Intelligent detection and investigation method and system for a subsea pipeline

Abstract

The application relates to the technical field of intelligent detection and analysis, in particular to an intelligent detection and investigation method and system for a submarine pipeline. The method comprises the following steps: collecting acoustic profile data, space-time coordinate data, ship body motion attitude data and design route planning data corresponding to each submarine pipeline in a target sea area; performing acoustic feature enhancement processing on the acoustic profile data and intelligently identifying the pipeline target, and calculating and generating a preliminary pipeline suspected target set; fusing the space-time coordinate data, the ship body motion attitude data and the preliminary pipeline suspected target set to convert from an instrument coordinate system to a geodetic coordinate system, and generating a pipeline target point cloud set; performing pipeline route correlation analysis, point cloud sorting and topological connection based on the design route planning data and the pipeline target point cloud set, simultaneously performing pipeline space state compliance analysis, and predicting and generating a pipeline long-term service performance degradation prediction curve and a maintenance priority list. The application can improve the spatial detection efficiency of the submarine pipeline.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and analysis technology, specifically to an intelligent detection and investigation method and system for submarine pipelines. Background Technology

[0002] With the deepening development of offshore oil and gas resources, subsea pipelines, as the core hub for oil and gas transportation, are directly related to the normal progress of oilfield development projects, the safety of the marine ecological environment, and the safety of personnel and property. Especially in near-shore oilfield areas, the seabed topography is complex, the water depth varies greatly, and natural factors such as ocean currents, tides, and seabed sediment migration can affect subsea pipelines, making them prone to subsidence, overhang, corrosion, and other hidden dangers. Therefore, regular exploration and surveying of subsea pipelines to determine their location, burial depth, spatial status, and potential risks has become an indispensable and crucial link in the offshore oil and gas development process. However, existing exploration methods have low automation levels, cumbersome data processing procedures, and rely on manual intervention. They often use a single data processing approach without achieving collaborative fusion and attitude compensation of multi-source data. This results in the inability to convert suspected pipeline targets detected in the initial exploration to a geodetic coordinate system, making it difficult to efficiently correlate them with the designed route. This affects the accuracy of pipeline spatial trajectory reconstruction and topological connection, thereby reducing the exploration efficiency of subsea pipelines. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide an intelligent detection and investigation method and system for submarine pipelines to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, an intelligent detection and survey method for subsea pipelines includes the following steps: Collect multi-source heterogeneous detection data corresponding to each submarine pipeline in the target sea area. The multi-source heterogeneous detection data includes acoustic profile data collected by a shallow seismic profiler mounted on the detection vessel, spatiotemporal coordinate data collected by a differential GNSS positioning system, ship motion attitude data collected by an attitude inertial navigation system, and design route planning data recorded by navigation software. Acoustic feature enhancement processing is performed on the acoustic profile data to generate an enhanced acoustic profile image sequence. Intelligent pipe target recognition is performed on the enhanced acoustic profile image sequence to extract abnormal reflection area pixel clusters that characterize the pipe contour in the image, and calculate their geometric center coordinates, equivalent diameter and reflection intensity distribution to generate a preliminary set of suspected pipe targets. By integrating spatiotemporal coordinate data, ship motion attitude data, and a preliminary set of suspected pipeline targets, the preliminary set of suspected pipeline targets is transformed from the instrument coordinate system to the geodetic coordinate system through coordinate transformation and attitude compensation algorithms, generating a cloud of pipeline target points. Based on the design route planning data and the pipeline target point cloud, pipeline route correlation analysis is performed. By calculating the shortest vertical distance between the point cloud and the design route line, point clouds with a shortest vertical distance less than a set threshold are associated with the same pipeline. The point clouds are sorted and topologically connected along the pipeline extension direction to generate a continuous three-dimensional spatial trajectory and burial depth attribute dataset for each subsea pipeline. Based on the continuous three-dimensional spatial trajectory and burial depth attribute datasets corresponding to each subsea pipeline, a pipeline spatial status compliance analysis is conducted to generate a pipeline spatial status assessment report. Using historical survey data and the pipeline spatial status assessment report, a pipeline status evolution prediction model is constructed to predict the pipeline's settlement trend, overhang risk, and corrosion rate within a specified future period, generating a pipeline long-term service performance degradation prediction curve and a maintenance priority list.

[0005] Furthermore, the present invention also provides an intelligent detection and survey system for submarine pipelines, including a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the intelligent detection and survey method for submarine pipelines as described above.

[0006] The beneficial effects of this invention are: The intelligent detection and survey method for subsea pipelines proposed in this invention, compared with existing technologies, offers the following advantages: It collects multi-source heterogeneous detection data from various subsea pipelines in the target sea area, encompassing acoustic profiles, spatiotemporal coordinates, ship motion attitude, and design route planning data. This step comprehensively collects four core types of multi-source heterogeneous data: acoustic profile data reflects the characteristics of the pipeline and surrounding geological strata; spatiotemporal coordinate data provides a benchmark for the detection location; ship motion attitude data compensates for errors caused by ship swaying; and design route planning data provides a reference for pipeline design standards. The synchronous acquisition of multi-source data breaks the limitations of single-data acquisition, constructing a complete and systematic detection dataset. This provides comprehensive data support for subsequent acoustic feature enhancement, coordinate transformation, route correlation, and status assessment, addressing the problems of data dispersion and insufficient support from the source, and laying the foundation for improving detection efficiency and accuracy. Secondly, by enhancing the acoustic profile data, an enhanced image sequence is generated. Then, through intelligent recognition, pixel clusters of abnormal reflection areas in the pipeline outline are extracted. Relevant geometric parameters are calculated to generate a preliminary set of suspected pipeline targets. This step, through acoustic feature enhancement processing, highlights the pipeline reflection signal and suppresses interference clutter, making the pipeline outline clearer and improving the identifiability of the pipeline target. Relying on intelligent recognition technology to automatically extract pixel clusters of abnormal reflection areas, replacing manual interpretation, significantly improves recognition efficiency. At the same time, key parameters such as geometric center coordinates and equivalent diameter are calculated to form a standardized preliminary set of suspected pipeline targets, clarifying the core features of suspected pipeline targets and providing a target basis for subsequent coordinate transformation and route association. By integrating spatiotemporal coordinates, ship motion and attitude data, and a preliminary set of suspected pipeline targets, coordinate transformation and attitude compensation algorithms are used to convert the suspected target set from the instrument coordinate system to the geodetic coordinate system, generating a pipeline target point cloud. This step, by fusing three types of data and using attitude compensation algorithms to correct errors caused by ship swaying, ensures the accuracy of suspected target positioning. Then, coordinate transformation is used to convert the suspected targets from the instrument coordinate system to the geodetic coordinate system, giving each suspected target a clear geospatial coordinate, generating a standardized pipeline target point cloud. This process overcomes the limitations of fragmented multi-source data, achieving collaborative data fusion and error correction, and solving the problem of suspected targets not corresponding to actual geographical locations, thereby improving the spatial detection efficiency of subsea pipelines. Then, route correlation analysis is conducted based on the design route planning data and the pipeline target point cloud. The shortest vertical distance is used to associate points along the same pipeline, and the point clouds are sorted and topologically connected. This step resolves the issues of point cloud chaos and mismatch by calculating the shortest vertical distance between the point cloud and the design route line. The point clouds are then sorted and topologically connected along the pipeline extension direction to reconstruct a continuous and complete 3D spatial trajectory of the pipeline. Simultaneously, burial depth attributes are extracted to generate a burial depth attribute dataset. This process achieves efficient correlation between actual probe point clouds and design routes, overcoming the limitation of point clouds being disconnected from design standards and solving the problems of discontinuous pipeline trajectory reconstruction and missing burial depth attributes.Finally, a spatial status compliance analysis was conducted based on the pipeline's continuous three-dimensional spatial trajectory and burial depth attribute dataset. An assessment report was generated, and a status evolution prediction model was constructed using historical data to predict pipeline hazards and generate degradation curves and a maintenance priority list. This step, through compliance analysis, comprehensively assesses whether the pipeline's spatial status meets design standards, generates a standardized assessment report, and clarifies the existing visible hazards. The prediction model, built using historical survey data, predicts future pipeline settlement trends, overhang risks, and corrosion rates, generating degradation curves and a maintenance priority list. This provides a scientific basis for maintenance work, achieving a closed-loop process encompassing pipeline detection, status assessment, risk prediction, and maintenance guidance. This enhances the foresight and targeting of maintenance work, ensures the long-term safe and stable operation of subsea pipelines, and reduces the risk of safety accidents. Attached Figure Description

[0007] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the steps of the intelligent detection and investigation method for submarine pipelines of the present invention; Figure 2 for Figure 1 A detailed flowchart illustrating the steps involved in generating the pipeline target point cloud as described in S03. Detailed Implementation

[0008] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0009] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides an intelligent detection and investigation method for subsea pipelines. In the embodiments of this invention, please refer to... Figure 1 The diagram shown is a flowchart illustrating the steps of the intelligent detection and survey method for submarine pipelines according to the present invention. In this example, the intelligent detection and survey method for submarine pipelines includes the following steps: S01: Collect multi-source heterogeneous detection data corresponding to each submarine pipeline in the target sea area. The multi-source heterogeneous detection data includes acoustic profile data collected by the shallow seismic profiler mounted on the detection vessel, spatiotemporal coordinate data collected by the differential GNSS positioning system, ship motion attitude data collected by the attitude inertial navigation system, and design route planning data recorded by the navigation software. In this embodiment of the invention, multi-source heterogeneous detection data corresponding to seven subsea pipelines within the target sea area are collected. The detection vessel navigates along a preset route at a speed controlled at 4 knots, and the detection coverage area completely covers the designed routes of the seven pipelines and the surrounding 50m area. The specific acquisition method and parameter settings for the multi-source heterogeneous detection data are as follows: a shallow seismic profiler is mounted on the bottom of the detection vessel, with a transmission frequency set to 100kHz and a sampling interval set to 0.1ms, collecting acoustic profile data within a 50m range below the seabed. The data format is continuous analog signal, recording seabed strata layering information and abnormal reflection signals generated by the pipelines; a differential GNSS positioning system is mounted on the deck of the detection vessel, with a positioning frequency set to 10Hz, using the CGCS2000 geodetic coordinate system and UTM projection (scale factor 0.9996), collecting real-time spatiotemporal coordinate data of the detection vessel, including north coordinates (X), east coordinates (X), and east coordinates (X). The data includes Y), elevation (Z), and acquisition timestamp, with coordinates in meters (m). The attitude inertial navigation system is fixedly connected to the hull of the probe vessel, with a sampling frequency set to 20Hz. It collects hull motion attitude data, including roll angle, pitch angle, and heading angle, in degrees, to compensate for the impact of hull turbulence on the probe data. The navigation equipment records design route planning data, including the coordinates of the design route inflection points, pipe diameter, and laying depth for each of the seven pipelines. The design route inflection point coordinates use the same coordinate system as the differential GNSS positioning system. The pipe diameter is uniformly 0.6m, and the design laying depth ranges from -11.8m to -13.6m. All probe data are recorded synchronously according to the acquisition timestamp. Each set of synchronous data includes acoustic profile data, spatiotemporal coordinate data, hull motion attitude data, and the corresponding pipeline design route planning data at the corresponding time, ensuring data temporal consistency and correlation.

[0010] S02: Perform acoustic feature enhancement processing on the acoustic profile data to generate an enhanced acoustic profile image sequence; perform intelligent pipe target recognition on the enhanced acoustic profile image sequence to extract abnormal reflection area pixel clusters that characterize the pipe contour in the image, and calculate their geometric center coordinates, equivalent diameter and reflection intensity distribution to generate a preliminary set of suspected pipe targets. In this embodiment of the invention, the original acoustic profile data is denoised using a moving average filtering method with a filter window size of 5 sampling points. The average value of the sampling points within the window is used to replace the center sampling point, thus eliminating seabed environmental noise and instrument noise. Then, contrast enhancement is performed using histogram equalization to stretch the grayscale distribution of the acoustic profile data to the 0-255 range, improving the grayscale difference between the abnormal reflection area of ​​the pipeline and the seabed background. Finally, edge enhancement is performed using the Sobel operator, which calculates the grayscale gradient between adjacent sampling points to enhance the grayscale changes at the pipeline contour edges, generating an enhanced acoustic profile image sequence with a resolution of 1024×768 pixels and a frame rate of 1 frame / second. Intelligent pipe target recognition is performed on the enhanced acoustic profile image sequence. The recognition process is as follows: Based on the enhanced image sequence, pixels with gray values ​​greater than 200 are extracted (these pixels correspond to strong reflection signals generated by the pipe), forming abnormal reflection area pixel clusters; through morphological closing operations (structural element is a 3×3 pixel square), tiny holes inside the pixel clusters are filled, and small noise pixel clusters with an area of ​​less than 50 pixels are removed; the geometric center coordinates of each effective pixel cluster are calculated (with the upper left corner of the image as the origin, the horizontal axis as the x-axis, and the vertical axis as the y-axis, and the coordinate unit as pixels), and the equivalent diameter (the area of ​​the pixel cluster is calculated). Clusters are approximately circular, with the equivalent diameter calculated based on the cluster area (in pixels). Reflectance intensity distribution is also considered (by statistically analyzing the grayscale value of each pixel within the cluster and calculating the average and standard deviation of the grayscale values). Clusters with geometric center coordinates corresponding to a range of 0.5-2m below the seabed along the vertical axis of the image, an equivalent diameter corresponding to an actual size of 0.5-0.7m, and a standard deviation of reflection intensity less than 30 are identified as the clusters corresponding to pipeline targets. All such clusters are then aggregated to generate a preliminary set of suspected pipeline targets. This set includes the image coordinates, equivalent diameter, reflection intensity distribution, and corresponding acquisition timestamp for each suspected target.

