A method, apparatus, equipment and medium for detecting pipeline damage

Abstract

This invention discloses a method, apparatus, equipment, and medium for detecting pipeline damage, relating to the field of pipeline safety monitoring technology. The method includes: acquiring images of the inner wall of the pipeline using an image acquisition device inside the pipeline; identifying candidate defect categories and corresponding visual defect locations using a defect hierarchy classification model; acquiring signals using multi-channel acoustic emission sensors arranged on the outer wall of the pipeline, converting them into continuous wavelet transform acoustic maps, and inputting them into a temporal convolutional network to extract acoustic anomaly features and their corresponding time windows; acquiring vibration signals along the pipeline collected by distributed fiber optic sensors, identifying disturbance events and fiber optic positioning results; mapping the visual defect locations, acoustic anomaly time windows, and fiber optic positioning results to a unified pipe section coordinate system to perform cross-modal consistency determination; and, when conditions are met, outputting the final pipeline damage level based on the aforementioned features.

Description

Technical Field

[0001] This invention relates to the field of pipeline safety monitoring technology, and in particular to a pipeline damage detection method, device, equipment, and medium. Background Technology

[0002] With the acceleration of industrialization and the continuous improvement of urban infrastructure, long-distance oil and gas pipelines and urban underground water supply and drainage networks play a fundamental role in economic operations. However, due to factors such as increased service life, changes in the geological environment, and third-party construction work, pipelines face the risks of corrosion perforation, rupture and leakage, and even breakage. To ensure the safe operation of the pipeline network, regular damage and leakage detection has become a routine maintenance procedure.

[0003] Existing pipeline inspection technologies largely rely on single-modal sensing methods. The first common approach involves using closed-circuit television or pipeline endoscopy robots to acquire images of the pipe interior. This method relies on human visual inspection or basic image recognition algorithms to locate cracks or holes. However, the interior of pipelines is often obscured by large amounts of dirt or water accumulation, making it prone to missed detections when relying solely on visual information. Furthermore, image-based methods primarily reflect existing physical deformations and are insufficient for early warning of stress concentration zones impending rupture or external third-party construction disturbances. The second common approach utilizes acoustic emission sensors or distributed fiber optic sensors deployed externally to the pipeline to identify leaks and damage by capturing abnormal sound waves or vibration signals. However, in real-world operating environments, background noise from fluid transport and environmental interference such as traffic vibrations can easily overlap with actual damage signals. Detection methods based on a single acoustic or vibration mode typically employ fixed threshold alarm mechanisms, leading to a high false alarm rate under complex operating conditions.

[0004] In summary, most existing detection technologies operate independently, with data collected by different modal sensors being fragmented in both time and space. Faced with complex interference environments, existing systems struggle to integrate multi-dimensional information for cross-validation, limiting the accuracy of damage identification and leading to false alarms or missed detections. This increases communication costs and resource investment in pipeline maintenance. Therefore, the industry needs a pipeline inspection method that can integrate multi-source data and address complex interference issues. Summary of the Invention

[0005] This invention provides a method, apparatus, equipment, and medium for detecting pipeline damage, which solves the problems of single-modal detection being susceptible to environmental interference, high false alarm rate, and fragmented multi-source data in existing technologies.

[0006] The first aspect of this invention provides a method for detecting pipeline damage, the method comprising: Images of the inner wall of the pipe are acquired by an image acquisition device inside the pipe, and a pre-set defect hierarchy classification model is used to identify the inner wall images of the pipe to obtain candidate defect categories and the visual defect locations corresponding to the candidate defect categories. Acoustic emission signals during operation are collected by multi-channel acoustic emission sensors arranged on the outer wall of the pipe. The acoustic emission signals are converted into continuous wavelet transform acoustic maps, and the continuous wavelet transform acoustic maps are input into a pre-trained temporal convolutional network for recognition, extracting acoustic anomaly features and their corresponding acoustic anomaly time windows. The vibration signals along the pipeline are collected by distributed optical fiber sensors laid along the pipeline. The spatiotemporal features of the vibration signals along the pipeline are extracted to identify the disturbance events and their corresponding optical fiber positioning results. The location of visual defects, the time window of acoustic anomalies, and the fiber optic positioning results are mapped to a preset unified pipe section coordinate system. Cross-modal consistency judgment is performed, and when the location of visual defects, the time window of acoustic anomalies, and the fiber optic positioning results meet the preset spatiotemporal consistency conditions, the final pipe damage level is output by combining the candidate defect category, acoustic anomaly characteristics, and disturbance events.

