A method for monitoring micro-leakage of a submarine pipeline

Abstract

The application provides a kind of submarine pipeline micro leakage monitoring method.The method comprises: obtaining the acoustic vibration signal of distributed optical fiber sensor arranged along submarine pipeline, constructing two-dimensional space-time strain phase matrix;Two-dimensional space-time strain phase matrix is respectively input into the space spectrum attention flow branch and time series attention flow branch of double flow depth neural network;Wherein, the space-frequency domain feature map after suppressing low-frequency environmental noise is output by space spectrum attention flow branch;The time series feature vector is output by time series attention flow branch;Space-frequency domain feature map and time series feature vector are input into feature fusion module to carry out cross-dimension feature fusion step, and the micro leakage occurrence probability of each spatial position channel along submarine pipeline is output;Based on micro leakage occurrence probability, space integration anomaly detection algorithm based on double sliding window is used for space-time joint determination.The application realizes high sensitivity extraction and high reliability discrimination of submarine pipeline micro leakage signal in complex marine environment.

Description

Technical Field

[0001] This invention relates to the field of submarine pipeline safety monitoring technology, specifically to a method for monitoring micro-leakage in submarine pipelines. Background Technology

[0002] Submarine oil and gas pipelines are critical infrastructure for transporting marine oil and gas resources, and their operational safety is directly related to marine ecological environment protection and energy supply stability. Due to the long-term exposure to high pressure, corrosion, and geological movements, submarine pipelines are highly susceptible to developing micro-perforations or cracks in their walls, leading to micro-leakage accidents. If these micro-leakages are not detected and addressed in time, they will gradually expand, eventually causing catastrophic ruptures and marine pollution incidents. Therefore, implementing high-precision, low-false-detection real-time monitoring of micro-leakages in submarine pipelines has significant engineering value and practical implications.

[0003] Among existing technologies, Distributed Acoustic Sensing (DAS) has become an important means of monitoring subsea pipelines due to its ability to acquire acoustic vibration signals over long distances and in continuous spatial coverage. This technology uses optical fibers laid along the pipeline as the sensing medium, and senses external acoustic vibration signals by detecting the phase change of Rayleigh scattered light in the optical fiber. It has advantages such as resistance to electromagnetic interference, corrosion resistance, and the ability to cover ultra-long distances.

[0004] However, applying DAS technology to monitor micro-leakage in subsea pipelines still faces several technical challenges. The marine environment is subject to strong background noise interference, primarily including low-frequency pressure pulsations caused by periodic wave fluctuations, low-frequency vibrations from propellers and mechanical components during ship navigation, marine biological activity, and ocean current erosion. The energy of this background noise is often far greater than the weak high-frequency signal generated by high-pressure fluid jets through tiny perforations in the pipe wall, resulting in an extremely low signal-to-noise ratio. Traditional monitoring methods typically employ fixed threshold discrimination or simple frequency domain filtering, which struggles to effectively extract micro-leakage characteristic signals in complex, time-varying marine noise environments, leading to persistently high false alarm and missed alarm rates.

[0005] In recent years, deep learning-based signal processing methods have been introduced into this field, attempting to improve detection performance through the feature extraction capabilities of neural networks. However, existing solutions mostly employ a single network structure, compressing spatiotemporal information into a one-dimensional sequence or a simple two-dimensional image for processing, failing to fully utilize the joint distribution characteristics of micro-leakage signals in the spatiotemporal dimensions. Micro-leakage signals have specific sweep ranges and attenuation patterns in space, and exhibit continuous non-stationary random processes in time, while background noise displays significant spatiotemporal heterogeneity. A single network struggles to simultaneously address noise suppression in the spatial frequency domain and capture weak signals in the temporal domain, and lacks dynamic adaptability to the physical environment, resulting in insufficient robustness of the model under harsh sea conditions. Furthermore, existing methods often employ single-point instantaneous probability thresholds in the output layer, failing to fully consider the spatiotemporal continuity of micro-leakage as a continuous physical process, making transient interference highly susceptible to triggering false alarms, severely hindering the engineering practicality of the technology.