[0011] S03: Integrate spatiotemporal coordinate data, ship motion attitude data and preliminary pipeline target set, and transform the preliminary pipeline target set from the instrument coordinate system to the geodetic coordinate system through coordinate transformation and attitude compensation algorithms to generate pipeline target point cloud; In this embodiment of the invention, spatiotemporal coordinate data collected by a differential GNSS positioning system, ship motion attitude data collected by an attitude inertial navigation system, and a preliminary set of suspected pipeline targets are fused. Coordinate transformation and attitude compensation algorithms are then used to achieve coordinate system transformation of the preliminary set of suspected pipeline targets. First, attitude compensation is performed. Based on the roll angle, pitch angle, and heading angle in the ship motion attitude data, the image coordinates of the preliminary set of suspected pipeline targets are compensated and corrected to eliminate target position offsets caused by ship turbulence. Then, coordinate transformation is performed, establishing a transformation relationship between the instrument coordinate system and the geodetic coordinate system. The instrument coordinate system has its origin at the shallow seismic profiler transmitter, with the horizontal axis parallel to the ship's heading, the vertical axis perpendicular to the seabed plane downwards, and the vertical axis perpendicular to the plane formed by the horizontal and vertical axes. Through the above attitude compensation and coordinate transformation, the image coordinates of all preliminary suspected pipeline targets are transformed to the geodetic coordinate system, obtaining the geodetic coordinates of each suspected target. All geodetic coordinates are then aggregated. The system generates a cloud of target points for the pipeline. Each point cloud contains geodetic coordinates, equivalent diameter, reflection intensity, and acquisition timestamp to ensure that the point cloud data is consistent with the actual pipeline location.

[0012] S04: Based on the design route planning data and the pipeline target point cloud, conduct pipeline route correlation analysis. By calculating the shortest vertical distance between the point cloud and the design route line, point clouds with a shortest vertical distance less than a set threshold are associated with the same pipeline. Then, sort and connect the point clouds along the pipeline extension direction to generate a continuous three-dimensional spatial trajectory and burial depth attribute dataset for each subsea pipeline. In this embodiment of the invention, seven design route lines for each pipeline are extracted from the design route planning data. Each design route line is formed by connecting a series of ordered inflection point coordinates, with the inflection points connected sequentially along the pipeline extension direction to form a continuous three-dimensional spatial polyline. Then, the shortest vertical distance from each point cloud within the pipeline target point set to each design route line is calculated. The calculation method uses the formula for the vertical distance from a spatial point to a spatial polyline: assuming the design route line starts from an inflection point... , ... Point clouds are formed by connecting them sequentially. ,calculate To each line segment ( =1,2,..., The minimum vertical distance among all distances is taken as the shortest vertical distance from the point cloud to the design route. A distance association threshold of 0.8m is set, determined based on the pipeline diameter and detection error. Point clouds with a shortest vertical distance less than 0.8m are associated with the pipeline corresponding to the design route, while those with a shortest vertical distance greater than or equal to 0.8m are considered interference points and discarded. For each pipeline-associated point cloud, it is sorted along the pipeline's design route extension direction. The sorting method is as follows: each point cloud is projected onto the corresponding design route line, the cumulative distance between the projected point and the starting point of the design route is calculated, and the point clouds are sorted in ascending order of cumulative distance, resulting in an ordered sequence of point clouds along the pipeline extension direction. Topological connections are performed on the ordered point cloud sequence, and a cubic B-spline curve fitting method is used. Based on the geodetic coordinates of the point clouds, a continuous and smooth three-dimensional spatial curve is generated, which represents the continuous three-dimensional spatial trajectory of the corresponding subsea pipeline. Simultaneously, based on the geodetic elevation of the point clouds... By combining acoustic profile data collected by a shallow seismic profiler, the depth difference between the top of the pipeline and the seabed interface at the corresponding point cloud location was interpreted. According to the formula Calculate the pipe burial depth ( The elevation of the seabed interface at the location corresponding to the point cloud (obtained by interpreting acoustic profile data) is used to summarize the burial depth data of all point clouds and generate a burial depth attribute dataset for each subsea pipeline. This dataset contains point cloud coordinates, burial depth, equivalent diameter, and reflection intensity information.

[0013] S05: Based on the continuous three-dimensional spatial trajectory and burial depth attribute datasets corresponding to each subsea pipeline, conduct pipeline spatial status compliance analysis and generate pipeline spatial status assessment reports; utilize historical survey data and pipeline spatial status assessment reports to construct a pipeline status evolution prediction model to predict the pipeline's settlement trend, overhang risk, and corrosion rate within a specified future period, and generate a pipeline long-term service performance degradation prediction curve and maintenance priority list.

[0014] In this embodiment of the invention, the actual pipeline route offset is the maximum vertical distance between the actual three-dimensional spatial trajectory and the designed route line, and the offset is specified not to exceed 1.0m; the burial depth compliance is the difference between the actual burial depth and the designed burial depth, and the absolute value of the difference is specified not to exceed 0.5m; the pipeline spacing compliance is the minimum distance between the actual trajectories of two adjacent pipelines, and the minimum distance is specified not to be less than 5.0m. Each compliance indicator for the seven pipelines is calculated, and the locations and values ​​of exceeding the standards are recorded. A pipeline spatial status assessment report is generated, which includes a continuous three-dimensional spatial trajectory map of each pipeline, a burial depth distribution histogram, compliance indicator detection results, and details of the locations of exceeding the standards. Using historical survey data (pipeline trajectory, burial depth, corrosion detection, and marine environmental data for the past 10 years) from the pipeline life cycle status evolution database and the current status data from the pipeline spatial status assessment report, a pipeline status evolution prediction model is constructed. This model integrates sub-models for predicting settlement rate, fatigue life of suspended sections, and corrosion rate. Using the current pipeline condition (current trajectory, burial depth, location of suspended sections, pipe wall thickness, etc.) as initial conditions, and driven by predicted marine environmental data for the target sea area over the next five years (ocean current velocity, sediment transport rate, wave load, seawater chemistry, etc.), it simulates the pipeline's evolution over the next five years. The settlement rate prediction sub-model calculates the cumulative settlement at each mileage point, generating a settlement trend map; the suspended section fatigue life prediction sub-model calculates the cumulative fatigue damage of the suspended section, generating a suspension risk probability distribution map; and the corrosion rate prediction sub-model calculates the pipe wall thickness loss, generating a remaining wall thickness distribution map and a long-term service performance degradation prediction curve. Finally, all prediction results are integrated, and maintenance priorities are assigned according to risk level. High-risk suspended sections, high-corrosion-rate areas, and high-settlement-rate areas are classified as Level 1 maintenance priorities; medium-risk areas as Level 2 maintenance priorities; and low-risk areas as Level 3 maintenance priorities. This generates a maintenance priority list, clearly defining the location, maintenance content, and maintenance time nodes for each maintenance area, achieving full life-cycle management of the pipeline.

[0015] Furthermore, the acoustic feature enhancement processing of the acoustic profile data includes the following steps: Short-time Fourier transform is performed on historical acoustic profile data to convert one-dimensional time-domain signals into two-dimensional time-frequency spectrograms, generating a time-frequency domain acoustic feature matrix. In the time-frequency domain acoustic feature matrix, an adaptive thresholding method is used to identify and mark continuous low-frequency high-energy regions formed by seawater reverberation and random noise, generating a noise energy mask map. In this embodiment of the invention, historical acoustic profile data is collected during the intelligent detection and investigation of subsea pipelines. The data duration is determined according to the actual needs of the subsea pipeline detection operation. The sampling frequency is derived from the main frequency range of the subsea pipeline acoustic signal—the main frequency of the commonly used acoustic signal for subsea pipeline detection is 0-5kHz. According to the Nyquist sampling theorem, the sampling frequency must be greater than twice the highest main frequency. Therefore, the sampling frequency is set to 10kHz. The number of sampling points per data segment is calculated from the sampling frequency and the data duration, i.e., number of sampling points = sampling frequency × data duration. Based on this, a one-dimensional time-domain signal sequence is constructed. ,in The time variable is used, and the sampling interval is the reciprocal of the sampling frequency. A short-time Fourier transform is performed on this one-dimensional time-domain signal. The sliding window length is derived based on the time resolution requirements of the acoustic signal. Considering the signal's dominant frequency characteristics, an integer power of 2 is chosen as the window length to improve transform efficiency. The final window length is determined, and a Hanning window is used to reduce spectral leakage. The overlap rate is set based on the balance between time and frequency resolution. Experiments have verified that when the overlap rate is 50%, time-domain information loss can be avoided while maintaining frequency resolution. That is, the number of overlapping points between adjacent sliding windows is 50% of the window length. This is achieved through the formula... Complete the transformation, where For time-frequency domain signals, The time frame number, For frequency point number, This represents the number of overlapping sampling points. For the Hanning window function, For the length of the window, The unit is the imaginary unit. The transformed two-dimensional time-frequency spectrogram has the number of frames in the time dimension calculated from the data duration, window length, and overlap rate: frame number = (data duration × sampling frequency - window length) / (window length × (1 - overlap rate)) + 1. The number of points in the frequency dimension is 1 / 2 + 1 of the window length. The amplitude matrix of this time-frequency spectrogram is used as the time-frequency domain acoustic feature matrix. Based on this feature matrix, an adaptive thresholding method is used to identify noise regions and calculate the average energy within the frequency range of each frame. and standard deviation Set threshold This threshold setting is based on the characteristics of a normal distribution and can cover more than 95% of the normal noise energy range. When the amplitude at a certain frequency point is greater than... When the signal is determined to be valid, the region is less than or equal to the specified value. The noise region is identified as a continuous low-frequency, high-energy region formed by seawater reverberation and random noise. This region is marked as 1, while the effective signal region is marked as 0. A noise energy mask map with the same dimension as the feature matrix is ​​generated to achieve accurate positioning of the noise region and lay the foundation for effective identification of submarine pipeline detection signals.

[0016] A denoising model based on a deep convolutional neural network is constructed. The time-frequency domain acoustic feature matrix is ​​used as input, and the noise energy mask is used as a supervision label. The model is trained to learn the distribution differences between noise and effective signal, and a well-trained intelligent denoising model is generated. In this embodiment of the invention, a lightweight deep convolutional neural network denoising model is constructed. The overall structure of the model includes an input layer, three convolutional blocks, one pooling layer, one deconvolutional layer, and an output layer. The total number of model parameters is controlled within 5 million. This upper limit of parameters is derived based on the computing power requirements of the submarine pipeline detection equipment. Combined with the computing power level of the embedded processor of the on-site detection equipment, it has been determined through testing that 5 million parameters can ensure the denoising effect while realizing the model's rapid inference, thus adapting to the lightweight operation requirements of intelligent detection and investigation of submarine pipelines. The input layer receives the generated time-frequency domain acoustic feature matrix, the dimension of which is determined by the short-time Fourier transform parameters. The number of convolutional kernels in the first convolutional block is set according to the complexity of the input features. Considering the sparsity of the acoustic feature dimensions, 16 3×3 convolutional kernels are selected. 3×3 convolutional kernels can retain local features while reducing parameters. A stride of 1 is used to avoid feature loss, and the padding method is "same" to ensure that the output dimension is consistent with the input. The activation function is ReLU, which can effectively alleviate the gradient vanishing problem and improve the model training efficiency. The second convolutional block increases the number of convolutional kernels to 32 based on the first convolutional block to extract more complex acoustic features. Features; the third convolutional block further increases the number of convolutional kernels to 64, realizing deep extraction of high-level features; the pooling layer uses 2×2 max pooling with a stride of 2, which can compress the data dimension while retaining key feature information and reducing the computational cost of the model; the deconvolution layer uses 64 3×3 deconvolutional kernels with a stride of 2 and the padding method is same, restoring the pooled feature map to the input dimension, ensuring that the output dimension is consistent with the input dimension; the output layer uses a 1×1 convolutional kernel and the activation function is sigmoid, which can map the output value to the 0-1 interval, corresponding to the discrimination result of noise and effective signal, and the output is a denoised time-frequency feature matrix with the same dimension as the input. The training data consisted of historical acoustic profiles from multiple segments of intelligent detection and survey data for subsea pipelines. The data volume was determined based on the convergence requirements of the model training. Experiments verified that when the data volume reached 10,000 segments, the model could fully learn the distribution patterns of noise and effective signals. Each data segment was divided into training and validation sets in an 8:2 ratio using the corresponding time-frequency domain acoustic feature matrix and noise energy mask. This ratio, determined according to the conventional principles of training and validation set division in machine learning, ensures sufficient training data while effectively monitoring the model's generalization ability. The training set was used for model parameter updates, and the validation set was used to monitor the model training effect. The training process employed the Adam optimizer, which combines momentum gradient descent and adaptive learning rate adjustment to accelerate model convergence. The learning rate was derived based on the model's convergence performance; experiments determined that 0.001 was the optimal learning rate, ensuring both convergence speed and avoiding overfitting. The batch size was set based on the memory capacity of the training device. Considering the memory limitations of embedded devices, a batch size of 32 was determined. The cross-entropy loss function was used, with the formula: [Formula omitted for brevity]. ,in The true label for the noise energy mask image. The predicted labels output by the model are used. The number of training iterations is determined based on the model's convergence. Training stops when the validation set loss no longer decreases for 10 consecutive epochs. This stopping condition can effectively avoid model overfitting and generate a well-trained intelligent denoising model. This model can accurately filter out noise in the acoustic signals of submarine pipeline detection by learning the differences in amplitude and frequency distribution between noise and effective signals.