[0007] In one optional implementation, a pre-defined defect hierarchy classification model is used to identify the inner wall image of the pipe, obtaining candidate defect categories and the corresponding visual defect locations, including: Visual features of the pipe inner wall image are extracted, and candidate regions in the pipe inner wall image are located through a candidate box detection network; The candidate regions are input into the defect hierarchical classification model, and the first-level classifier classifies the candidate regions into normal or abnormal states. The second-level classifier, trained by the hierarchical loss function, further subdivides the candidate regions in the abnormal state into at least one of cracks, holes, breaches, and sediment occlusion to obtain candidate defect categories. The hierarchical loss function introduces a category weight matrix calculated based on the historical defect dataset to compensate for the classification bias of a small number of defect types. The odometer data is obtained when the image acquisition device acquires images of the inner wall of the pipe. The odometer data and the pixel coordinates of the candidate region in the image of the inner wall of the pipe are combined to calculate the three-dimensional physical coordinates of the physical region of the inner wall of the pipe corresponding to the candidate region in the real pipe. The three-dimensional physical coordinates are used as the location of the visual defect.

[0008] In one optional implementation, the acoustic emission signal is converted into a continuous wavelet transform acoustic map, and the continuous wavelet transform acoustic map is input into a pre-trained temporal convolutional network for recognition, extracting acoustic anomaly features and their corresponding acoustic anomaly time windows, including: The acoustic emission signal is continuously sampled using a sliding time window to obtain the time-domain amplitude signal within the current sliding time window. A Gaussian filter is used to suppress background noise in the time-domain amplitude signal. The noise-reduced time-domain amplitude signal is then subjected to continuous wavelet transform to generate a continuous wavelet transform acoustic map containing the distribution of energy of multiple channels with time and frequency. The continuous wavelet transform acoustic map is sequentially input into a temporal convolutional network containing spatial convolutional layers and long short-term memory networks according to the time series. The system determines whether there are leakage features within the current sliding time window and uses the leakage features as acoustic anomalies when they are present. If a leakage feature exists within the current sliding time window, and the duration of the leakage feature exceeds a preset abnormal duration threshold within multiple consecutive sliding time windows, an abnormal state is confirmed, and the start and end time periods corresponding to the multiple consecutive sliding time windows are confirmed as acoustic abnormal time windows.

[0009] In one alternative implementation, after acquiring acoustic emission signals during operation using multi-channel acoustic emission sensors arranged on the outer wall of the pipe, and inputting the continuous wavelet transform acoustic map into a pre-trained temporal convolutional network for recognition, the method further includes: High-dimensional spatial features corresponding to each acquisition channel are extracted from the continuous wavelet transform acoustic map. Feature splicing and dimensionality reduction operations are performed on the high-dimensional spatial features of each acquisition channel to obtain the channel fusion feature vector. The channel fusion feature vector is input into a preset leakage aperture regression network to calculate the leakage aperture estimate corresponding to the leakage feature. The estimated leakage aperture is mapped to the initial leakage level parameter, and the leakage level parameter is used as an auxiliary evaluation index to characterize the severity of acoustic anomalies.

[0010] In one optional implementation, vibration signals along the pipeline are acquired from distributed optical fiber sensors laid along the pipeline. Spatiotemporal features of the vibration signals are extracted to identify disturbance events and their corresponding optical fiber positioning results, including: It receives the optical phase change signal returned by a distributed optical fiber sensor based on the coherent Rayleigh scattering principle and uses the optical phase change signal as a vibration signal along the line; The energy distribution features of the vibration signal along the line in the spatial dimension and the frequency distribution features in the time dimension are extracted, and spatiotemporal features are spliced ​​to obtain a multidimensional disturbance feature set; The multidimensional disturbance feature set is input into a pre-set event recognition network. The support vector machine classification layer is used to filter environmental noise signals and identify and distinguish third-party construction disturbance events from real damage precursor events. The distinguished events are taken as disturbance events. Based on the fiber optic scattering distance point corresponding to the identified disturbance event, the longitudinal physical distance of the disturbance event along the axial direction of the long-distance pipeline is determined, and the longitudinal physical distance is used as the fiber optic positioning result.