[0006] Therefore, there is an urgent need in this field for a new method for monitoring micro-leaks in subsea pipelines that can achieve spatiotemporal joint analysis and has strong environmental adaptability. Summary of the Invention

[0007] Based on the above background, the purpose of this invention is to provide a method for monitoring micro-leakage in subsea pipelines, which can achieve high-sensitivity extraction and high-reliability discrimination of micro-leakage signals in subsea pipelines under complex marine environments, and effectively reduce false alarm rate and missed alarm rate.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A method for monitoring micro-leakage in subsea pipelines, the method comprising:

[0010] Step S1: Acquire the acoustic vibration signal from the distributed optical fiber sensor installed along the seabed pipeline, discretize it according to the set spatial resolution and time sampling rate, and construct a two-dimensional spatiotemporal strain phase matrix.

[0011] Step S2: The two-dimensional spatiotemporal strain phase matrix is ​​input into the spatial spectrum attention flow branch and the time series attention flow branch of the dual-flow deep neural network, respectively; wherein, the spatial spectrum attention flow branch performs spatiotemporal frequency domain transformation and frequency masking steps, and outputs a spatial-frequency domain feature map after suppressing low-frequency environmental noise; the time series attention flow branch sequentially performs high-pass filtering and time series modeling steps to extract the time series features of micro-leakage sound and outputs a time series feature vector;

[0012] Step S3: Input the spatial-frequency domain feature map and the temporal feature vector into the feature fusion module to perform a cross-dimensional feature fusion step, and output the probability of micro-leakage at each spatial location along the submarine pipeline.

[0013] Step S4: Based on the probability of micro-leakage occurrence, a spatiotemporal joint determination is performed using a dual sliding window spatial integral anomaly detection algorithm. When the preset spatiotemporal persistence condition is met, a micro-leakage warning is triggered and the leak location is output.

[0014] Preferably, in step S2, the steps performed by the spatial spectral attention flow branch specifically include:

[0015] A short-time Fourier transform is performed on the two-dimensional spatiotemporal strain phase matrix along the time axis to generate a three-dimensional space-frequency-time tensor.

[0016] The external physical environment data of the sea area where the subsea pipeline is located is obtained, the corresponding low-frequency cutoff frequency is calculated based on the external physical environment data, and a low-frequency attention mask is constructed.

[0017] The low-frequency attention mask is multiplied element-wise with the spatial-frequency-time three-dimensional tensor to obtain the noise reduction tensor;

[0018] The noise reduction tensor is input into a feature extraction module that includes a two-dimensional residual network and a spatial attention network, and the spatial-frequency domain feature map is output.

[0019] Preferably, the external physical environment data includes sea state remote sensing data or the root mean square amplitude of the background noise of the distributed optical fiber sensor within a preset historical time period.

[0020] Preferably, in step S2, the steps performed by the time-series attention flow branch include:

[0021] The two-dimensional spatiotemporal strain phase matrix is ​​subjected to high-pass filtering to remove low-frequency noise bands and obtain one-dimensional high-frequency time-series signals corresponding to each spatial location channel.

[0022] A gated recurrent network is used to perform time-series modeling on the one-dimensional high-frequency time-series signal.

[0023] The time dependencies within the one-dimensional high-frequency time-series signal are calculated using a multi-head attention network, and the time-series feature vector is output.

[0024] Preferably, in step S3, the steps performed by the feature fusion module specifically include:

[0025] The spatial-frequency domain feature map and the temporal feature vector are tensor-concatenated and then split into a first processing branch and a second processing branch.

[0026] In the first processing branch, channel scaling is performed through a first 1×1 convolutional layer;

[0027] In the second processing branch, after channel mapping is performed through the second 1×1 convolutional layer, the features are input to a local depth feature extractor containing N reparameterized residual blocks.

[0028] The output features of the first processing branch are added and fused with the output features of the local deep feature extractor, and then the probability of micro-leakage is output through a multilayer perceptron and a normalized exponential function layer.

[0029] Preferably, during the model training phase, the internal structure of each reparameterized residual block includes parallel 3×3 convolution branches, 1×1 convolution branches, and identity mapping branches. The output feature of the reparameterized residual block is the element-wise sum of the output features of the three parallel branches.

[0030] Preferably, in step S4, the step of performing spatiotemporal joint determination using the dual sliding window spatial integral anomaly detection algorithm specifically includes:

[0031] Within the first spatial sliding window, the width of the spatial sliding window is adjusted based on the historical acoustic background noise variance of each spatial location channel, and the spatial integral value of the probability of the micro-leakage occurring within the spatial sliding window is calculated.

[0032] Within the second time window, a time sliding buffer queue with a preset frame count of k is established;

[0033] If the spatial integral value of a certain local area is greater than the set safety threshold for k consecutive time frames in the time sliding buffer queue, then the micro-leakage warning is executed.