[0017] The real-time acquired acoustic profile data is input into the trained intelligent denoising model. The model outputs the clean time-frequency features after filtering out continuous background noise, generating a denoised time-frequency spectrum. An inverse short-time Fourier transform is performed on the denoised time-frequency spectrum to reconstruct a time-domain acoustic signal with suppressed reverberation noise, generating preliminary denoised acoustic profile data. In this embodiment of the invention, during the intelligent detection and survey of subsea pipelines, acoustic profile data is collected in real time. The collected parameters are consistent with historical data to ensure data consistency and comparability. The sampling frequency and data duration are consistent with the derived parameters, resulting in a one-dimensional real-time time-domain signal. This real-time signal is then subjected to the same transformation operation according to the short-time Fourier transform parameters to obtain a real-time time-frequency domain acoustic feature matrix, which is directly input into a trained intelligent denoising model. The model, based on the learned distribution patterns of noise and effective signals, discriminates each frequency point in the real-time feature matrix, filters out the amplitude information corresponding to continuous background noise, and outputs a clean time-frequency feature matrix after noise removal. An inverse short-time Fourier transform is performed on this clean time-frequency feature matrix, using the same window function, window length, and overlap rate as the forward transform to ensure the accuracy of time-domain signal reconstruction. Complete the time-domain signal reconstruction, where The reconstructed time-domain acoustic signal, The time-frequency domain signal after denoising is reconstructed to obtain a time-domain acoustic signal with suppressed reverberation noise. This signal is the acoustic profile data for preliminary denoising, which can clearly present the effective acoustic characteristics related to subsea pipeline detection, eliminate the interference of seawater reverberation and random noise, and improve the accuracy of pipeline detection.

[0018] To address the residual multipath reflection interference in the acoustic profile data after initial noise reduction, a matched filtering technique based on a prior model of pipe reflection is employed. A filter matching the typical reflection characteristics of the pipe is designed, and convolution operations are performed on the signal to further suppress reflection interference unrelated to the prior model, thereby generating an enhanced acoustic profile image sequence.

[0019] In this embodiment of the invention, a priori model of pipeline reflection is constructed to address the residual multipath reflection interference in the acoustic profile data after preliminary noise reduction. This model is based on the actual physical parameters of the subsea pipeline. The pipeline diameter and material are actual parameters that have been determined before the subsea pipeline exploration and investigation. The propagation speed of sound waves in the pipeline is derived from the pipeline material, the temperature and salinity of the surrounding seawater medium. The propagation speed of sound waves in the seawater medium is positively correlated with temperature and salinity. Combined with the seawater environmental parameters of the on-site exploration area, the propagation speed of sound waves is calculated. Then, combined with the laying depth of the subsea pipeline and the acoustic characteristics of the surrounding medium (seawater, seabed sediments), the propagation path and amplitude attenuation law of the multipath reflection signal are determined through acoustic propagation theory. The delay threshold of the reflection signal is calculated based on the pipeline diameter and the propagation speed of sound waves, i.e., delay threshold = 2 × pipeline diameter / propagation speed of sound waves. The amplitude attenuation coefficient is derived from the sound absorption coefficient and propagation distance of the seabed sediments, and the specific value is determined in combination with the sediment characteristics of the on-site exploration area. Based on this prior model, a matched filter is designed. The impulse response of the filter perfectly matches the characteristic waveform of the pipeline reflection signal, which is derived from the pipeline reflection prior model. The acoustic signal after preliminary noise reduction is input into the filter, and the effective signal reflected by the pipeline is enhanced through convolution operation. At the same time, the multipath reflection interference signal that does not match the prior model is suppressed, and the acoustic profile data with thorough noise reduction and clear effective signal is output, providing high-quality acoustic data support for the detection and investigation of submarine pipeline location, damage detection and other related work.

[0020] Furthermore, the intelligent identification of pipeline targets on the enhanced acoustic profile image sequence includes the following steps: Adaptive grayscale normalization and contrast stretching are performed on the enhanced acoustic profile image sequence to unify the brightness and contrast levels of images under different voyages and sea conditions, thereby generating a standardized acoustic image sequence. In this embodiment of the invention, an enhanced acoustic profile image sequence is obtained, consisting of multiple frames of acoustic profile images. Each frame corresponds to a seabed area at a fixed detection distance, and the grayscale value range varies with the detection voyage and sea state. Adaptive grayscale normalization processing is performed on the image sequence, and the maximum grayscale value of each frame is calculated. Minimum value The image grayscale values ​​are mapped to the 0-255 range using a linear normalization formula, which is: ,in The original grayscale value. This formula, used to normalize grayscale values, eliminates grayscale fluctuations caused by differences in illumination across different voyages and sea states, achieving grayscale uniformity. After normalization, a contrast stretching operation is performed, employing histogram equalization to statistically analyze the grayscale histogram of each frame and calculate the cumulative grayscale distribution function. ,in grayscale The probability density, through the mapping relationship This method stretches areas with concentrated grayscale distribution to the full grayscale range, enhancing the grayscale difference between the pipeline target and the background and seabed interface. Ultimately, it generates a standardized acoustic image sequence, ensuring that subsequent segmentation and recognition processes are not affected by differences in image brightness and contrast, thus laying the foundation for accurate pipeline target extraction.

[0021] An improved U-Net network architecture is used to perform pixel-level semantic segmentation on standardized acoustic image sequences. The network encoder extracts multi-scale contextual features, the decoder recovers spatial details, and an attention gate mechanism is introduced in the skip connections to focus on the pipe target region, generating a preliminary segmentation map containing three categories of labels: "background", "pipe", and "seabed interface". In this embodiment of the invention, a modified U-Net network architecture is used to perform pixel-level semantic segmentation on standardized acoustic image sequences. The overall network structure consists of three parts: an encoder, a decoder, and skip connections. The total number of model parameters is controlled within 8 million, adapting to the lightweight computing power requirements of subsea pipeline detection equipment. The encoder part contains four convolutional blocks, each consisting of two 3×3 convolutional layers, one batch normalization layer, and one ReLU activation function. The stride of each convolution is 1, and the padding method is "same". The first convolutional block outputs a feature map with 64 channels, and the number of channels doubles in each subsequent convolutional block, to 128, 256, and 512 respectively. Each convolutional block is followed by a 2×2 max pooling layer (stride of 2) to compress the feature map dimension and extract multi-scale contextual features, achieving a gradual extraction from shallow texture to deep semantic features. The decoder section is symmetrical to the encoder and contains four deconvolutional blocks. Each deconvolutional block consists of one 2×2 deconvolutional layer (stride of 2), two 3×3 convolutional layers, one batch normalization layer, and one ReLU activation function. The deconvolutional layers are used to recover spatial details of the feature maps. The number of channels gradually decreases from 512 to 64, with a final output of 3 channels, corresponding to the three labels: "background," "pipe," and "seabed interface." An attention gate mechanism is introduced in the skip connections. The attention gate consists of a 1×1 convolutional layer and a sigmoid activation function. It calculates the correlation weights between the encoder and decoder feature maps using the following formula: ,in For attention weights, For the sigmoid function, , This is the weight matrix. For bias terms, For encoder feature maps, The decoder feature map is generated by multiplying the weights with the encoder feature map and then feeding it into the decoder to focus on the target region in the pipeline and suppress background interference. The training data consists of 10,000 frames of standardized acoustic images and corresponding manually labeled images, divided into training and validation sets in a 7:3 ratio. Training uses the Adam optimizer, with the learning rate determined through model convergence experiments. The loss function is the cross-entropy loss function. Iterative training continues until the validation set loss stabilizes, generating a trained segmentation model. Inputting a sequence of standardized acoustic images into this model outputs a preliminary segmentation map containing three types of labels, achieving initial localization of the pipeline target.

[0022] Connectivity analysis is performed on the pixel regions marked as "pipes" in the preliminary segmentation map to remove noise regions whose pixel area is smaller than the lower limit of the physical size of the pipe, generating a set of candidate pipe connected regions. For each candidate pipe connected region, the aspect ratio of its minimum bounding rectangle, the average reflection intensity of the pixels in the region, and the skewness and kurtosis of the reflection intensity distribution are calculated to generate a set of geometric and reflection feature descriptors. In this embodiment of the invention, connected component analysis is performed on the pixel regions marked as "pipes" in the preliminary segmentation map. The eight-neighbor connectivity criterion is used, meaning that "pipe" pixels within a range of eight adjacent pixels are considered to belong to the same connected component. The entire segmentation map is traversed, and all independent "pipe" connected components are marked. Based on the actual physical dimensions of the seabed pipeline and the resolution and detection distance of the detection equipment, the minimum pixel area threshold for the pipeline in the image is derived. This threshold is calculated from the actual diameter of the pipeline and the detection scale, i.e., minimum pixel area = π × (actual pipeline diameter × detection scale / 2). 2 Connected components with pixel areas smaller than a threshold are removed, as these are formed by noise interference. This process generates a set of candidate pipeline connected components. For each candidate pipeline connected component, its minimum bounding rectangle is calculated using image geometric analysis. The length and width of the rectangle are determined by traversing the extreme coordinates of all pixels within the connected component, i.e., length = ... Width = Then, the aspect ratio is calculated as length / width, which can distinguish between pipe-like and non-pipe-like elongated interference; the average reflection intensity of all pixels within the connected component is calculated using the formula: ,in Let be the reflection intensity of pixel (x, y). The total number of pixels within the connected region; calculate the skewness and kurtosis of the reflection intensity distribution, the skewness formula is: The kurtosis formula is ,in The aspect ratio, average reflection intensity, skewness, and kurtosis together constitute the geometric and reflection feature descriptors, which fully characterize the feature information of the candidate pipeline.

[0023] The geometric and reflection feature descriptors are input into a pre-trained pipeline target classifier, which is built based on support vector machines to distinguish the reflection of real pipelines from the reflection of similar-shaped geological bodies or debris. The classifier outputs the probability value of each candidate connected component as a real pipeline, generating a pipeline target probability set. A probability threshold is set, and candidate connected components with probability values ​​exceeding the threshold are filtered out. The geometric center coordinates of their pixel clusters are calculated as the target location. The equivalent diameter is calculated based on the circumscribed rectangle, and the intensity distribution histogram of pixels in the region is statistically analyzed to generate a preliminary set of suspected pipeline targets.

[0024] In this embodiment of the invention, a pipeline target classifier based on support vector machines is constructed. This classifier employs a linear kernel function, the expression of which is: This approach reduces computational complexity while maintaining classification accuracy, making it suitable for lightweight detection needs. The classifier training data uses generated feature descriptors and corresponding labels, categorized into "real pipes" and "interference targets." A training set of 5000 feature descriptors is selected, with 3000 representing real pipe features and 2000 representing interference targets such as geological bodies and debris. The optimal classification hyperplane is determined by solving a convex quadratic programming problem during training. The hyperplane equation is... ,in For the weight vector, As a bias term, by minimizing Achieve hyperplane optimization The penalty coefficient is determined by the training experiment. The slack variables are used to handle outlier samples, generating a pre-trained pipeline target classifier after training. The generated geometric and reflection feature descriptors are input into this classifier. The classifier calculates the distance between the feature descriptors and the hyperplane, outputting the probability value of each candidate connected component being a real pipeline, generating a pipeline target probability set. The probability threshold is determined experimentally based on the accuracy and recall of the classifier training to ensure that all selected candidate connected components are real pipelines. Candidate connected components with probability values ​​exceeding the threshold are then selected. All pixel coordinates within the connected component are traversed, and the geometric center coordinates are calculated. Use these coordinates as the target location of the pipeline; calculate the equivalent diameter based on the length and width of the minimum bounding rectangle. The equivalent diameter has a fixed proportional relationship with the actual diameter of the pipe, which can reflect the actual size of the pipe. The distribution histogram of pixel reflection intensity in the connected domain is statistically analyzed. The horizontal axis is the reflection intensity interval and the vertical axis is the number of pixels in the corresponding interval. The reflection intensity characteristics of the pipe are fully recorded, and a preliminary set of suspected pipe targets is generated, which provides data support for the subsequent accurate positioning and detection of the pipe.

[0025] Furthermore, the pixel-level semantic segmentation of the standardized acoustic image sequence using the improved U-Net network architecture includes the following steps: Construct a U-Net network backbone with encoder-decoder symmetry, where the encoder contains four downsampling stages, each of which uses two convolutional layers and one max pooling layer to progressively extract and compress image features, generating multi-level downsampled feature maps; In this embodiment of the invention, a U-Net network backbone with encoder-decoder symmetry is constructed to meet the lightweight computing power requirements of intelligent detection and surveying of subsea pipelines. The total number of network parameters is controlled within 8 million, ensuring fast inference on the embedded processor of the detection equipment. The encoder contains four downsampling stages, which are progressively advanced to achieve gradual extraction and compression of image features. Each downsampling stage consists of two 3×3 convolutional layers and one 2×2 max pooling layer. The stride of each convolutional layer is set to 1, and the padding method is set to "same" to ensure that the size of the feature map after convolution is consistent with the input, avoiding loss of feature information. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The batch normalization layer is used to accelerate model training and alleviate gradient vanishing, while the ReLU activation function is used to introduce nonlinear features and enhance the network's feature extraction capability. The first downsampling stage takes a standardized acoustic image as input. After processing through two convolutional layers, it outputs a feature map with 64 channels. Then, a max pooling layer (with a stride of 2) compresses the feature map size to half that of the input, generating the first-level downsampled feature map. The second downsampling stage receives the first-level downsampled feature map, convolves it, and doubles the number of output channels to 128. After compression by a max pooling layer, it generates the second-level downsampled feature map. The third downsampling stage outputs 256 channels through convolution and generates the third-level downsampled feature map after max pooling. The fourth downsampling stage outputs 512 channels through convolution and generates the fourth-level downsampled feature map after max pooling. This feature map is the semantic feature map output by the encoder, containing deep feature information of the pipeline target, realizing the gradual extraction and compression from shallow texture to deep semantic features.