[0011] In one optional implementation, the visual defect location, acoustic anomaly time window, and fiber optic positioning results are mapped to a preset unified pipe segment coordinate system. Cross-modal consistency determination is performed, and the final pipe damage level is output when the visual defect location, acoustic anomaly time window, and fiber optic positioning results meet preset spatiotemporal consistency conditions. This level includes: A three-dimensional pipe network model is established with the fixed starting point of the pipeline as the absolute origin, and the three-dimensional pipe network model is used as a unified pipe segment coordinate system. Based on the preset spatial tolerance parameters, the location of visual defects is mapped to the first spatial coordinate interval in the three-dimensional pipeline network model, and the fiber optic positioning result is mapped to the second spatial coordinate interval in the three-dimensional pipeline network model. The image acquisition device is positioned inside the pipeline when the acoustic anomaly time window occurs, and the third spatial coordinate interval corresponding to the acoustic anomaly time window is calculated based on the position information. Calculate the spatial intersection-union ratio of the first, second, and third spatial coordinate intervals in the three-dimensional pipeline network model, extract the event time intervals where disturbance events occur, and obtain the time overlap between the acoustic anomaly time window and the event time interval. When the spatial intersection-to-union ratio is greater than the preset spatial consistency threshold and the temporal overlap is greater than the preset temporal consistency threshold, the location of the visual defect, the time window of the acoustic anomaly, and the fiber optic positioning result are determined to meet the spatiotemporal consistency condition.

[0012] In one alternative implementation, the final pipe damage level is output by combining candidate defect categories, acoustic anomaly features, and disturbance events, including: When determining whether the spatiotemporal consistency condition is met, the corresponding visual risk value, acoustic risk value, and fiber optic risk value are determined based on the candidate defect category, acoustic anomaly characteristics, and disturbance event, respectively. The visual risk value is assigned a first confidence weight, the acoustic risk value is assigned a second confidence weight, and the fiber optic risk value is assigned a third confidence weight. Based on the first confidence weight, the second confidence weight, and the third confidence weight, the visual risk value, acoustic risk value, and fiber optic risk value are weighted and summed to obtain the comprehensive damage risk value. Based on the preset range of the comprehensive damage risk value, the final pipeline damage level is determined, and an automatic alarm command with a maintenance priority label is generated based on the final pipeline damage level.

[0013] A second aspect of this invention provides a pipeline damage detection device, the device comprising: The image recognition module is used to acquire images of the inner wall of the pipe through an image acquisition device inside the pipe, and to identify the inner wall images of the pipe using a pre-set defect hierarchy classification model to obtain candidate defect categories and the visual defect locations corresponding to the candidate defect categories. The acoustic feature extraction module is used to collect acoustic emission signals during operation through multi-channel acoustic emission sensors arranged on the outer wall of the pipe, convert the acoustic emission signals into continuous wavelet transform acoustic maps, and input the continuous wavelet transform acoustic maps into a pre-trained temporal convolutional network for recognition, extracting acoustic anomaly features and their corresponding acoustic anomaly time windows. The fiber optic positioning module is used to acquire vibration signals along the pipeline from distributed fiber optic sensors laid along the pipeline, extract spatiotemporal features from the vibration signals along the pipeline, and identify disturbance events and their corresponding fiber optic positioning results. The consistency determination and output module is used to map the visual defect location, acoustic anomaly time window and fiber optic positioning result to a preset unified pipe section coordinate system, perform cross-modal consistency determination, and when the visual defect location, acoustic anomaly time window and fiber optic positioning result meet the preset spatiotemporal consistency conditions, it combines candidate defect category, acoustic anomaly characteristics and disturbance event to output the final pipe damage level.

[0014] A third aspect of this invention provides an electronic device, comprising: a memory and a processor; Memory is used to store computer programs; The processor is used to execute computer programs to implement the aforementioned pipeline damage detection method.

[0015] The fourth aspect of this invention provides an electronic device and a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned pipeline damage detection method.

[0016] Beneficial Effects: The pipeline damage detection method, apparatus, equipment, and medium provided in this invention integrate multi-source data acquired by image acquisition equipment, multi-channel acoustic emission sensors, and distributed fiber optic sensors to construct a unified cross-modal detection architecture, improving upon the limitations of existing single detection methods, which suffer from high false alarm frequencies and susceptibility to environmental interference. A defect hierarchy classification model is used to identify and extract visual defect locations from inner wall images. Simultaneously, acoustic emission signals are converted into continuous wavelet transform acoustic maps, and acoustic anomaly time windows are extracted using a temporal convolutional network. These are then combined with fiber optic vibration signals to identify and locate disturbance events. By mapping the aforementioned multi-source heterogeneous results to a unified pipe section coordinate system and performing cross-modal consistency determination in the spatiotemporal dimension, cross-validation can be performed between visual spatial features, acoustic time windows, and fiber optic positioning results. This mechanism can filter misleading information caused by occasional environmental noise or unidirectional visual occlusion, improving the accuracy of identifying real damage events under complex operating conditions. Furthermore, by combining multi-dimensional risk indicators to output the final pipeline damage level, an objective assessment of the pipeline's health status is achieved, providing objective data reference for subsequent resource scheduling and maintenance planning. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart of the steps of a pipeline damage detection method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the functional modules of a pipeline damage detection device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. 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.