[0034] Preferably, the step of adjusting the width of the spatial sliding window based on the historical acoustic background noise variance of each spatial location channel specifically includes:

[0035] When the historical acoustic background noise variance of a certain spatial location channel is greater than the mean threshold, the width of the spatial sliding window is reduced.

[0036] When the variance of the historical acoustic background noise is less than or equal to the mean threshold, the width of the spatial sliding window is increased.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] This invention discloses a method for monitoring micro-leakage in subsea pipelines. By constructing a two-dimensional spatiotemporal strain phase matrix, it elevates the one-dimensional signal to a spatiotemporal joint analysis space, facilitating subsequent deep learning processing. Through a dual-stream asynchronous architecture, the spatial-spectral attention flow branch suppresses dynamic environmental noise, while the time-series attention flow branch focuses on extracting weak leak signals. Cross-dimensional feature reparameterization fusion integrates spatial frequency domain features with temporal features, avoiding information dilution. Dual sliding-window spatial integral anomaly detection maps the output probability to a spatiotemporal joint judgment conforming to physical laws, effectively reducing false alarm and missed alarm rates. This invention deeply integrates prior physical knowledge with the advantages of deep learning, possessing strong environmental adaptability and high engineering practical value, providing an effective technical solution for monitoring micro-leakage in subsea pipelines. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the overall process architecture of a method for monitoring micro-leakage in a subsea pipeline provided by an embodiment of the present invention;

[0041] Figure 2 This is a detailed processing flowchart of the spatial spectrum attention flow branch provided in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the internal structure of the feature fusion module provided in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the processing logic of the dual sliding window spatial integration anomaly detection algorithm provided in an embodiment of the present invention;

[0044] Figure 5 This is a comparison chart of the effects of dynamic low-frequency cutoff frequency under different sea state conditions provided by embodiments of the present invention;

[0045] Figure 6 This is a comparison of the subject operating characteristic curves of the multi-method microleakage detection performance provided in the embodiments of the present invention;

[0046] Figure 7 This is an engineering performance verification diagram of the dual sliding window spatiotemporal integral anomaly detection algorithm provided in an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0048] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0049] like Figure 1 As shown, an embodiment of the present invention discloses a method for monitoring micro-leakage in subsea pipelines. The process architecture of the method consists of four processing stages, namely, a data field construction stage, a dual-flow depth feature extraction stage, a cross-dimensional feature fusion stage, and a dual sliding window spatiotemporal joint determination stage.

[0050] In the data field construction phase, distributed fiber optic sensors are laid along the seabed pipeline. These sensors acquire acoustic vibration signals along the pipeline using a Distributed Acoustic Sensing (DAS) system. Specifically, the optical fibers laid along the seabed pipeline serve as the sensing medium. The interrogator of the DAS injects coherent light pulses into the optical fibers, sensing external acoustic vibrations by detecting the phase change of the backscattered Rayleigh light. The acquired raw signals are discretized according to a set spatial resolution and temporal sampling rate. The spatial resolution determines the sampling point density along the pipeline length and is typically set based on the pipeline diameter, burial depth, and the spatial attenuation characteristics of the expected leakage signal. The temporal sampling rate, based on the Nyquist sampling theorem, needs to cover the main frequency components of the micro-leakage jet noise. After discretization, a two-dimensional spatiotemporal strain phase matrix, denoted as S(z, t), is constructed, where z represents the spatial position dimension along the pipeline length, and t represents the time dimension. Each row of this two-dimensional spatiotemporal strain phase matrix corresponds to the strain phase sequence of a spatial position channel at different times, and each column corresponds to the strain phase distribution of all spatial position channels at a given time. By upscaling a one-dimensional time-series signal into a two-dimensional spatiotemporal data field, the subsequent neural network can simultaneously capture the spatial extent and temporal evolution characteristics of the leaked signal.

[0051] In the dual-stream deep feature extraction stage, the above two-dimensional spatiotemporal strain phase matrix S(z, t) is input into two parallel branches of the dual-stream deep neural network, namely the spatial spectrum attention flow branch and the time series attention flow branch, and differentiated feature extraction processing is performed respectively.

