cost by about 35%.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention inputs multi-component magnetic flux leakage signals into a deep convolutional recurrent neural network model, realizing automated feature extraction and automatic defect identification, classification, and quantification, reducing data analysis time from two months to less than seven days.

[0029] 2. This invention applies the TBCRNN model (TB-CNN+CNN+RNN) to magnetic flux leakage signal analysis, solving the problem that traditional deep learning models in the image domain cannot directly process one-dimensional magnetic flux leakage time-series signals.

[0030] 3. This invention improves the ability to identify complex defect morphologies by extracting structured and temporal features through TB-CNN tree convolution + CNN + RNN respectively.

[0031] 4. This invention achieves triaxial magnetic flux leakage signal data fusion through cross-correlation convolution, and improves the comprehensiveness of feature extraction by utilizing the complementarity between components.

[0032] 5. This invention uses the DRCRNN model (40+ layers) to achieve simultaneous qualitative and quantitative analysis through classification and regression dual-branch output, thereby reducing cascade errors.

[0033] 6. This invention captures long-distance temporal dependencies through an LSTM / GRU gating mechanism, thereby improving classification and regression accuracy.

[0034] 7. This invention employs an 80 / 20 training and validation strategy for iterative optimization, resulting in a model with over a million parameters and strong generalization ability.

[0035] 8. This invention eliminates cumulative drift and transient noise errors through coordinate transformation and fusion calculation using gyroscope / accelerometer / odometer.

[0036] 9. This invention effectively distinguishes between single defects and compound defects by automatically aligning and verifying deformation and magnetic leakage signals, thus avoiding misjudgment.

[0037] 10. The five modules of this invention are integrated to achieve unified intelligent analysis, shortening on-site operation time by 30% and reducing operation costs by 35%. Attached Figure Description

[0038] Figure 1 This is a diagram illustrating the overall architecture of the AI-based data analysis system.

[0039] Figure 2 This is a schematic diagram of the TBCRNN model structure;

[0040] Figure 3 Here is the overall flowchart of the algorithm;

[0041] Figure 4 This is a functional module architecture diagram for data analysis software.

[0042] Figure 5 Flowchart for IMU trajectory calculation;

[0043] Figure 6 This is a flowchart of the model training and validation process. Detailed Implementation

[0044] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0045] Example 1: Pipeline In-Process Inspection Data Analysis Method Based on TBCRNN Model

[0046] like Figure 1As shown, the detection data AI analysis system of this embodiment includes a pipe in-situ detector 1, a data acquisition module 2, a data preprocessing module 3, a model inference module 4, an IMU trajectory calculation module 5, and a result output module 6. The pipe in-situ detector 1 integrates a deformation detection sensor, a magnetic flux leakage detection sensor, and an inertial measurement unit (IMU), and simultaneously collects deformation detection data, magnetic flux leakage signals, and IMU data during a single pipe in-situ inspection operation.

[0047] Step S1: Acquire in-pipe detection data. When the in-pipe detector 1 operates inside the pipe, the deformation detection sensor collects real-time data on the deformation of the pipe's inner diameter, the magnetic flux leakage detection sensor collects magnetic flux leakage signal data from the pipe wall, and the IMU simultaneously collects data from the three-axis acceleration channel, the three-axis angular rate channel, and the odometer pulse count. The data is stored in digital signal form through the in-pipe detector's storage unit and exported to the data acquisition module 2 after detection is complete.

[0048] Step S2: Data Preprocessing. The data preprocessing module 3 first denoises the original magnetic flux leakage signal, using bandpass filtering to remove low-frequency baseline drift and high-frequency electromagnetic interference noise. Then, the denoised signal is normalized, mapping the signal amplitude to the [0, 1] interval. Finally, multi-component magnetic flux leakage signals are extracted, including axial, circumferential, and radial components. The magnetic flux leakage signal is a high-dimensional (>1000 channels), high-frequency (>1kHz), large dynamic range (>70dB), and extremely low signal-to-noise ratio (<3dB) time-series signal data.

[0049] Step S3: Model Input. The preprocessed multi-component magnetic flux leakage signal is input into the pre-trained TBCRNN model in model inference module 4. For example... Figure 2 As shown, the TBCRNN model consists of a TB-CNN tree-structured convolutional network, a CNN convolutional neural network, and an RNN recurrent neural network connected sequentially. TB-CNN uses tree-structured convolutional kernels to extract hierarchical structural features from the magnetic flux leakage signal. The CNN achieves data fusion of the axial, circumferential, and radial components of the multi-component magnetic flux leakage signal through multi-input multi-output cross-correlation convolution calculations.