[0019] The present invention will be further described below with reference to the accompanying drawings.

[0020] Reference Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0021] like Figure 1As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (Wi-Fi). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and computer programs.

[0024] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the pipeline damage detection device stored in the memory 1005 through the processor 1001 and executes the pipeline damage detection method provided in the embodiment of the present invention.

[0025] The first aspect of this invention proposes a pipeline damage detection method. The execution entity of this method can be a pipeline damage detection device with data processing and analysis capabilities, a local workstation, an edge computing node, or a cloud server. (See attached figure.) Figure 2 The method provided in this embodiment of the invention includes the following steps: S101, the image of the inner wall of the pipe is acquired by the image acquisition device inside the pipe, and the image of the inner wall of the pipe is identified by the preset defect hierarchy classification model to obtain the candidate defect category and the visual defect location corresponding to the candidate defect category.

[0026] In specific industrial applications, long-distance oil and gas pipelines or urban underground water supply and drainage networks are buried deep underground for extended periods. Continuous erosion by the internal media, chemical corrosion, and external geological subsidence can all cause physical damage to the pipeline's inner wall. Conventional external manual inspections are insufficient to directly observe structural damage within the pipeline. Therefore, it is essential to rely on image acquisition equipment deployed inside the pipeline to obtain firsthand visual data. This image acquisition equipment is typically mounted on autonomous pipeline robots or closed-circuit television (CCTV) tractors. As the equipment travels axially inside the pipeline, its built-in high-definition camera module continuously captures images of the pipeline's inner wall at a preset fixed frame rate, thus obtaining a continuous sequence of images. Because the interior of a pipeline is a closed, confined space where natural light cannot reach, the image acquisition equipment is usually equipped with an adaptively adjustable ring-shaped light-emitting diode (LED) light source array for active illumination. After receiving the raw images of the pipeline's inner wall, image preprocessing algorithms are typically invoked to eliminate light spots or shadows caused by uneven illumination. For example, the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm can be used to perform local contrast enhancement and illumination compensation processing on the original pipe inner wall image, thereby highlighting the fine texture of the pipe wall, the direction of cracks, and the outline of surface deposits, laying a high-quality data foundation for subsequent deep learning model recognition.

[0027] In one optional implementation, a pre-defined defect hierarchy classification model is used to identify the inner wall image of the pipe, obtaining candidate defect categories and the corresponding visual defect locations, including: Visual features of the pipe inner wall image are extracted, and candidate regions in the pipe inner wall image are located through a candidate box detection network; The candidate regions are input into the defect hierarchical classification model, and the first-level classifier classifies the candidate regions into normal or abnormal states. The second-level classifier, trained by the hierarchical loss function, further subdivides the candidate regions in the abnormal state into at least one of cracks, holes, breaches, and sediment occlusion to obtain candidate defect categories. The hierarchical loss function introduces a category weight matrix calculated based on the historical defect dataset to compensate for the classification bias of a small number of defect types. The odometer data is obtained when the image acquisition device acquires images of the inner wall of the pipe. The odometer data and the pixel coordinates of the candidate region in the image of the inner wall of the pipe are combined to calculate the three-dimensional physical coordinates of the physical region of the inner wall of the pipe corresponding to the candidate region in the real pipe. The three-dimensional physical coordinates are used as the location of the visual defect.

[0028] Specifically, the above steps can be implemented through the following sub-steps S1011 to S1014: S1011. Extract visual features from the pipe inner wall image and locate candidate regions in the pipe inner wall image using a candidate box detection network.

[0029] After image preprocessing, a deep convolutional neural network (CNN) is used as the backbone feature extractor to extract multi-scale visual features from the pipe inner wall image. The backbone feature extractor can employ a ResNet network with deep residual connections or a multi-stage local network structure. After the image is input into the backbone feature extractor, it undergoes multi-layer convolution, batch normalization, and non-linear activation function mapping, outputting a set of feature maps that fuse high-level semantic abstraction information with low-level spatial texture details. These feature map sets are then input into a region proposal network (RPN). The RPN generates preset anchor boxes of various scales and aspect ratios by sliding them across each spatial location of the feature maps, and calculates the probability score of the presence of pipe defect features within each preset anchor box, as well as the positional offset of the bounding box. Furthermore, by combining the Non-Maximum Suppression (NMS) algorithm, redundant bounding boxes with spatial overlap exceeding a set threshold are eliminated in descending order of probability scores, thereby accurately locating the candidate regions in the pipe inner wall image that are most likely to have structural anomalies or foreign object occlusion.