[0026] An attention gate module is introduced at the skip connection of the network. This module takes the high-level semantic feature map passed from the encoder side as input and generates a spatial attention weight map. This map is used to weight the encoder features before feature fusion, so that the network pays more attention to the pipeline target region related to the current task of the decoder and generates weighted encoder features. In this embodiment of the invention, an attention gate module is introduced at the network skip connections. This module corresponds one-to-one with the four downsampling stages of the encoder, processing the downsampling feature maps at each level to achieve precise focusing of the target region in the pipeline. The attention gate module takes the corresponding high-level semantic feature map passed from the encoder as its sole input. Internally, the module consists of two 1×1 convolutional layers, an addition unit, and a sigmoid activation function. The first 1×1 convolutional layer compresses the number of channels in the encoder feature map to 1 / 4 of the original number, reducing computational complexity. The second 1×1 convolutional layer restores the number of channels in the compressed feature map to the original number, ensuring dimensionality matching with the subsequent decoder feature map. (The formula is used to...) ,in For attention weights, For the sigmoid function, , This is the weight matrix. For bias terms, For encoder feature maps, For the decoder feature map, the sigmoid activation function maps weight values ​​to the 0-1 range, with higher weight values ​​indicating a greater probability that the corresponding region is a pipe target. The generated spatial attention weight map is then multiplied pixel-by-pixel with the original encoder feature map to obtain weighted encoder features. This weighting operation suppresses background region features and enhances pipe target region features, enabling the network to focus more on pipe target regions relevant to the decoder's current upsampling task during subsequent feature fusion, thus reducing the impact of seabed sediments and water interference on segmentation accuracy.

[0027] The decoder consists of four upsampling stages. Each stage receives weighted encoder features from the corresponding layer, upsamples them through transposed convolutions, concatenates them with the weighted encoder features, and then performs feature fusion and refinement through two convolutional layers to gradually restore spatial resolution and generate the final high-resolution feature map. In the final layer of the network, a 1x1 convolution is used to map the number of channels of the high-resolution feature map to the number of categories, and the Softmax function outputs the probability of each pixel belonging to "background", "pipe", and "seabed interface" to generate a preliminary segmentation map. In this embodiment of the invention, the decoder comprises four upsampling stages, symmetrically distributed with the encoder's four downsampling stages, to achieve gradual restoration of the feature map's spatial resolution. Each upsampling stage receives weighted encoder features from the corresponding level of the encoder, and simultaneously receives the feature map output from the previous upsampling stage. First, a 2×2 transposed convolutional layer is used to upsample the received feature map from the previous stage, with a stride of 2, doubling the feature map size to ensure the upsampled feature map size is identical to the corresponding level's weighted encoder feature map size. The upsampled feature map and the corresponding level's weighted encoder feature map are then concatenated, with the number of channels in the concatenated feature map being the sum of the number of channels in both layers, achieving the fusion of shallow texture features and deep semantic features. After concatenation, two 3×3 convolutional layers refine the fused feature map, with a stride of 1 and a "same" padding method. A batch normalization layer and a ReLU activation function are then connected after the convolution to further extract pipeline target information from the fused features and suppress invalid interference features. The four upsampling stages proceed sequentially: the first stage outputs 256 channels, the second 128, the third 64, and the fourth 32, gradually restoring the spatial resolution of the feature map. The final output is a high-resolution feature map with dimensions identical to the input normalized acoustic image. In the final layer of the network, a 1×1 convolutional layer maps the high-resolution feature map to 3 channels, corresponding to the three labels: "background," "pipe," and "seabed interface." The Softmax function is then used to calculate the probability of each pixel belonging to one of the three labels. The Softmax function formula is... ,in For the first The probability of class labels, The output of a 1×1 convolutional layer The class label feature values ​​are used to generate a preliminary segmentation map in which each pixel corresponds to a class probability, thus achieving a preliminary distinction between the pipeline target and the seabed interface and background.

[0028] During the model training phase, an acoustic image dataset containing labeled pipe contours is used. The weighted sum of the cross-entropy loss function and the Dice loss function is used as the optimization objective. The network parameters are adjusted through the backpropagation algorithm until the segmentation accuracy of the model on the validation set reaches the preset standard, thus generating a trained improved U-Net segmentation model.

[0029] In this embodiment of the invention, during the model training phase, an acoustic image dataset containing labeled pipeline outlines is used. The dataset consists of acoustic images collected during actual operations of intelligent subsea pipeline detection and survey. The images cover scenes from different voyages, sea states, and pipeline operating conditions, ensuring the diversity and representativeness of the dataset. The dataset contains 10,000 frames, each with manually labeled "background," "pipeline," and "seabed interface," ensuring the labeling accuracy matches actual detection needs and accurately reflects pipeline outlines and boundary information. The dataset is divided into a training set and a validation set in a 7:3 ratio. The training set is used for model parameter adjustment, while the validation set is used to monitor model segmentation accuracy, ensuring the model has good generalization ability. A weighted sum of the cross-entropy loss function and the Dice loss function is used as the model optimization objective, with weighting coefficients set to 0.5 to balance class balance and segmentation boundary accuracy. The cross-entropy loss function formula is... ,in One-hot encoding for the tag. The class probabilities are the output of the model; the Dice loss function formula is: This is used to address training bias caused by the low proportion of target pixels in the pipeline. The total loss function formula is: Training employs the Adam optimizer, with the learning rate determined through model convergence experiments. The batch size is set based on the memory capacity of the detection device. During training, the gradient of the total loss function is propagated backward to each layer of the network via backpropagation, gradually adjusting the weight matrices and bias terms of the convolutional layers and attention gate modules. The segmentation accuracy of the validation set is monitored in real time during iterative training. The preset segmentation accuracy standard is a validation set intersection-union ratio greater than 0.85. When the validation set segmentation accuracy remains stable above the preset standard for 10 consecutive training epochs, model training is stopped, generating a trained improved U-Net segmentation model. This model can accurately achieve pixel-level segmentation of pipeline targets, laying the foundation for subsequent pipeline target selection.

[0030] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1 A detailed flowchart illustrating the steps for generating a pipeline target point cloud as described in S03 is shown. In this embodiment, generating the pipeline target point cloud includes the following steps: S031: Obtain the latitude, longitude, elevation, and timestamp of the probe's center of mass in the geodetic coordinate system from the differential GNSS positioning system to generate the absolute attitude time series of the hull; obtain the roll angle, pitch angle, bow angle, and three-dimensional acceleration data in the hull coordinate system with the center of mass as the origin from the attitude inertial navigation system to generate the relative motion attitude time series of the hull; measure and calculate the three-dimensional installation offset of the transducer phase center relative to the center of mass of the hull based on the installation position of the shallow seismic profiler transducer on the hull, and generate the instrument installation parameter matrix; In this embodiment of the invention, during the intelligent detection and survey operation of the subsea pipeline, the differential GNSS positioning system collects the position and time information of the probe's center of mass in the geodetic coordinate system in real time. The sampling frequency is consistent with the acoustic data acquisition frequency to ensure time synchronization between the positioning data and the acoustic data. The acquisition parameters are fixed to output a set of data every 100ms, generating a time series of the probe's absolute pose. This series contains the latitude and longitude in the geodetic coordinate system at each acquisition moment. Elevation and corresponding timestamp The latitude and longitude coordinates are in degrees, minutes, and seconds format, the elevation is based on mean sea level, and the timestamps use the UTC unified time format to accurately correspond to the acquisition time of each set of acoustic data. The attitude inertial navigation system and the differential GNSS positioning system work synchronously to acquire motion attitude data in the ship's coordinate system with the ship's center of mass as the origin. Similarly, a set of data is output every 100ms to generate a time series of the ship's relative motion attitude, which includes the roll angle. Pitch angle Bow angle and three-dimensional acceleration data The roll angle is the rotation angle of the hull about the longitudinal axis, the pitch angle is the rotation angle of the hull about the transverse axis, and the bow angle is the angle between the bow of the hull and true north. The three-dimensional accelerations correspond to the accelerations in the longitudinal, transverse, and vertical directions in the hull coordinate system, respectively, and are used to correct the influence of hull motion on target positioning. The installation position of the transducer on the hull is determined by precise measurement. A hull coordinate system is established with the hull's center of mass as the origin (x-axis forward along the longitudinal direction of the hull, y-axis to the right along the transverse direction of the hull, and z-axis perpendicular to the hull and upward). The offsets of the transducer's phase center in the x, y, and z directions in this coordinate system are measured, and these three offsets are organized into a 3×1 instrument installation parameter matrix. ,in This is the vertical offset. This is the lateral offset. This is the vertical offset, and this matrix is ​​used for position correction of the transducer and the center of mass of the hull in subsequent coordinate transformations.

[0031] S032: At each acoustic data acquisition moment, synchronize the absolute pose time series of the hull, the relative motion attitude time series of the hull, and the image coordinates of the target points in the preliminary pipeline suspected target set according to the timestamp; In this embodiment of the invention, a time synchronization mechanism is established based on the acoustic data acquisition time to ensure precise synchronization of the ship's absolute pose time series, the ship's relative motion attitude time series, and the image coordinates of target points in the initial pipeline suspected target set. The acoustic data acquisition time is recorded by the built-in clock of the shallow subsurface profiler, outputting a UTC timestamp consistent with the differential GNSS positioning system and the attitude inertial navigation system. Each acoustic data acquisition time corresponds to a unique timestamp. Regarding this timestamp Searching for and matching the absolute pose time series of the hull The closest pose data is used to correct for time deviations using linear interpolation. The absolute position of the hull at that moment Perform the same operation on the time series of the ship's relative motion attitude to retrieve and correct the results. The relative motion attitude of the ship at that moment Simultaneously, extract the generated preliminary pipeline suspected target set, corresponding to Image coordinates of all target points in the acoustic image at any given time ,in The horizontal pixel coordinates of the image. The vertical pixel coordinates of the image ensure a one-to-one correspondence between the absolute pose, relative pose, and target point image coordinates at the same acquisition time, eliminating positioning errors caused by time deviations, providing a synchronous data foundation for subsequent coordinate transformations, and ensuring the accuracy of pipeline target positioning.

[0032] S033: Based on the instrument installation parameter matrix and the synchronized time series of the ship's relative motion attitude, the target point is transformed from the instrument coordinate system centered on the transducer to the ship coordinate system centered on the ship's center of mass through three-dimensional coordinate rotation and translation transformation, thereby generating the target point coordinates in the ship coordinate system. In this embodiment of the invention, the method is based on the instrument installation parameter matrix. By combining the synchronized time series of the ship's relative motion attitude, the target point is transformed from the instrument coordinate system to the ship coordinate system through three-dimensional coordinate rotation and translation transformation. First, an instrument coordinate system is established with the transducer phase center as the origin. The initial attitude of the instrument coordinate system is consistent with that of the ship coordinate system. The target point's coordinates in the instrument coordinate system are... From its image coordinates Based on the imaging parameters derived from the shallow seismic profiling instrument, these parameters are fixed by the instrument's hardware and calculated through the mapping relationship between image pixels and actual distances. Due to the ship's roll, pitch, and bow movements, attitude deviations need to be corrected using a rotation matrix. This three-dimensional rotation matrix is ​​derived from the roll angle... Pitch angle Bow angle The construction and rotation sequence is: first roll, then pitch, and finally bow; rotation matrix. The expression is: ,in For the roll rotation matrix, For the pitch and rotation matrix, The forward rotation matrix is ​​expressed as follows: , , The coordinate transformation process first involves rotating and correcting the target point coordinates in the instrument coordinate system to obtain the rotated coordinates. In addition, the instrument installation parameter matrix After performing a translation transformation, the coordinates of the target point in the ship's coordinate system are finally obtained. This process transforms the target point from the instrument coordinate system to the ship coordinate system, eliminating the influence of ship attitude motion and transducer installation offset on the target coordinates.

[0033] S034: Combining the synchronized ship absolute pose time series, the target point coordinates in the ship coordinate system are transformed to the CGCS2000 geodetic coordinate system through geodetic theme calculation, generating pipeline target point coordinates with absolute latitude, longitude and elevation, and generating a pipeline target point aggregator after integrating all time data.