[0052] The core design goal of the spatial-spectral attention flow branch is to suppress interference from low-frequency noise in the marine environment while preserving and enhancing the spatial frequency domain characteristics associated with micro-leakage. For example... Figure 2 As shown, this branch first performs a short-time Fourier transform (STFT) on the two-dimensional spatiotemporal strain phase matrix S(z,t) along the time axis to generate a three-dimensional space-frequency-time tensor, denoted as X. SF(z, f, τ) Where f represents the frequency dimension and τ represents the time window index. The short-time Fourier transform obtains the joint distribution information of the signal in the two-dimensional plane of time and frequency by segmenting the time signal, windowing it, and performing Fourier transform. Combined with the original spatial dimension, a complete three-dimensional data representation is formed.

[0053] Subsequently, external physical environment data of the sea area where the subsea pipeline is located is acquired. Based on this data, the corresponding low-frequency cutoff frequency is calculated, and a low-frequency attention mask is constructed. The external physical environment data comes from two sources: one is sea state remote sensing data obtained through satellite remote sensing or ocean buoys, including parameters such as wind speed, wave height, and tidal cycle; the other is the root mean square amplitude of background noise collected by distributed fiber optic sensors within a preset historical time period. Based on this physical environment data, the main energy distribution boundary of low-frequency noise under the current sea state conditions, i.e., the low-frequency cutoff frequency, is calculated using a preset mapping relationship or a lightweight prediction model. For example, in severe sea states with large waves, the frequency of pressure pulsations caused by waves may shift to a higher frequency band, at which point the low-frequency cutoff frequency is adjusted upwards accordingly; while in calm sea states, the low-frequency cutoff frequency remains at a lower level. Based on the calculated low-frequency cutoff frequency, a low-frequency attention mask M(f) is constructed. This mask takes a small value or zero value in the low-frequency noise band and a value of 1 in the effective signal band, thus forming a selective weighting of the frequency dimension of the three-dimensional tensor.

[0054] The low-frequency attention mask M(f) is coupled with the three-dimensional spatial-frequency-time tensor X. SF(z, f, τ) Perform element-wise multiplication to obtain the denoised tensor X' SF(z, f, τ) This operation enables dynamic suppression of low-frequency environmental noise, and the suppression boundary changes in real time with the physical environment, avoiding the defect of traditional fixed filters whose performance deteriorates sharply when sea state changes abruptly.

[0055] The denoised tensor is input into a feature extraction module containing a 2D residual network ResNet18 and a Spatial Attention Module (SAM). The 2D residual network employs a residual learning framework, progressively extracting high-level abstract features through the stacking of multiple residual blocks, while utilizing skip connections to alleviate the gradient vanishing problem in deep networks. The Spatial Attention Module dynamically weights the feature map along the channel dimension, focusing the network on spatial regions more critical for leak detection. After the above processing, the Spatial-Spectral Attention Flow branch finally outputs a spatial-frequency domain feature map, denoted as F. A .

[0056] The design of the time-series attention flow branch focuses on extracting the high-frequency temporal characteristics of micro-leakage signals. This branch first performs high-pass filtering on the two-dimensional spatiotemporal strain phase matrix S(z, t), setting an appropriate cutoff frequency to remove low-frequency noise, thus obtaining the one-dimensional high-frequency temporal signal corresponding to each spatial location channel. Micro-leakage is physically caused by fluid turbulence jetting due to the high pressure difference inside and outside the pipe. Its acoustic characteristics are continuous, non-periodic broadband noise, with energy mainly concentrated in the mid-to-high frequency range. Therefore, high-pass filtering can effectively preserve the leakage-related signal components. Subsequently, a gated recurrent unit (GRU) is used to perform time-series modeling on the aforementioned one-dimensional high-frequency temporal signal. The GRU effectively captures long-term dependencies in the sequence data through the coordinated control of update and reset gates. Compared to long short-term memory networks, it has fewer parameters and faster computation speed, making it more suitable for online monitoring scenarios. After hidden-state encoding by the GRU, a high-dimensional feature representation containing the temporal evolution law is obtained. Then, a multi-head attention network is used to calculate the temporal dependencies within the aforementioned one-dimensional high-frequency temporal signal. Multi-head attention mechanisms project query, key, and value vectors into multiple subspaces, calculate attention weights for each, and then concatenate them. This allows for the extraction of dependency structures within the signal from different perspectives, and it is particularly adept at capturing long-range temporal correlations. Through processing by the multi-head attention network, the stable and continuous temporal characteristics of micro-leakage signals are enhanced, while occasional high-frequency transient interferences are effectively distinguished due to their lack of sustained temporal correlation. The final output of this branch is a temporal feature vector, denoted as F. B .