[0050] Step S4: Feature Extraction. The RNN (Recurrent Neural Network) uses an LSTM or GRU structure to receive the spatial feature sequence output by the CNN and model and extract the temporal dependency features in the magnetic flux leakage signal. The gating mechanism of LSTM / GRU can effectively capture long-distance temporal dependencies, solving the gradient vanishing problem of traditional RNNs when processing long sequence signals.

[0051] Step S5: Feature Fusion and Result Output. The spatial features extracted by the CNN and the temporal features extracted by the RNN are concatenated and fused in the feature fusion layer to form a high-dimensional feature vector containing joint spatial-temporal information. The fused feature vector is then processed through a dense fully connected layer to finally output the defect analysis results, including defect identification results, defect classification results, and defect quantification results.

[0052] Model training methods

[0053] like Figure 6 As shown, the training process of the TBCRNN model is as follows: First, a sample dataset is constructed, including simulated magnetic flux leakage signals generated through finite element simulation and actual detected magnetic flux leakage signals collected through tensile tests and actual pipeline inspections. Defect types and sizes are manually labeled frame-by-frame in the sample data. 80% of the sample dataset is used as the training dataset, and the remaining 20% ​​as the validation dataset. The model is trained using the training dataset via forward propagation, and the loss function value between the output result and the labeled ground truth is calculated. The model recognition error is calculated using the validation dataset, and the error is fed back to the input layer. Backpropagation is then used to adjust the weight parameters in the network model through optimization algorithms such as Adam or SGD. This process is iterated repeatedly until the loss function on the validation dataset converges to a preset threshold, resulting in the trained deep convolutional recurrent neural network model. The model parameter scale can exceed the millions.

[0054] Example 2: Pipeline In-Process Inspection Data Analysis Method Based on DRCRNN Model

[0055] The difference between this embodiment and Embodiment 1 is that the deep convolutional recurrent neural network model in the model inference module 4 adopts the DRCRNN model. The DRCRNN model is a deep residual convolutional recurrent neural network with more than 40 layers, consisting of multiple convolutional layers, activation layers, pooling layers, recurrent layers, and dense layers stacked together. The recurrent layers include gated LSTM units and / or GRU units, as well as basic recurrent units without gates. The dense layers of the DRCRNN model contain two parallel output branches: a classification output branch outputs defect identification and classification results through a Softmax activation function; and a regression output branch outputs defect size fitting results through a linear activation function. This technical solution enables the DRCRNN model to simultaneously perform qualitative and quantitative analysis of defects.

[0056] IMU trajectory calculation process

[0057] like Figure 5As shown, the IMU trajectory calculation module 5 processes the IMU data as follows: First, the hexadecimal IMU data downloaded from the internal detector is converted to decimal data. The IMU data includes clock counts, three-axis acceleration channel data, three-axis angular rate channel data, odometer pulse counts, three-axis acceleration increment data, and three-axis angular acceleration increment data. Since the acceleration data is represented in the Body coordinate system, it needs to be transformed from the Body coordinate system to the world coordinate system using the attitude quaternion or direction cosine matrix at the corresponding moment. Before the transformation, the accelerometer bias is subtracted, and after the transformation, the gravitational acceleration is subtracted. Based on this, a fusion algorithm is used to fuse and calculate the gyroscope data, accelerometer data, and odometer data. The high dynamic response of the gyroscope is used to track rapid attitude changes, the long-term stability of the accelerometer is used to compensate for the cumulative drift error of the gyroscope, and the odometer provides a displacement reference. A stable and accurate attitude estimate is obtained through Kalman filtering or complementary filtering. Based on accurate attitude estimation, the displacement is obtained by quadratic integration of acceleration data, and then corrected by odometer data. Finally, the three-dimensional spatial trajectory of the pipeline centerline is obtained, and abnormal deformation of the pipeline is detected by curvature analysis.

[0058] Data verification process

[0059] The result output module 6 automatically aligns and cross-verifies the deformation detection data and magnetic flux leakage signal. For example... Figure 4 As shown, using pipeline mileage as a common benchmark, deformation detection data and magnetic flux leakage signals are automatically aligned spatially using mileage wheel data and IMU trajectory data. A point-by-point comparative analysis is then performed on the aligned data: if an abnormal signal is found only in the deformation data at a certain location, it is identified as a single deformation defect; if an abnormal signal is found only in the magnetic flux leakage signal, it is identified as a single metal loss defect; if both abnormal signals are present, it is identified as a combined deformation and metal damage defect. The system automatically generates a pipeline integrity evaluation report, including a defect statistics list, defect statistics charts, pipeline information list, remaining life calculation, ERF calculation, and on-site excavation recommendations. This integrated intelligent analysis process reduces the total analysis time for each 100 kilometers of pipeline from two months of traditional manual analysis to less than seven days.