[0034] In this embodiment of the invention, by combining the synchronized absolute pose time series of the hull and using geodetic calculation, the target point coordinates in the hull coordinate system are transformed to the CGCS2000 geodetic coordinate system, thus achieving the absolute positioning of the pipeline target. (Synchronized absolute pose of the hull) The absolute position of the ship's center of mass in the CGCS2000 geodetic coordinate system, and the coordinates of the target point in the ship's coordinate system. To determine the relative position of the target point with respect to the ship's center of mass, the relative position needs to be converted to an absolute position using geodetic calculations. First, the geodetic coordinates of the ship's center of mass are... Convert to spatial rectangular coordinate system The conversion formula is based on the CGCS2000 ellipsoid parameters, and the semi-major axis of the ellipsoid. Flatness For a fixed value, the conversion formula is: , , ,in Let be the radius of curvature of the ellipsoid's prime meridian. , For the eccentricity of the ellipsoid, , It is the semi-major axis of the ellipse. This is the minor semi-axis of the ellipsoid. Then, the target point in the ship's coordinate system is relative to the coordinates... By rotation matrix Convert to relative coordinates in a Cartesian coordinate system The rotation matrix of the hull coordinate system relative to the spatial rectangular coordinate system is given by the roll angle. Pitch angle Bow angle The construction follows the same logic as the derivation of the rotation matrix. The transformed relative coordinates are added to the Cartesian coordinates of the ship's center of mass to obtain the Cartesian coordinates of the target point. Then convert the spatial rectangular coordinates Reverse conversion to latitude and longitude in CGCS2000 geodetic coordinate system With elevation Complete the absolute coordinate transformation. Repeat the above operation to transform the coordinates of the pipeline target points at all acquisition times, integrate the absolute coordinate data of all target points, and generate a pipeline target point cloud. This point cloud can accurately reflect the actual spatial location of the subsea pipeline, providing accurate positioning data support for the location confirmation and route analysis of the subsea pipeline.

[0035] Furthermore, the pipeline route correlation analysis based on the design route planning data and the pipeline target point cloud includes the following steps: The design routes of each subsea pipeline are extracted from the design route planning data. The design routes are represented by a three-dimensional spatial polyline formed by connecting a series of ordered design inflection point coordinates, and a three-dimensional line set of design routes is generated. In this embodiment of the invention, during the intelligent detection and survey of subsea pipelines, the route planning data is the preset design data before pipeline laying, containing detailed route information for all subsea pipelines to be detected. The designed route of each pipeline is presented in the form of a three-dimensional spatial polyline, which is composed of a series of ordered design inflection point coordinates. The inflection point coordinates all adopt the CGCS2000 geodetic coordinate system, including latitude and longitude. With elevation In this example, it is applied to a real-world scenario of a pipeline post-investigation project. For instance, for the seven corresponding subsea pipelines (AE mixed transport, EA water injection, BA mixed transport, AB water injection, CA mixed transport, AC water injection, and DE mixed transport), the design routes of each of the seven subsea pipelines are extracted from the design route planning data of the pipeline post-investigation project. Each design route is represented by a three-dimensional spatial polyline, which is formed by connecting a series of ordered design inflection point coordinates. The inflection point coordinates adopt the CGCS2000 geodetic coordinate system, UTM projection (scale factor 0.9996), and the coordinate unit is m, including three dimensions: north coordinate (X), east coordinate (Y), and elevation (Z).The design route of the AE mixed-transport pipeline extracts 12 design inflection points, with coordinates as follows: (4251301.58, 415764.42, -12.3), (4251500.21, 416000.15, -12.5), (4251720.35, 416230.89, -12.7), (4251950.12, 416450.76, -12.8), (4252180.45, 416680.33, -12.9), (4252410.67, 416910.88, -13.0), (4252640.23, 41714). 0.55, -13.1), (4252870.56, 417370.12, -13.2), (4253100.34, 417600.67, -13.3), (4253330.78, 417830.24, -13.4), (4253400.11, 418050.89, -13.5), (4253432.55, 418242.33, -13.6); 11 design inflection points were extracted from the EA water injection pipeline design route, and the coordinates of the inflection points are (4253398.76, 418233.71, -13.6), (4253168 ... .43, 418003.16, -13.5), (4252938.11, 417772.61, -13.4), (4252707.78, 417542.06, -13.3), (4252477.45, 417311.51, -13.2), (4252247.12, 417080.96, -13.1), (4252016.79, 416850.41, -13.0), (4251786.46, 416619.86, -12.9), (4251556.13, 416389.31, -12. 8), (4251325.80, 416158.76, -12.7), (4251291.62, 415783.51, -12.5); the remaining 5 pipeline design routes all extract the corresponding ordered design inflection point coordinates according to the same rules. After all the design inflection point coordinates are sorted in order according to the pipeline extension direction, adjacent inflection point coordinates are connected pairwise to form design route segments. The design route segments of the 7 pipelines are summarized to form a design route 3D line set. Each design route segment is assigned a unique ID, and the ID encoding rule is "pipeline number-segment number". For example, the ID of the first design route segment of the AE mixed transmission pipeline is "AE-01", and so on. Among them, the design route 3D line sets corresponding to some segments are shown in Table 1 below: Table 1. Design Route 3D Line Set (Partial Line Segments)

[0036] For each target point in the pipeline target point cloud, traverse each design route segment in the design route three-dimensional line set, calculate the three-dimensional Euclidean vertical distance from the target point to each line segment, and record the minimum vertical distance and its corresponding design route line segment ID to generate a point-line minimum distance association table. In this embodiment of the invention, a pipeline target point set is obtained. This point set contains all the detection target points of the 7 pipelines, and each target point contains the north coordinates in the CGCS2000 geodetic coordinate system. East coordinates Elevation Three-dimensional parameters, with coordinates in meters (m). For each target point in the point cloud, traverse all design route segments in the three-dimensional line set of the design routes, calculating the three-dimensional Euclidean vertical distance from the target point to each design route segment. The three-dimensional Euclidean vertical distance calculation uses the formula for the vertical distance from a spatial point to a spatial line segment, assuming the two endpoints of the spatial line segment are... and The target point is First, calculate the vector. ,vector Calculate vector In vector Projection coefficients on , ,when When ≤0, the perpendicular distance from the target point to the line segment is ,Right now ;when When ≥1, the perpendicular distance from the target point to the line segment is ,Right now When 0 < When <1, the perpendicular distance from the target point to the line segment is ,Right now After calculation, record the three-dimensional Euclidean vertical distances from the target point to all design route segments. Filter out the smallest vertical distance value and record the corresponding design route segment ID. Enter the target point number, minimum vertical distance, and corresponding design route segment ID into a table to generate a point-to-line minimum distance association table. This table contains six items: target point number, north coordinate, east coordinate, elevation, minimum vertical distance, and corresponding segment ID. The target point numbers are sequentially encoded as "D001, D002, D003…" according to the detection order, ensuring that each target point has a unique corresponding record without omissions or duplicates. Partial data from the generated point-to-line minimum distance association table is shown in Table 2 below. Table 2. Point-to-Line Minimum Distance Relationship Table (Partial Data)

[0037] Set a distance association threshold, filter out all target points whose minimum vertical distance in the point-line minimum distance association table is less than the threshold, and initially classify these target points as "points related to the design route" to generate a preliminary route-related point set; for each design route, extract all target points associated with it from the preliminary route-related point set, and spatially sort them based on the timestamps of these target points or the projection distance along the design route to generate a sequence of target points arranged in an orderly manner according to the pipeline extension direction; In this embodiment of the invention, the distance association threshold is derived based on the accuracy requirements of subsea pipeline detection and the allowable deviation of the design route. Combined with the construction error of the subsea pipeline and the positioning error of the detection equipment, and verified through field tests, the distance association threshold is determined to be a fixed value. This threshold ensures that only target points that actually overlap with the design route are selected. Based on this distance association threshold, the generated point-to-line minimum distance association table is filtered, removing target points with a minimum vertical distance greater than or equal to the threshold. These target points are considered interference points or non-target pipeline points. All target points with a minimum vertical distance less than the threshold are retained and categorized as "points related to the design route," generating a preliminary route-related point set. For each design route in the three-dimensional line set, based on its route ID and line segment ID, all target points associated with it are extracted from the preliminary route-related point set; that is, all target points whose minimum distance corresponds to the line segment ID of that route. Based on the acquisition timestamps of these target points, the target points are spatially sorted in chronological order. Target points with earlier timestamps correspond to earlier pipeline locations traversed by the probe. The sorted sequence yields a target point sequence consistent with the actual extension direction of the pipeline. For target points with abnormal timestamps, the sorting is based on the projected distance along the designed route segment. The projected distance is calculated using vector projection. For example, the projected distance... The target points are sorted by projection distance from smallest to largest, and finally a sequence of target points corresponding to each design route is generated, arranged in an orderly manner according to the pipeline extension direction.

[0038] A three-dimensional B-spline curve is fitted to the ordered sequence of target points to smooth out coordinate fluctuations caused by measurement errors or ocean current disturbances, generating a smooth and continuous three-dimensional spatial curve as the continuous three-dimensional spatial trajectory of the subsea pipeline. The elevation value corresponding to each target point and the depth difference from the top of the pipeline to the seabed interface interpreted from the acoustic image are extracted from the ordered sequence of target points to calculate the burial depth of the pipeline and generate the burial depth attribute dataset of the subsea pipeline.

[0039] In this embodiment of the invention, a three-dimensional B-spline curve is fitted to the ordered sequence of target points for each pipeline. The three-dimensional B-spline curve fitting adopts... B-spline basis function construction, basis function degree The curve is fitted to three degrees (a cubic B-spline curve can achieve smoothness, continuity, and a close fit to the target point distribution). The fitting process is based on the three-dimensional coordinates of the target point sequence. Construct the equation of the three-dimensional B-spline curve: ( ),in The coordinates of the i-th target point in an ordered sequence of target points , The number of points in the target point sequence. for The B-spline basis functions are calculated using the de Boer recurrence formula: when When =0, ( ),otherwise ;when When >0, ,in The parameter has a range of values. Node vectors Set according to uniform distribution, ( This formula is used to fit an ordered sequence of target points, generating a smooth and continuous three-dimensional spatial curve. This curve represents the continuous three-dimensional spatial trajectory of the corresponding subsea pipeline, accurately reflecting the pipeline's actual route and spatial distribution. Simultaneously, the elevation value of each target point is extracted from the ordered sequence. The depth difference between the top of the pipe and the seabed interface at the target point, obtained by interpreting acoustic detection images. Pipeline burial depth According to the formula Calculation, where The seabed interface elevation corresponding to the target point is obtained (by combining water depth measurement data with tidal level correction). The corresponding burial depth value is calculated for each target point. The number of each target point, the corresponding pipeline number, the elevation value, the depth difference, and the burial depth are entered into a table in sequence to generate the burial depth attribute dataset of the subsea pipeline. The corresponding burial depth attribute datasets are generated for each of the 7 pipelines, and the burial depth information of each pipeline at different locations is completely recorded.

[0040] Furthermore, the three-dimensional B-spline curve fitting of the ordered sequence of target points includes the following steps: Input an ordered sequence of target points, where each point contains three-dimensional coordinates. Parameterize the sequence and calculate the parameter value of each data point in the cumulative chord length parameter space to generate a parameterized point sequence. Set the degree of the B-spline curve, the number of control points, and the node vector. Approximate the parameterized point sequence using the least squares method. Calculate a set of coordinates of the control vertices by solving the system of equations to minimize the overall deviation between the fitted curve and the data points, thus generating an initial B-spline curve. In this embodiment of the invention, the ordered target point sequence of the AE mixed transport pipeline is used as the object. This sequence contains 68 target points, each of which contains three-dimensional coordinates in the CGCS2000 geodetic coordinate system. The coordinate unit is meters, and the coordinates are arranged sequentially from the starting point to the ending point according to the direction of the pipeline extension. to The sequence is parameterized using the cumulative chord length parameterization method. The parameter value of each data point in the cumulative chord length parameter space is calculated. The specific calculation process is as follows: First, the chord length between two adjacent target points is calculated. The formula for chord length calculation is... ( Then calculate the distance of each target point relative to the start of the sequence. cumulative chord length , The cumulative chord length of each target point Divide by the total cumulative chord length of the sequence To obtain the parameter values ​​for each data point. The parameter values ​​range from [0,1], thus generating a parameterized point sequence. , , , The B-spline curve is set to degree 3, with 14 control points (the number of control points is set to 1 / 5 of the target number to ensure the fitted curve has both smoothness and good fit), and node vectors... Assuming a uniform distribution, the number of nodes is the number of control points + curve degree + 1 = 18, and the node values ​​are as follows: =0、 =0、 =0、 =0、 =0.0625、 =0.125、 =0.1875、 =0.25、 =0.3125、 =0.375、 =0.4375、 =0.5、 =0.5625、 =0.625、 =0.6875、 =0.75、 =1、 =1、 =1 (repeated 3 times at each end to ensure the curve endpoints coincide with the start and end points of the target point sequence). The parameterized point sequence is approximated using the least squares method to construct the objective function. ( =0,1,...,67), where P For B-spline curves in parameters The coordinates of the location To obtain the coordinates of the target point in the parameterized point sequence, we obtain the normal equation system by differentiating the objective function and setting the derivative to zero. ,in The matrix is ​​of order n×m (n is the number of target points, 68, and m is the number of control points, 14), and its elements are... ( Index the target point. (for control point indexes). The coordinate matrix of control points is of order m×3. For an n×3 matrix, the matrix elements are... Solving the system of equations yields the three-dimensional coordinates of 14 control vertices. Based on the control vertices and the set node vectors and curve order, an initial B-spline curve is generated.