[0057] like Figure 3 As shown, in the cross-dimensional feature fusion stage, the spatial-frequency domain feature map F output by the spatial-spectral attention flow branch is... A The temporal feature vector F output by the time series attention flow branch B The input is fed into the Attention-guided Feature Fusion Module (AFFM) for cross-dimensional feature fusion.

[0058] Specifically, firstly, the spatial-frequency domain feature map F A With time series feature vector F B Perform a tensor concatenation operation. Due to the differences in data dimensions and physical meanings between the two, direct concatenation requires splitting into two parallel processing paths: a first processing branch and a second processing branch.

[0059] In the first processing branch, channel scaling is performed through the first 1×1 convolutional layer. The 1×1 convolution linearly transforms the number of channels without changing the spatial dimension, thereby achieving feature dimensionality compression and information integration, while preserving global shallow feature information.

[0060] In the second processing branch, channel mapping is first performed through a second 1×1 convolutional layer, transforming the concatenated features into a channel space suitable for depth processing. Subsequently, the mapped features are input into a local depth feature extractor containing N reparameterized residual blocks (RepBlocks). During model training, the internal structure of each reparameterized residual block includes parallel 3×3 convolutional branches, 1×1 convolutional branches, and an identity mapping branch. Specifically, the 3×3 convolutional branches are responsible for capturing large-scale spatial neighborhood features, the 1×1 convolutional branches realize information interaction and integration between channels, and the identity mapping branch directly passes the input features to the output, ensuring smooth gradient backpropagation and the integrity of the original information. The output feature of this reparameterized residual block is the element-wise sum of the output features of the three parallel branches, i.e.:

[0061] ;

[0062] in, , , These represent the output features of the 3×3 convolution branch, the 1×1 convolution branch, and the identity mapping branch, respectively. This represents the final output feature of the reparameterized residual block.

[0063] The output features of the first processing branch are added and fused with the output features of the local depth feature extractor. This addition operation achieves effective complementarity between global shallow features and deep local features. The fused features are then passed sequentially through a Multi-Layer Perceptron (MLP) and a Normalized Exponential Function (Softmax) layer to finally output the probability of micro-leakage at each spatial location along the subsea pipeline, denoted as P(z). i ).

[0064] It is important to note that the aforementioned reparameterized residual block employs a multi-branch parallel structure during the training phase to ensure rich feature extraction and network training stability. However, during actual engineering inference deployment, based on the principle of structural reparameterization, the parallel 3×3 convolutions, 1×1 convolutions, and identity mapping branches are equivalently fused into a single 3×3 convolution kernel. This characteristic enables the present invention to maintain high detection accuracy while meeting the real-time requirements of online monitoring of subsea pipelines, achieving a balance between accuracy and efficiency.

[0065] like Figure 4 As shown, in the dual sliding window spatiotemporal joint decision-making stage, based on the aforementioned micro-leakage occurrence probability P(z) i The algorithm uses a dual sliding window spatial integral anomaly detection algorithm to perform spatiotemporal joint judgment to determine whether to trigger a micro-leakage warning and output the leak location.

[0066] Specifically, the dual-sliding-window spatial integral anomaly detection algorithm includes two levels of processing: a first spatial sliding window and a second temporal sliding window.

[0067] Within the first spatial sliding window, the width of the sliding window is adjusted based on the historical acoustic background noise variance of each spatial location channel, and the spatial integral value of the probability of micro-leakage occurring within the sliding window is calculated. The historical acoustic background noise variance is calculated based on pure noise data collected by distributed fiber optic sensors within a historical time window, reflecting the intensity of background noise fluctuations in the area. The dynamic adjustment rules for the width of the spatial sliding window are as follows: when the historical acoustic background noise variance of a certain spatial location channel is greater than a preset mean threshold, it indicates that the background noise in the area fluctuates drastically and has strong spatial heterogeneity. In this case, the width of the spatial sliding window is reduced to narrow the spatial field of view, focus potential signal sources, and avoid excessive accumulation of surrounding noise energy; when the historical acoustic background noise variance is less than or equal to the mean threshold, it indicates that the background in the area is relatively stable. In this case, the width of the spatial sliding window is increased to expand the spatial field of view and collect weaker diffused signals.