[0060] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based AI analysis method for pipeline in-situ detection data, characterized in that, Includes the following steps: S1: Acquire internal pipeline detection data, which includes magnetic flux leakage detection signal data and inertial measurement unit (IMU) data. The magnetic flux leakage detection signal data is referred to as magnetic flux leakage signal. S2: Preprocess the magnetic flux leakage signal to extract multi-component magnetic flux leakage signal, which includes axial component, circumferential component and radial component. S3: Input the preprocessed multi-component magnetic flux leakage signal into a pre-trained deep convolutional recurrent neural network model, wherein the deep convolutional recurrent neural network model includes convolutional layers, recurrent layers, and dense layers; S4: Extract spatial features from the multi-component magnetic flux leakage signal through the convolutional layers, and extract temporal features from the multi-component magnetic flux leakage signal through the recurrent layers; S5: The extracted spatial and temporal features are fused through the dense layer to output defect analysis results, which include defect identification results, defect classification results, and defect quantification results.

2. The method according to claim 1, characterized in that, The leakage magnetic signal is a time-series signal data with high dimension, high frequency, large dynamic range, and extremely low signal-to-noise ratio. It has more than 1,000 channels, a frequency greater than 1,000 kHz, a dynamic range greater than 70 dB, and a signal-to-noise ratio less than 3 dB.

3. The method according to claim 1, characterized in that, The deep convolutional recurrent neural network model is the TBCRNN model, which includes a tree-structured convolutional network TB-CNN, a convolutional neural network CNN, and a recurrent neural network RNN ​​connected in sequence. The TB-CNN is used to extract tree-structured features from the input multi-component magnetic flux leakage signal, the CNN is used for further convolutional spatial feature extraction, and the RNN is used to extract the temporal dependency features of the magnetic flux leakage signal.

4. The method according to claim 3, characterized in that, In step S4, the data fusion of the axial, circumferential and radial components in the multi-component leakage magnetic field signal is achieved through the multi-input multi-output cross-correlation convolution calculation in the convolutional neural network.

5. The method according to claim 1, characterized in that, The deep convolutional recurrent neural network model is the DRCRNN model, which is a deep residual convolutional recurrent neural network with more than 40 layers. It is composed of convolutional layers, activation layers, pooling layers, recurrent layers, and dense layers stacked together. The recurrent layers include recurrent units with gate control and recurrent units without gate control. The DRCRNN model can simultaneously realize qualitative and quantitative analysis of defects.

6. The method according to claim 5, characterized in that, The recurrent layer in the DRCRNN model includes LSTM units and / or GRU units, and the dense layer includes a classification output branch and a regression output branch. The classification output branch outputs defect identification results and defect classification results, and the regression output branch outputs defect size fitting results. The defect size fitting results are a component of the defect quantification results.

7. The method according to claim 1, characterized in that, The training method for the deep convolutional recurrent neural network model includes: constructing a sample dataset, which includes simulated magnetic leakage signals and actual detected magnetic leakage signals; using 80% of the sample dataset as the training dataset and 20% as the validation dataset; using the training dataset to train the model through forward propagation; using the validation dataset to calculate the model recognition error; feeding the model recognition error back to the input layer, adjusting the weight parameters in the network model through an optimization algorithm, iterating repeatedly until convergence, and obtaining the trained deep convolutional recurrent neural network model.

8. The method according to claim 1, characterized in that, The method further includes an IMU trajectory calculation step: converting the IMU data from hexadecimal format to decimal format; converting the IMU acceleration data from the body coordinate system to the world coordinate system; using a fusion algorithm to calculate the gyroscope data, accelerometer data, and odometer data to obtain the IMU attitude estimate and the pipeline centerline trajectory; and detecting abnormal deformation of the pipeline based on the pipeline centerline trajectory.

9. The method according to claim 1, characterized in that, The method also includes a data verification step: by automatically aligning and mutually verifying the deformation detection data and leakage magnetic signals of the same pipe section, the defect type is determined to be a single deformation defect, a single metal loss defect, or a composite defect of deformation and metal damage.

10. A deep learning-based AI analysis system for pipeline in-situ detection data, characterized in that, include: The data acquisition module is used to acquire internal pipeline detection data, which includes leakage magnetic field detection signal data and inertial measurement unit (IMU) data. The data preprocessing module is used to preprocess the magnetic flux leakage detection signal data and extract multi-component magnetic flux leakage signals. The model inference module is used to input the preprocessed multi-component magnetic flux leakage signal into a pre-trained deep convolutional recurrent neural network model and output the defect analysis results. The IMU trajectory calculation module is used to perform coordinate transformation and fusion calculation on the IMU data to obtain the pipeline centerline trajectory and deformation detection results; The results output module is used to output defect identification results, defect classification results, defect quantification results, and pipeline integrity evaluation reports.

Images