[0041] Calculate the vertical distance from each original data point to the initial B-spline curve to obtain a set of fitting residuals. Calculate the standard deviation based on the distribution of the fitting residuals to generate a fitting accuracy evaluation index. In this embodiment of the invention, the vertical distance from each original data point in the ordered target point sequence of the AE mixed transport pipeline to the initial B-spline curve is calculated based on the generated initial B-spline curve; this distance is the fitting residual. The fitting residual is calculated using the formula for the vertical distance from a spatial point to the B-spline curve, for each target point... First, in the B-spline curve Searching for and The nearest point P , The parameter value corresponding to the minimum distance is obtained by solving Newton's iteration method. The iterative formula is The iteration terminates when the parameter difference between two iterations is less than 1e-6, yielding... Then, the fitting residuals ,in , , They are respectively P The north and east coordinates and elevation values ​​were obtained. 68 fitting residuals were calculated, and their standard deviation was calculated based on the distribution of these residuals. This standard deviation was used as an indicator of fitting accuracy. The formula for calculating the standard deviation is: ,in The average of the 68 fitted residuals. =68 represents the number of target points. This indicator is used to measure the overall fit between the initial fitted curve and the original target point sequence, providing a basis for subsequent curve optimization. The remaining six pipelines were processed using the same method, calculating their respective initial B-spline curve fitting residuals and standard deviations to complete the fitting accuracy evaluation.

[0042] Curve optimization is performed using iterative reweighted least squares. Each data point is assigned a weight that is inversely proportional to its fitting residual. Outliers with fitting residuals greater than or equal to the residual scale estimate obtained from the fitting residuals are assigned a weight of 0. Then, the B-spline curve is refitted to generate the optimized B-spline curve. In this embodiment of the invention, the initial B-spline curve is optimized using an iterative reweighted least squares method. The core logic is to reduce the impact of outliers on the fitting results through weight allocation, thereby improving the robustness of the curve. Taking the AE mixed-transport pipeline as an example, based on the calculated 68 fitting residuals, a weight is assigned to each data point. The weight is inversely proportional to its fitting residual. The Tukey weight function is used to calculate the weight value. ,in For residual scaling estimator, For the first The fitting residuals for each target point, where the fitting residuals are greater than or equal to the residual scaling estimate obtained from the fitting residuals. outliers ≥1, weight =0, assigning the lowest weight to completely exclude its influence on the fit; for fitting residuals smaller than the residual scaling estimate The normal point, <1, assign The weights approach 1, meaning the interval is (0,1], preserving their contribution to the fit. After weight allocation, the weighted least squares objective function is reconstructed. ( =0,1,...,67), and solve the weighted system of equations using the same method as described above. ,in For a weighted matrix, the elements , For the new control point coordinate matrix, For a weighted matrix, the elements The new control vertex coordinates are obtained by solving the problem. Combined with the original node vectors and curve order, the B-spline curve is refitted to generate an optimized B-spline curve. The number of iterations is set to 3. In each iteration, the weights are recalculated based on the residuals of the previous fitted curve, and then the curve is fitted again. After the third iteration, the weight distribution tends to stabilize, and the optimized B-spline curve is obtained, which effectively eliminates the influence of outliers caused by seabed measurement interference.

[0043] The optimized B-spline curve is checked to ensure that its curvature change is continuous and physically reasonable, avoiding unnatural and severe bending, and finally outputting a smooth, continuous three-dimensional spatial curve that accurately reflects the actual direction of the pipeline.

[0044] In this embodiment of the invention, the generated optimized B-spline curve is inspected, with the core inspection focusing on the continuity and physical rationality of curvature changes. This ensures the curve avoids unnatural, abrupt bends and closely conforms to the actual route of the subsea pipeline. The curvature calculation uses the B-spline curve curvature formula. For the optimized B-spline curve... First, calculate its first derivative. and second derivative First derivative ( =0,1,...,13), second derivative ( =0,1,...,13), curvature , where × represents the cross product of vectors, and |·| represents the magnitude of the vector. In the parameters Within the range [0,1], 500 sampling points are taken at intervals of 0.01. The curvature value of each sampling point is calculated, and a curvature change curve is plotted. The curvature change is checked to see if it is continuous without abrupt jumps, and the curvature value is controlled between 0.001 and 0.005m. -1 Within the specified range (determined based on the subsea pipeline laying process, conforming to the actual bending characteristics of the pipeline, and avoiding excessive bending), the optimized B-spline curve of the AE mixed-transport pipeline, after inspection, showed that the curvature values ​​of all sampling points were within the specified range, the curvature changes were continuous, and there were no sharp bends. Physically, it conformed to the laying characteristics of subsea pipelines and could accurately reflect the actual pipeline route. The optimized B-spline curves of the remaining six pipelines were checked for curvature using the same method to ensure that each curve met the requirements of smoothness, continuity, and physical rationality. Finally, the optimized B-spline curve for each subsea pipeline was output as the pipeline's continuous three-dimensional spatial trajectory.

[0045] Furthermore, the curve optimization using the iterative reweighted least squares method includes the following steps: After completing the initial B-spline curve fitting, the vertical distance from each original data point to the curve is calculated and denoted as the initial residual. The median absolute deviation is calculated based on all the initial residuals to generate the residual scaling estimate. In this embodiment of the invention, taking the AE mixed-transport pipeline as the object, after completing the initial B-spline curve fitting, for each of the 68 original data points in the ordered target point sequence of the pipeline, the vertical distance from each data point to the initial B-spline curve is calculated one by one. This distance is the initial residual. ( (e.g., 0, 1, ..., 67), the initial residual calculation method is consistent with the fitting residual calculation method, that is, finding the nearest point on the initial B-spline curve for each data point using Newton's iteration method, and then calculating the three-dimensional Euclidean distance between the two points to obtain 68 initial residual values. Based on the median absolute deviation calculated from all initial residuals, the absolute value of each of the 68 initial residuals is first taken to obtain... ( (e.g., 0, 1, ..., 67) These absolute values ​​are sorted in ascending order, and the value at the middle position after sorting is taken as the median, denoted as . Median absolute deviation Residual scaling estimator generated based on median absolute bias The coefficient 1.4826 is a correction factor used to convert the median absolute deviation into a scaling estimate equivalent to the standard deviation of the normal distribution. This residual scaling estimate measures the overall dispersion of the initial residuals, providing a quantitative basis for subsequent robust weight calculations. This avoids interference from single residual inconsistencies in weight allocation, ensuring the rationality and stability of weight calculations. The remaining six pipelines follow the same process, calculating their respective initial residuals, median absolute deviations, and residual scaling estimates to provide basic parameters for the iterative optimization of curves in each pipeline.

[0046] Based on the residual scale estimate, a robust weight is calculated for each data point. The weight function uses the Tukey double weight function, which ensures that points with initial residuals greater than or equal to the residual scale estimate receive zero weights, while points with initial residuals less than the residual scale estimate receive weights in the interval (0,1). This generates a robust weight value for each data point. The robust weight value is then used as the weight matrix for the weighted least squares method. The normal equations are reconstructed and solved to obtain a new set of control vertex coordinates, thereby generating the B-spline curve after the first iteration of optimization. In this embodiment of the invention, the residual scaling estimate is obtained through calculation. Robust weights were calculated for each of the 68 original data points of the AE mixed-transport pipeline. The Tukey double-weight function was used, and its expression is as follows: ,in For the first Robust weight values ​​for each data point For the first The initial residuals of each data point This is a residual scaling estimator. When the initial residuals of the data points... Larger than the residual scaling estimator hour, >1, at this time <0, assign weight =0, this data point is identified as an outlier, and its influence on the fitting results is largely eliminated; when the initial residual of the data point is 0, the data point is considered an outlier. Smaller than the residual scaling estimator hour, <1, at this time >0, assign weight The data point within the interval (0,1] is considered a normal point, retaining its dominant role in the fitting result; when hour, =0, which perfectly eliminates the influence of that data point. After calculation, robust weight values ​​are obtained for 68 data points. All robust weight values ​​are then constructed into a 68×68 diagonal weight matrix. The diagonal elements of the matrix are the robust weight values ​​of each data point. The off-diagonal elements are all 0. This weight matrix serves as the core parameter of the weighted least squares method, used to reconstruct the normal equations. The new normal equations are: ,in It is an n×m matrix (n=68, m=14). The robust weight matrix is The coordinate matrix of the control points. Given the 3D coordinate matrix of 68 original data points, a new set of 14 control vertex 3D coordinates is obtained by solving the normal equations. Combined with the set node vectors and the 3rd order B-spline basis function, the curve is refitted to generate the B-spline curve after the first iteration optimization. This curve has initially reduced the interference of outliers and the fit is significantly improved compared with the initial B-spline curve.

[0047] Calculate the residual from the data point to the curve after the first iteration of optimization, update the residual scaling estimate, and recalculate the robust weight of each data point accordingly to generate the updated weight value. In this embodiment of the invention, the vertical distance from the 68 original data points of the AE mixed-transport pipeline to the generated B-spline curve after the first iteration optimization is calculated. This distance is the residual of the first iteration. ( =0,1,...,67), the calculation method is consistent with the initial residuals. Newton's iteration method is used to find the nearest point on the first iteration optimization curve for each data point, and the three-dimensional Euclidean distance between the two points is calculated, resulting in 68 first iteration residuals. Based on this set of iterative residuals, the residual scaling estimate is updated: first, the absolute value of the 68 first iteration residuals is taken, resulting in... Sort them in ascending order and take the median. Calculate the absolute deviation of the median. Updated residual scaling estimator Based on the updated residual scaling estimator Recalculate the robust weights for each data point, still using the Tukey double-weight function, expressed as follows: ,in For the updated robustness weight values, For the residual of the first iteration, This is the updated residual scaling estimate. Since the curve has eliminated some outliers after the first iteration of optimization, the dispersion of the iterative residuals is significantly reduced compared to the initial residuals. The updated residual scaling estimate... Less than the initial The corresponding robust weight allocation is more accurate. Some suspicious points that originally had low weights will have their weights increased if the iteration residual decreases, while the weights of the originally determined outliers will remain at 0, ensuring that the weight allocation always matches the actual distribution of the data points.

[0048] Using the updated weight values, perform weighted least squares fitting again to generate the B-spline curve optimized in the second iteration. Repeat this iterative process until the change in control vertex coordinates between two adjacent iterations is less than the preset convergence threshold. At this point, the final optimized B-spline curve is output.

[0049] In this embodiment of the invention, the updated robust weight value is used... Construct a new 68×68 diagonal weight matrix Using the weighted least squares fitting method, the normal equation system is reconstructed. ,in For the new control point coordinate matrix, solving the system of equations yields the 3D coordinates of the 14 control vertices after the second iteration. Combining this with the original node vectors and the 3rd-order B-spline basis functions, the optimized B-spline curve for the second iteration is generated. This iterative process is repeated: each iteration first calculates the iterative residual from the current data point to the current optimized curve, updates the residual scale estimate, recalculates the robust weights, constructs the weight matrix, solves the normal equations to obtain new control vertices, and generates a new optimized curve. A preset convergence threshold of 1e-5m is used to determine whether the iteration terminates, based on the trajectory accuracy requirements for subsea pipeline detection. The convergence threshold is determined by the maximum difference in the 3D coordinates of the corresponding vertices between two adjacent iterations being less than 1e-5m. During the iteration process, the change in control vertex coordinates gradually decreases after each iteration. After the fourth iteration, the change in control vertex coordinates between two adjacent iterations (the third and fourth) is calculated. The maximum value is 8.6 × 10⁻⁶ m, which is less than the preset convergence threshold 1e⁻⁵ m, at which point the iteration terminates. At this point, the B-spline curve optimized by the fourth iteration is output as the final optimized B-spline curve for the AE mixed-transport pipeline. The remaining six pipelines follow the same iteration process, setting the same convergence threshold, completing iterative optimization, and outputting their respective final optimized B-spline curves to ensure that each curve can effectively eliminate outlier interference and closely match the actual pipeline route.

[0050] Furthermore, the construction of a pipeline condition evolution prediction model using historical survey data and pipeline spatial condition assessment reports includes: Collect historical survey data of each pipeline in the target sea area, including the three-dimensional spatial trajectory of the pipeline, burial depth data, pipeline corrosion detection reports and marine environmental data obtained from each survey, and construct a database of the pipeline's full life cycle status evolution; In this embodiment of the invention, nearly 10 years of survey data were collected for seven subsea pipelines (AE mixed transport, EA water injection, BA mixed transport, AB water injection, CA mixed transport, AC water injection, and DE mixed transport) within the target sea area. The survey was conducted annually, resulting in 10 rounds of complete survey data collection. The three-dimensional spatial trajectory of the pipeline obtained from each survey was optimized curve data generated through three-dimensional B-spline curve fitting, including three-dimensional coordinates (CGCS2000 geodetic coordinate system, UTM projection, coordinate unit m) sampled at 0.5m intervals along the entire pipeline mileage. The burial depth data were calculated values ​​for the burial depth of the corresponding sampling points, according to the formula... The calculation yielded, where The elevation of the seabed interface at the corresponding location (obtained by combining water depth measurement data with tidal level correction). The elevation of the target point for the pipeline. The depth difference from the top of the pipeline to the seabed interface is calculated (derived from acoustic image interpretation). The pipeline corrosion detection report includes pipe wall thickness data, corrosion defect location and size data, measured at 1-meter intervals along the entire pipeline length. The marine environmental data includes ocean current velocity (average value based on 4 daily samplings, unit: m / s), sediment transport rate (average value based on 1 monthly sampling, unit: kg / (m·s)), and wave load (load value corresponding to significant wave height based on 4 daily samplings, unit: N / m) for the target sea area within the corresponding survey period. All collected data are categorized and organized by pipeline number and survey year, with each pipeline corresponding to an independent data file. The file is divided into four main categories: 3D spatial trajectory, burial depth data, pipeline corrosion detection data, and marine environmental data. All data are archived chronologically to construct a database of the pipeline's full life cycle status evolution. Each data entry in the database is uniquely identified using the coding rule "pipeline number-survey year-data type-serial number" to ensure data traceability, completeness, and non-duplication.