[0068] The calculated spatial integral value is denoted as I(z), which physically represents the cumulative amount of micro-leakage probability within the local spatial range at the current moment, effectively suppressing the random fluctuations in single-point probability determination.

[0069] Within the second time window, a time-sliding buffer queue with a preset frame count of k is established. This buffer queue stores the spatial integral value sequence of a local region within the most recent k time frames. It is determined whether the spatial integral value of this local region is greater than the set safety threshold for all k consecutive time frames in the time-sliding buffer queue. If so, a micro-leakage warning is executed and the corresponding leakage location is output; otherwise, it is determined to be transient interference or noise fluctuation, and no warning is triggered.

[0070] The design of the aforementioned dual sliding window mechanism fully considers the essential characteristic of micro-leakage as a continuous physical process. Once micro-leakage occurs, the resulting acoustic signal will persist in time and will not disappear within a single frame or a few frames; while transient interference may cause a high probability output in a single frame, it is difficult to maintain a high integral value over multiple consecutive time frames. Therefore, through the physical adaptation of the spatial sliding window and the continuity constraint of the temporal sliding window, the probability value of the model output is mapped to a spatiotemporal joint decision logic that conforms to engineering practice.

[0071] In summary, this invention achieves high sensitivity and high reliability monitoring of micro-leakage in subsea pipelines under complex marine environments through four steps: data field dimensionality enhancement, dual-stream asynchronous decoupling feature extraction, cross-dimensional reparameterization fusion, and dual sliding window spatiotemporal joint determination.

[0072] To verify the monitoring performance of the above-mentioned technical solution of the present invention, a continuous 72-hour simulation test experiment was conducted. Figure 5 Figure (a) illustrates the correlation between sea state changes and wave frequency drift during the test. The wave intensity in the test area fluctuated dramatically between force 0 and 8, with the corresponding wave frequency drifting back and forth between 0.05 Hz and 0.55 Hz. During this period, two significant sea state deterioration events occurred: Strong Wave Event I and Extremely Strong Wave Event II. In such a dynamically changing marine environment, Figure 5 Subgraph (b) compares the performance differences between the dynamic low-frequency cutoff frequency of the present invention and the fixed threshold scheme. The fixed threshold scheme sets the low-frequency cutoff frequency to a constant 80 Hz, maintaining the same value regardless of changes in sea state; while the dynamic low-frequency cutoff frequency of the present invention, calculated in real time based on external physical environment data, can closely track the changing trend of the theoretically optimal cutoff frequency. During calm periods with weak winds and waves, it automatically lowers to around 100 Hz, and during strong wind and wave events I and extremely strong wind and wave events II, it responds rapidly and rises, reaching a maximum of the preset upper limit of 350 Hz, achieving real-time dynamic alignment of the noise spectrum boundary. Figure 5 Subplot (c) further provides a quantitative comparison of the signal-to-noise ratio (SNR) improvement between the two schemes. The fixed-threshold scheme, shown by the gray dashed line, shows a significant decrease in SNR improvement as sea state deteriorates, generally remaining in the low range of 5 dB to 18 dB. In contrast, the dynamic scheme of this invention, shown by the green solid line, consistently maintains a high SNR improvement of 15 dB to 28 dB. The shaded area clearly indicates the all-time performance gain of the dynamic scheme compared to the fixed scheme, with a maximum gain of 18.9 dB. These results fully demonstrate that the dynamic masking mechanism, by introducing physical environmental information such as sea state remote sensing data and fiber optic background noise, can effectively overcome the noise spectrum drift problem caused by the time-varying characteristics of the marine environment, providing high-quality noise reduction input for the spatial-spectral attention flow branch and greatly improving the signal preprocessing quality under severe sea conditions.