[0051] From the pipeline's full life cycle state evolution database, the burial depth data of the same pipeline at different periods are extracted, the vertical settlement at each mileage point of the pipeline between two adjacent surveys is calculated, and a pipeline historical settlement rate sequence is generated. In this embodiment of the invention, the burial depth data of each pipeline at different periods is extracted from the pipeline life cycle status evolution database according to the pipeline number. Taking the AE mixed transport pipeline as an example, the burial depth data for each year of the past 10 years (2014-2023) is extracted. Each annual burial depth data includes the burial depth value sampled at 0.5m intervals along the entire pipeline mileage. The sampling point numbers are sequentially coded from the pipeline start point to the end point as M0001, M0002, ..., M1360 (corresponding to a total pipeline length of 680m). The vertical settlement at each mileage point of the pipeline between two adjacent surveys is calculated. The vertical settlement is calculated by subtracting the burial depth value of the previous survey from the burial depth value of the later survey. If the calculation result is positive, it indicates that the pipeline has settled; if it is negative, it indicates that the pipeline has risen. The unit of settlement is meters. Taking 2014-2015 as an example, the burial depth values ​​for all sampling points from M0001 to M1360 of the AE mixed-transport pipeline in 2014 and 2015 were extracted, and the vertical settlement of each sampling point was calculated one by one to obtain the settlement data of the pipeline in 2014-2015. Using the same method, the settlement data for each year from 2015-2016, 2016-2017 to 2022-2023 were calculated in sequence. The settlement of each sampling point in different years was arranged in chronological order to generate the historical settlement rate sequence of the sampling point. The settlement rate was calculated by dividing the annual settlement by the survey interval (1 year), with the unit being m / year. The historical settlement rate sequences of all sampling points were summarized to form the historical settlement rate sequence of the AE mixed-transport pipeline. The corresponding burial depth data and vertical settlement were extracted and the historical settlement rate sequence of the other 6 pipelines were generated in the same way to ensure that each mileage point has a complete historical settlement rate record.

[0052] The correlation between the historical settlement rate sequence of the pipeline and the marine environmental data of the same period was analyzed. The marine environmental data included ocean current velocity, sediment transport rate and wave load. A settlement rate prediction sub-model based on multiple linear regression was established. In this embodiment of the invention, the historical settlement rate sequence of the AE mixed-transport pipeline is used as the analysis object. Simultaneously, marine environmental data (ocean current velocity, sediment transport rate, and wave load) for the corresponding period are extracted from the database. Correlation analysis is then performed between the historical settlement rate sequence and the corresponding marine environmental data. The Pearson correlation coefficient method is used for correlation analysis, calculating the correlation coefficients between settlement rate and ocean current velocity, settlement rate and sediment transport rate, and settlement rate and wave load. The correlation coefficient ranges from -1 to 1; the closer the absolute value is to 1, the stronger the correlation, and the closer the absolute value is to 0, the weaker the correlation. The calculated correlation coefficients are 0.72 for settlement rate and ocean current velocity, 0.85 for sediment transport rate, and 0.68 for wave load, all showing strong positive correlations. That is, the greater the ocean current velocity, the faster the sediment transport rate, and the greater the wave load, the greater the pipeline settlement rate. Based on these correlation results, a settlement rate prediction sub-model based on multiple linear regression is established. The model structure is as follows: ,in The pipeline settlement rate is given in m / year. The velocity of the ocean current is (m / s). , where is the sediment transport rate (kg / (m·s)). Wave load (N / m). , , For regression coefficients, The constant term is used. The model training data uses historical settlement rate data of the AE mixed-transport pipeline over the past 8 years and corresponding marine environmental data for the same period, totaling 8 sets of annual data. Each set of data includes the average settlement rate and the corresponding average values ​​of ocean current velocity, sediment transport rate, and wave load. The regression coefficients and constant term are solved using the least squares method. The solution process is as follows: Construct the objective function. ( =1,2,...,8), calculate the objective function respectively. , , , The partial derivatives of the equations are calculated and set to zero to obtain a system of linear equations. Solving these equations yields the corresponding sub-model for predicting settlement rate. The other six pipelines are analyzed using the same method to determine their correlations and establish their own multiple linear regression sub-models for predicting settlement rate, ensuring that the model parameters match the actual conditions of the pipelines.

[0053] Extract the location and length of the current pipeline span from the pipeline spatial state assessment report, calculate the vortex-induced vibration frequency and amplitude acting on the span by combining ocean current data, and query the material fatigue characteristic curve of the pipeline to establish a fatigue life prediction sub-model of the span based on fracture mechanics. In this embodiment of the invention, the locations and lengths of the suspended sections of each pipeline in the current year (2023) are extracted from the pipeline spatial status assessment report. Taking the AE mixed transport pipeline as an example, three suspended sections are identified: Suspended section 1 (location: mileage 120m-135m, length 15m), Suspended section 2 (location: mileage 350m-368m, length 18m), and Suspended section 3 (location: mileage 520m-532m, length 12m). Current and historical ocean current data for the target sea area are collected, and the average ocean current velocity corresponding to each suspended section location is extracted (the average of four daily samples over the past year, in m / s). The ocean current velocity corresponding to suspended section 1 is 1.2 m / s, that of suspended section 2 is 1.5 m / s, and that of suspended section 3 is 1.1 m / s. The vortex-induced vibration frequency is calculated using the formula... Calculation, where The value is the Strauhal number (0.2, determined based on the pipe cross-sectional shape and the seawater medium). The velocity of the ocean current is (m / s). The pipe diameter is 0.6m; the vortex-induced vibration amplitude is calculated using the formula... Calculation, where The length of the suspended span is (m). The bending stiffness of the pipeline (value taken as 2.8 × 10⁻⁶) 9 N·m 2 (Based on pipe materials and specifications) The force exerted by the ocean current on the suspended section (according to the formula) Calculation, where The density of seawater is taken as 1025 kg / m³. 3 , The drag coefficient is set to 1.2, determined based on the pipe surface roughness. The fatigue limit of the pipe material is found to be 1.2 × 10⁻⁶, based on its fatigue characteristic curve. 7 Next, a sub-model for predicting the fatigue life of the suspended segment based on fracture mechanics is established. The model structure is as follows: ,in Fatigue life (number of cycles). For the fracture toughness of the pipeline material, The stress intensity factor amplitude (according to the formula) calculate, The alternating stress generated by vortex-induced vibration is calculated according to the formula. calculate, For pipe wall thickness, The height of the pipe cross-section; The initial crack length of the pipeline is given. The fatigue life of the corresponding suspension section is calculated by substituting the parameters of each suspension section. The other 6 pipelines are treated in the same way: the parameters of the suspension section are extracted, the relevant parameters of vortex-induced vibration are calculated, and the fatigue characteristic curves are queried to establish their respective suspension section fatigue life prediction sub-models.

[0054] Pipe wall thickness loss data were extracted from historical corrosion detection reports, a function of corrosion depth over time was fitted, and a sub-model for predicting pipeline corrosion rate based on physicochemical reactions was established by combining seawater chemical properties, protective layer status, and cathodic protection system operation data. In this embodiment of the invention, pipe wall thickness loss data for each pipeline is extracted from historical corrosion inspection reports in the pipeline lifecycle status evolution database. Taking the AE mixed-transport pipeline as an example, pipe wall thickness inspection data for each year over the past 10 years is extracted. Each inspection point (interval of 1m) has a corresponding initial pipe wall thickness (0.08m) and the annual inspection thickness. The pipe wall thickness loss for each inspection point in each year is calculated, where the thickness loss is the initial thickness minus the annual inspection thickness, in meters. Based on the thickness loss data for each inspection point, a function of corrosion depth changing over time is fitted using an exponential function, with the function form being: ,in for The corrosion depth (m) at any given time. The initial corrosion depth (valued at 0.001m, representing the corrosion depth detected during the initial stage of pipeline installation). The corrosion rate constant is 1 / year. The time period is in years. The corrosion depth data for each detection point over the past 10 years is substituted into the function, and the solution is obtained using the least squares method. The average corrosion rate constant of the pipeline was obtained by calculating the value. Data on the chemical properties of seawater in the target sea area (pH 7.8, salinity 35‰, dissolved oxygen 6.2 mg / L), the condition of the pipeline protective layer (protective layer integrity rate 92%, calculated based on the percentage of damaged protective layer area detected annually), and the operation data of the cathodic protection system (protective potential -0.85V, averaged based on daily monitoring data) were collected. These data were used as influencing factors and substituted into the pipeline corrosion rate prediction sub-model based on physicochemical reactions. The model structure is as follows: ,in The corrosion rate is expressed in m / year. The base corrosion rate (valued at 0.008 m / year, determined based on the annual average corrosion rate calculated using a fitting function) is used. The salinity of seawater is expressed in per mille (‰). The pH value of seawater. The protective layer integrity rate (decimal form). This is the cathodic protection potential (V, absolute value). , , , For the influencing factor coefficients, the model training data used corrosion rate data of the pipeline over the past 8 years, along with corresponding seawater chemical properties, protective layer status, and cathodic protection system operation data, totaling 8 sets of annual data. The least squares method was used to solve for each influencing factor coefficient to ensure that the model prediction results are consistent with the actual corrosion data. The other 6 pipelines were trained using the same method, fitting corrosion depth functions, collecting relevant influencing factor data, and establishing their own corrosion rate prediction sub-models.

[0055] By integrating the sub-models for predicting settlement rate, fatigue life of the suspended span, and corrosion rate, a comprehensive pipeline state evolution prediction model is constructed. This model uses the current state as the initial condition and is driven by future marine environmental prediction data to simulate the state evolution of the subsea pipeline over a specified period in the future. It outputs prediction results including a settlement trend map, a suspended span risk probability distribution map, and a remaining wall thickness distribution map, and generates a maintenance priority list accordingly.

[0056] In this embodiment of the invention, a comprehensive pipeline state evolution prediction model is constructed by integrating the established sub-models for settling rate prediction, fatigue life prediction, and corrosion rate prediction. The model uses the current (2023) pipeline state as the initial condition. The current state includes the three-dimensional spatial trajectory of each pipeline, the current burial depth, the current location and length of the suspended section, the current pipe wall thickness, and current marine environmental data, all extracted from the pipeline's full life cycle state evolution database and spatial state assessment report. The model is driven by the marine environmental prediction data of the target sea area for the next five years (2024-2028). The future marine environmental prediction data includes the annual predicted average values ​​of ocean current velocity, sediment transport rate, and wave load, as well as seawater chemical properties and wave load change trend data, all determined based on historical marine environmental data of the target sea area and marine environmental prediction standards. During model execution, the first step involves using a sub-model to predict settlement rate. Substituting future marine environmental prediction data, the sub-model calculates the settlement rate and cumulative settlement at each mileage point of the pipeline over the next five years, generating a settlement trend map. This map plots time on the horizontal axis and settlement amount on the vertical axis, marking the settlement change curves at each mileage point. The second step involves using a sub-model to predict the fatigue life of the suspended span. Combined with future ocean current prediction data, this sub-model calculates the cumulative fatigue damage of each suspended span over the next five years. Based on the fatigue life calculation results, a suspension risk probability distribution map is generated. Risk levels are categorized according to the proportion of cumulative fatigue damage to fatigue life (≥80% is high risk, 50%-80% is medium risk, <50% is low risk), and these levels are then labeled. The risk level and distribution location of each suspended section are determined. Then, using a corrosion rate prediction sub-model combined with future seawater chemistry prediction data, the wall thickness loss at each monitoring point of each pipeline over the next 5 years is calculated, generating a remaining wall thickness distribution map, and marking the remaining wall thickness and corrosion degree at each monitoring point. Finally, the above prediction results are integrated to generate a maintenance priority list. Maintenance priorities are divided according to risk level: high-risk suspended sections, high corrosion rate areas, and high settlement rate areas are listed as Level 1 maintenance priorities; medium-risk areas are listed as Level 2 maintenance priorities; and low-risk areas are listed as Level 3 maintenance priorities. The list clearly specifies the location, maintenance content, and maintenance time nodes for each maintenance area to ensure targeted maintenance work. The future state evolution simulation of the remaining 6 pipelines is completed using this comprehensive prediction model, outputting corresponding prediction results and maintenance priority lists, realizing full life cycle state management of all pipelines.

[0057] Furthermore, the present invention also provides an intelligent detection and survey system for submarine pipelines, including a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the intelligent detection and survey method for submarine pipelines as described above.