[0073] Figure 6 This paper presents a comparison of the Receiver Operating Characteristic Curve (ROC) curves of various microleakage detection methods and a quantitative analysis of key indicators. The tests were conducted using the same dataset and consistent annotation conditions. The methods compared included the proposed scheme, a single-stream spatial-spectral network retaining only the spatial-spectral attention flow branch, a single-stream temporal network retaining only the time-series attention flow branch, and a traditional method employing fixed threshold joint frequency domain filtering. Figure 6 The ROC curve on the left uses the false positive rate (FPR) as the x-axis and the true positive rate (TPR) as the y-axis. The area under the curve (AUC) comprehensively represents the overall discrimination ability of the detector. The ROC curve of the complete scheme of this invention is located at the top left, with the largest area enclosed by the curve and the coordinate axis, and the AUC reaches 0.979. The AUC of the single-stream spatial spectrum network is 0.938; the AUC of the single-stream temporal network is 0.910; and the AUC of the traditional method is only 0.627, close to the diagonal of random guessing. At the low false alarm rate operating point, when the false positive rate is strictly controlled at 5%, the true positive rate of the complete scheme of this invention is as high as 90%, while the single-stream spatial spectrum network drops to 54.1%, the single-stream temporal network further drops to 40.5%, and the traditional method can only reach 16.7%. Figure 6 The bar chart on the right further quantifies and compares the results from three dimensions: AUC, true positive rate under low false positive rate conditions, and false positive rate under high true positive rate conditions. When a high detection level of 90% is required for the true positive rate, the false positive rate of the complete scheme of this invention is only 5.0%, while the false positive rates of single-stream spatial spectrum network, single-stream temporal network, and traditional methods deteriorate to 14.1%, 20.6%, and 83.8%, respectively. The above data clearly show that the detection performance of this invention, through the synergistic effect of the dual-stream asynchronous decoupling structure and the cross-dimensional reparameterization fusion mechanism, is significantly better than the simple superposition of the performance of two single-stream branches, effectively solving the technical problem that weak leakage features are easily submerged by marine background noise under low signal-to-noise ratio conditions.

[0074] Figure 7 The effectiveness of the dual sliding window spatiotemporal integral anomaly detection algorithm in reducing false alarms in real-world engineering scenarios was verified. Figure 7Subgraph (a) shows the raw micro-leakage instantaneous probability P(z, t) output by the traditional deep network and the alarm results using a fixed threshold single-point decision. In this 150-frame test sequence, the traditional single-point decision strategy sets a fixed alarm threshold of 0.75. Although it can capture real leakage events in the range of frames 120 to 145, it causes instantaneous spikes in the probability value near frame 25 due to strong transient interference and near frame 65 due to general transient disturbances, thus triggering false alarms. In addition, the fluctuation spikes in the probability value at the beginning and end of the real leakage event also generate additional alarm noise. Figure 7 Subgraph (b) shows the variation curve of the spatial integral value I(z) after the first spatial sliding window integration process, as well as the judgment logic of the second temporal sliding window. The spatial integral value is smoother and more stable than the original probability curve, and the set safety threshold λ is 0.65. Around frame 25, although the spatial integral value once exceeded the safety threshold, its duration was only 1 to 2 frames, which did not meet the preset time buffer frame requirement, so it was judged as transient interference and correctly filtered. After frame 120, the spatial integral value continuously and stably exceeded the safety threshold and remained at least 4 frames, satisfying the spatiotemporal persistence condition. Figure 7 Subplot (c) visually displays the comparison of the final early warning execution results in a time-axis distribution. Traditional single-point threshold judgment results are marked in red, showing multiple brief alarm segments scattered throughout the monitoring period. These segments, along with... Figure 7 The transient probability spikes in (a) correspond one-to-one and are typical false alarms; while the spatiotemporal joint judgment result of the dual sliding window of this invention is marked in green, outputting a single and continuous alarm interval only during the actual leakage duration from frame 120 to frame 145, successfully filtering out all transient interference and achieving zero false alarm output. The above experimental results fully verify the engineering effectiveness of the dual sliding window mechanism in mapping model probabilities to conform to the physical spatiotemporal joint judgment logic.

[0075] In summary, this invention achieves highly sensitive detection and highly reliable alarm of micro-leakage in subsea pipelines in complex and ever-changing marine environments by constructing a two-dimensional spatiotemporal data field, extracting asynchronous features from two streams, fusing cross-dimensional reparameterization, and combining dual sliding window spatiotemporal joint judgment with deep embedding of dynamic physical environment information. This provides an effective technical solution for the safety monitoring of subsea pipelines.

[0076] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for monitoring micro-leakage in subsea pipelines, characterized in that: The method includes: Step S1: Acquire the acoustic vibration signal from the distributed optical fiber sensor installed along the seabed pipeline, discretize it according to the set spatial resolution and time sampling rate, and construct a two-dimensional spatiotemporal strain phase matrix. Step S2: The two-dimensional spatiotemporal strain phase matrix is ​​input into the spatial spectrum attention flow branch and the time series attention flow branch of the dual-flow deep neural network, respectively; wherein, the spatial spectrum attention flow branch performs spatiotemporal frequency domain transformation and frequency masking steps, and outputs a spatial-frequency domain feature map after suppressing low-frequency environmental noise; the time series attention flow branch sequentially performs high-pass filtering and time series modeling steps to extract the time series features of micro-leakage sound and outputs a time series feature vector; Step S3: Input the spatial-frequency domain feature map and the temporal feature vector into the feature fusion module to perform a cross-dimensional feature fusion step, and output the probability of micro-leakage at each spatial location along the submarine pipeline. Step S4: Based on the probability of micro-leakage occurrence, a spatiotemporal joint determination is performed using a dual sliding window spatial integral anomaly detection algorithm. When the preset spatiotemporal persistence condition is met, a micro-leakage warning is triggered and the leak location is output.

2. The method for monitoring micro-leakage in a subsea pipeline according to claim 1, characterized in that: In step S2, the steps performed by the spatial spectrum attention flow branch specifically include: A short-time Fourier transform is performed on the two-dimensional spatiotemporal strain phase matrix along the time axis to generate a three-dimensional space-frequency-time tensor. The external physical environment data of the sea area where the subsea pipeline is located is obtained, the corresponding low-frequency cutoff frequency is calculated based on the external physical environment data, and a low-frequency attention mask is constructed. The low-frequency attention mask is multiplied element-wise with the spatial-frequency-time three-dimensional tensor to obtain the noise reduction tensor; The noise reduction tensor is input into a feature extraction module that includes a two-dimensional residual network and a spatial attention network, and the spatial-frequency domain feature map is output.

3. The method for monitoring micro-leakage in a subsea pipeline according to claim 2, characterized in that: The external physical environment data includes sea state remote sensing data or the root mean square amplitude of the background noise of the distributed optical fiber sensor within a preset historical time period.

4. The method for monitoring micro-leakage in a subsea pipeline according to claim 1, characterized in that: In step S2, the steps performed by the time-series attention flow branch include: The two-dimensional spatiotemporal strain phase matrix is ​​subjected to high-pass filtering to remove low-frequency noise bands and obtain one-dimensional high-frequency time-series signals corresponding to each spatial location channel. A gated recurrent network is used to perform time-series modeling on the one-dimensional high-frequency time-series signal. The time dependencies within the one-dimensional high-frequency time-series signal are calculated using a multi-head attention network, and the time-series feature vector is output.

5. The method for monitoring micro-leakage in a subsea pipeline according to claim 1, characterized in that: In step S3, the steps performed by the feature fusion module specifically include: The spatial-frequency domain feature map and the temporal feature vector are tensor-concatenated and then split into a first processing branch and a second processing branch. In the first processing branch, channel scaling is performed through a first 1×1 convolutional layer; In the second processing branch, after channel mapping is performed through the second 1×1 convolutional layer, the features are input to a local depth feature extractor containing N reparameterized residual blocks. The output features of the first processing branch are added and fused with the output features of the local deep feature extractor, and then the probability of micro-leakage is output through a multilayer perceptron and a normalized exponential function layer.

6. The method for monitoring micro-leakage in a subsea pipeline according to claim 5, characterized in that: During the model training phase, the internal structure of each reparameterized residual block includes parallel 3×3 convolution branches, 1×1 convolution branches, and identity mapping branches. The output feature of the reparameterized residual block is the element-wise sum of the output features of the three parallel branches.

7. The method for monitoring micro-leakage in a subsea pipeline according to claim 1, characterized in that: In step S4, the step of performing spatiotemporal joint determination using the dual sliding window spatial integral anomaly detection algorithm specifically includes: Within the first spatial sliding window, the width of the spatial sliding window is adjusted based on the historical acoustic background noise variance of each spatial location channel, and the spatial integral value of the probability of the micro-leakage occurring within the spatial sliding window is calculated. Within the second time window, a time sliding buffer queue with a preset frame count of k is established; If the spatial integral value of a certain local area is greater than the set safety threshold for k consecutive time frames in the time sliding buffer queue, then the micro-leakage warning is executed.

8. The method for monitoring micro-leakage in a subsea pipeline according to claim 7, characterized in that: The step of adjusting the width of the spatial sliding window based on the historical acoustic background noise variance of each spatial location channel specifically includes: When the historical acoustic background noise variance of a certain spatial location channel is greater than the mean threshold, the width of the spatial sliding window is reduced. When the variance of the historical acoustic background noise is less than or equal to the mean threshold, the width of the spatial sliding window is increased.

Images