[0058] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for intelligent detection and investigation of subsea pipelines, characterized in that, Includes the following steps: Collect multi-source heterogeneous detection data corresponding to each submarine pipeline in the target sea area. The multi-source heterogeneous detection data includes acoustic profile data collected by a shallow seismic profiler mounted on the detection vessel, spatiotemporal coordinate data collected by a differential GNSS positioning system, ship motion attitude data collected by an attitude inertial navigation system, and design route planning data recorded by navigation software. Acoustic feature enhancement processing is performed on the acoustic profile data to generate an enhanced acoustic profile image sequence; Intelligent pipe target recognition is performed on the enhanced acoustic profile image sequence to extract abnormal reflection area pixel clusters that characterize the pipe contour in the image, and calculate their geometric center coordinates, equivalent diameter and reflection intensity distribution to generate a preliminary set of suspected pipe targets. By integrating spatiotemporal coordinate data, ship motion attitude data, and a preliminary set of suspected pipeline targets, the preliminary set of suspected pipeline targets is transformed from the instrument coordinate system to the geodetic coordinate system through coordinate transformation and attitude compensation algorithms, generating a cloud of pipeline target points. Based on the design route planning data and the pipeline target point cloud, a pipeline route correlation analysis is performed. By calculating the shortest vertical distance between the point cloud and the design route line, point clouds with a shortest vertical distance less than a set threshold are associated with the same pipeline. The point clouds are then sorted and topologically connected along the pipeline extension direction to obtain an ordered sequence of target points arranged in the pipeline extension direction. A 3D B-spline curve is fitted to the ordered sequence of target points to smooth out coordinate fluctuations caused by measurement errors or ocean current disturbances, generating a smooth and continuous 3D spatial curve as the continuous 3D spatial trajectory of the subsea pipeline. The elevation value corresponding to each target point and the depth difference from the top of the pipeline to the seabed interface, interpreted from the acoustic image, are extracted from the ordered sequence of target points to calculate the burial depth of the pipeline, generating a burial depth attribute dataset for the subsea pipeline. Based on the continuous three-dimensional spatial trajectory and burial depth attribute datasets corresponding to each subsea pipeline, a pipeline spatial status compliance analysis is conducted to generate a pipeline spatial status assessment report. Using historical survey data and the pipeline spatial status assessment report, a pipeline status evolution prediction model is constructed to predict the pipeline's settlement trend, overhang risk, and corrosion rate within a specified future period, generating a pipeline long-term service performance degradation prediction curve and a maintenance priority list.

2. The intelligent detection and survey method for submarine pipelines according to claim 1, characterized in that, The acoustic feature enhancement processing of the acoustic profile data includes the following steps: Short-time Fourier transform is performed on historical acoustic profile data to convert one-dimensional time-domain signals into two-dimensional time-frequency spectrograms, generating a time-frequency domain acoustic feature matrix. In the time-frequency domain acoustic feature matrix, an adaptive thresholding method is used to identify and mark continuous low-frequency high-energy regions formed by seawater reverberation and random noise, generating a noise energy mask map. A denoising model based on a deep convolutional neural network is constructed. The time-frequency domain acoustic feature matrix is ​​used as input, and the noise energy mask is used as a supervision label. The model is trained to learn the distribution differences between noise and effective signal, and a well-trained intelligent denoising model is generated. The real-time acquired acoustic profile data is input into the trained intelligent denoising model. The model outputs the clean time-frequency features after filtering out continuous background noise, generating a denoised time-frequency spectrum. An inverse short-time Fourier transform is performed on the denoised time-frequency spectrum to reconstruct a time-domain acoustic signal with suppressed reverberation noise, generating preliminary denoised acoustic profile data. To address the residual multipath reflection interference in the acoustic profile data after initial noise reduction, a matched filtering technique based on a prior model of pipe reflection is employed. A filter matching the typical reflection characteristics of the pipe is designed, and convolution operations are performed on the signal to further suppress reflection interference unrelated to the prior model, thereby generating an enhanced acoustic profile image sequence.

3. The intelligent detection and survey method for submarine pipelines according to claim 2, characterized in that, The intelligent identification of pipeline targets in the enhanced acoustic profile image sequence includes the following steps: Adaptive grayscale normalization and contrast stretching are performed on the enhanced acoustic profile image sequence to unify the brightness and contrast levels of images under different voyages and sea conditions, thereby generating a standardized acoustic image sequence. An improved U-Net network architecture is used to perform pixel-level semantic segmentation on standardized acoustic image sequences. The network encoder extracts multi-scale contextual features, the decoder recovers spatial details, and an attention gate mechanism is introduced in the skip connections to focus on the pipe target region, generating a preliminary segmentation map containing three categories of labels: "background", "pipe", and "seabed interface". Connectivity analysis is performed on the pixel regions marked as "pipes" in the preliminary segmentation map to remove noise regions with pixel areas smaller than the lower limit of the physical size of the pipes, generating a set of candidate pipe connected regions. For each candidate pipe connected region, the aspect ratio of its minimum bounding rectangle, the average reflection intensity of the pixels in the region, and the skewness and kurtosis of the reflection intensity distribution are calculated to generate a set of geometric and reflection feature descriptors. The geometric and reflection feature descriptors are input into a pre-trained pipeline target classifier, which is built based on support vector machines to distinguish the reflection of real pipelines from the reflection of similar-shaped geological bodies or debris. The classifier outputs the probability value of each candidate connected component as a real pipeline, generating a pipeline target probability set. A probability threshold is set, and candidate connected components with probability values ​​exceeding the threshold are filtered out. The geometric center coordinates of their pixel clusters are calculated as the target location. The equivalent diameter is calculated based on the circumscribed rectangle, and the intensity distribution histogram of pixels in the region is statistically analyzed to generate a preliminary set of suspected pipeline targets.

4. The intelligent detection and survey method for submarine pipelines according to claim 3, characterized in that, The pixel-level semantic segmentation of standardized acoustic image sequences using the improved U-Net network architecture includes the following steps: Construct a U-Net network backbone with encoder-decoder symmetry, where the encoder contains four downsampling stages, each of which uses two convolutional layers and one max pooling layer to progressively extract and compress image features, generating multi-level downsampled feature maps; An attention gate module is introduced at the skip connection of the network. This module takes the high-level semantic feature map passed from the encoder side as input and generates a spatial attention weight map. This map is used to weight the encoder features before feature fusion, so that the network pays more attention to the pipeline target region related to the current task of the decoder and generates weighted encoder features. The decoder consists of four upsampling stages. Each stage receives weighted encoder features from the corresponding layer, upsamples them through transposed convolutions, and concatenates them with the weighted encoder features. Then, it passes through two convolutional layers for feature fusion and refinement, gradually restoring the spatial resolution and generating the final high-resolution feature map. In the final layer of the network, a 1x1 convolution is used to map the number of channels in the high-resolution feature map to the number of categories, and the Softmax function outputs the probability that each pixel belongs to "background", "pipe", or "seabed interface", generating a preliminary segmentation map. During the model training phase, an acoustic image dataset containing labeled pipe contours is used. The weighted sum of the cross-entropy loss function and the Dice loss function is used as the optimization objective. The network parameters are adjusted through the backpropagation algorithm until the segmentation accuracy of the model on the validation set reaches the preset standard, thus generating a trained improved U-Net segmentation model.

5. The intelligent detection and survey method for subsea pipelines according to claim 1, characterized in that, The process of generating the pipeline target point cloud includes the following steps: The latitude, longitude, elevation, and timestamp of the probe's center of mass in the geodetic coordinate system are obtained from the differential GNSS positioning system to generate an absolute attitude time series of the hull; the roll angle, pitch angle, bow angle, and three-dimensional acceleration data in the hull coordinate system with the center of mass as the origin are obtained from the attitude inertial navigation system to generate a relative motion attitude time series of the hull; based on the installation position of the transducer of the shallow seismic profiler on the hull, the three-dimensional installation offset of the transducer's phase center relative to the center of mass of the hull is measured and calculated to generate an instrument installation parameter matrix; At each acoustic data acquisition moment, the absolute pose time series of the hull, the relative motion attitude time series of the hull, and the image coordinates of the target points in the preliminary pipeline suspected target set are synchronized according to the timestamp. Based on the instrument installation parameter matrix and the synchronized time series of the ship's relative motion attitude, the target point is transformed from the instrument coordinate system centered on the transducer to the ship coordinate system centered on the ship's center of mass through three-dimensional coordinate rotation and translation transformation, generating the target point coordinates in the ship coordinate system. By combining the synchronized ship absolute pose time series, the target point coordinates in the ship coordinate system are transformed to the CGCS2000 geodetic coordinate system through geodetic theme calculation, generating pipeline target point coordinates with absolute latitude, longitude and elevation. After integrating all time data, a pipeline target point aggregator is generated.

6. The intelligent detection and survey method for subsea pipelines according to claim 1, characterized in that, The pipeline route correlation analysis based on design route planning data and pipeline target point aggregation includes the following steps: The design routes of each subsea pipeline are extracted from the design route planning data. The design routes are represented by a three-dimensional spatial polyline formed by connecting a series of ordered design inflection point coordinates, and a three-dimensional line set of design routes is generated. For each target point in the pipeline target point cloud, traverse each design route segment in the design route three-dimensional line set, calculate the three-dimensional Euclidean vertical distance from the target point to each line segment, and record the minimum vertical distance and its corresponding design route line segment ID to generate a point-line minimum distance association table. Set a distance association threshold, filter out all target points whose minimum vertical distance in the point-line minimum distance association table is less than the threshold, and initially classify these target points as "points related to the design route" to generate a preliminary route-related point set; for each design route, extract all target points associated with it from the preliminary route-related point set, and spatially sort them based on the timestamps of these target points or the projection distance along the design route to generate a target point sequence arranged in an orderly manner according to the pipeline extension direction.

7. The intelligent detection and survey method for submarine pipelines according to claim 1, characterized in that, The process of fitting a three-dimensional B-spline curve to an ordered sequence of target points includes the following steps: Input an ordered sequence of target points, where each point contains three-dimensional coordinates. Parameterize the sequence and calculate the parameter value of each data point in the cumulative chord length parameter space to generate a parameterized point sequence. Set the degree of the B-spline curve, the number of control points, and the node vector. Approximate the parameterized point sequence using the least squares method. Calculate a set of coordinates of the control vertices by solving the system of equations to minimize the overall deviation between the fitted curve and the data points, thus generating an initial B-spline curve. Calculate the vertical distance from each original data point to the initial B-spline curve to obtain a set of fitting residuals. Calculate the standard deviation based on the distribution of the fitting residuals to generate a fitting accuracy evaluation index. Curve optimization is performed using iterative reweighted least squares. Each data point is assigned a weight that is inversely proportional to its fitting residual. Outliers with fitting residuals greater than or equal to the residual scale estimate obtained from the fitting residuals are assigned a weight of 0. Then, the B-spline curve is refitted to generate the optimized B-spline curve. The optimized B-spline curve is checked to ensure that its curvature change is continuous and physically reasonable, avoiding unnatural and severe bending, and finally outputting a smooth, continuous three-dimensional spatial curve that accurately reflects the actual direction of the pipeline.

8. The intelligent detection and survey method for submarine pipelines according to claim 7, characterized in that, The curve optimization using the iterative reweighted least squares method includes the following steps: After completing the initial B-spline curve fitting, the vertical distance from each original data point to the curve is calculated and denoted as the initial residual. The median absolute deviation is calculated based on all the initial residuals to generate the residual scaling estimate. Based on the residual scale estimate, a robust weight is calculated for each data point. The weight function uses the Tukey double weight function, which ensures that points with initial residuals greater than or equal to the residual scale estimate receive zero weights, while points with initial residuals less than the residual scale estimate receive weights in the interval (0,1). This generates a robust weight value for each data point. The robust weight value is then used as the weight matrix for the weighted least squares method. The normal equations are reconstructed and solved to obtain a new set of control vertex coordinates, thereby generating the B-spline curve after the first iteration of optimization. Calculate the residual from the data point to the curve after the first iteration of optimization, update the residual scaling estimate, and recalculate the robust weight of each data point accordingly to generate the updated weight value. Using the updated weight values, perform weighted least squares fitting again to generate the B-spline curve optimized in the second iteration. Repeat this iterative process until the change in control vertex coordinates between two adjacent iterations is less than the preset convergence threshold. At this point, the final optimized B-spline curve is output.

9. The intelligent detection and survey method for submarine pipelines according to claim 1, characterized in that, The method of constructing a pipeline condition evolution prediction model using historical survey data and pipeline spatial condition assessment reports includes: Collect historical survey data of each pipeline in the target sea area, including the three-dimensional spatial trajectory of the pipeline, burial depth data, pipeline corrosion detection reports and marine environmental data obtained from each survey, and construct a database of the pipeline's full life cycle status evolution; From the pipeline's full life cycle state evolution database, the burial depth data of the same pipeline at different periods are extracted, the vertical settlement at each mileage point of the pipeline between two adjacent surveys is calculated, and a pipeline historical settlement rate sequence is generated. The correlation between the historical settlement rate sequence of the pipeline and the marine environmental data of the same period was analyzed. The marine environmental data included ocean current velocity, sediment transport rate and wave load. A settlement rate prediction sub-model based on multiple linear regression was established. Extract the location and length of the current pipeline span from the pipeline spatial state assessment report, calculate the vortex-induced vibration frequency and amplitude acting on the span by combining ocean current data, and query the material fatigue characteristic curve of the pipeline to establish a fatigue life prediction sub-model of the span based on fracture mechanics. Pipe wall thickness loss data were extracted from historical corrosion detection reports, a function of corrosion depth over time was fitted, and a sub-model for predicting pipeline corrosion rate based on physicochemical reactions was established by combining seawater chemical properties, protective layer status, and cathodic protection system operation data. By integrating the sub-models for predicting settlement rate, fatigue life of the suspended span, and corrosion rate, a comprehensive pipeline state evolution prediction model is constructed. This model uses the current state as the initial condition and is driven by future marine environmental prediction data to simulate the state evolution of the subsea pipeline over a specified period in the future. It outputs prediction results including a settlement trend map, a suspended span risk probability distribution map, and a remaining wall thickness distribution map, and generates a maintenance priority list accordingly.

10. An intelligent detection and survey system for subsea pipelines, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, for performing the intelligent detection and survey method for submarine pipelines as described in any one of claims 1-9.

Images