[0030] S1012. Input the candidate region into the defect hierarchical classification model, and use the first-level classifier to classify the candidate region into normal or abnormal states.

[0031] After acquiring the located candidate regions, state classification begins. In the massive amount of actual detection image data, healthy, normal pipe wall images usually account for the vast majority. Directly initiating complex multi-classification calculations for all candidate regions would lead to a significant waste of computing power and reduce the overall real-time performance of the detection. Therefore, this embodiment adopts a defect hierarchical classification model from shallow to deep. First, a first-level classifier with fewer parameters and extremely low computational overhead is used to quickly filter the features of the cropped candidate regions. The first-level classifier performs only a strict binary classification task, that is, by extracting shallow features such as color distribution, grayscale gradient changes, and local binary patterns within the candidate region, it determines whether the candidate region conforms to the continuous and smooth features of a healthy pipe wall. If the normal probability output by the first-level classifier is greater than a set threshold, it is determined to be in a normal state, and the candidate region is directly discarded, skipping subsequent processing; only when the abnormal probability output by the first-level classifier exceeds a set abnormal trigger threshold is the candidate region determined to be in an abnormal state, and its feature vector is passed to the next level for refined category classification.

[0032] S1013. The candidate regions in the abnormal state are further subdivided into at least one of cracks, holes, breaks and sediment occlusion by the second-level classifier generated by training through the hierarchical loss function to obtain candidate defect categories. The hierarchical loss function introduces a category weight matrix calculated based on the historical defect dataset to compensate for the classification bias of the few sample damage types.

[0033] After confirming that the candidate region is in an abnormal state, it is necessary to determine its specific physical damage type. In real pipeline maintenance databases, the probability of different types of defects exhibits a typical long-tail distribution. For example, minor surface cracks or localized silt deposits are very common, accumulating a large number of training samples; while severe through-hole breaches or large-area collapses are rare events, with extremely few samples. To address this problem of extremely imbalanced samples, the second-level classifier uses a specially constructed hierarchical loss function in the early offline training phase. The hierarchical loss function dynamically introduces a class weight matrix based on the traditional cross-entropy loss calculation. The class weight matrix is ​​calculated based on the inverse proportion of each class in the historical defect dataset or the number of effective samples, assigning higher penalty weight values ​​to rare damage types with extremely low occurrence frequencies. During the backpropagation optimization process of the network model, this weight allocation mechanism forces the neural network to pay more attention to the decision boundary features of minority class samples when updating parameters, effectively compensating for the classification bias of damage types with few samples. In the actual online detection phase, through the forward propagation process, the second-level classifier further subdivides the candidate regions in the abnormal state into specific categories such as cracks, holes, breaches, and sediment occlusion, and outputs the category with the highest classification confidence as the candidate defect category.

[0034] S1014. Obtain the odometer data when the image acquisition device acquires the image of the inner wall of the pipe. Combine the odometer data and the pixel coordinates of the candidate area in the image of the inner wall of the pipe to calculate the three-dimensional physical coordinates of the physical area of ​​the inner wall of the pipe corresponding to the candidate area in the real pipe. Use the three-dimensional physical coordinates as the location of the visual defect.

[0035] To map pixel-level defects found in a 2D image plane to a real-world 3D physical pipe, a rigid body coordinate transformation across space is required. During pipe inspection by the image acquisition device, its onboard encoder, odometer, and inertial measurement unit (IMU) operate synchronously. At the same moment each frame of the pipe's inner wall image is captured, the current odometer data is recorded. This data includes the cumulative axial travel distance of the image acquisition device from the pipe's starting manhole, as well as the device's current attitude angles, including pitch, yaw, and roll. Subsequently, using a pre-calibrated camera intrinsic parameter matrix, including the camera's focal length, principal point coordinates, and distortion coefficients, the pixel coordinates in the 2D image coordinate system are converted to 3D coordinates in the camera's physical coordinate system. Next, using the extrinsic parameter matrix and the real-time 3D pose parameters provided by the odometer data, a homogeneous coordinate transformation matrix is ​​constructed, mapping the coordinates in the camera's physical coordinate system to a 3D spatial coordinate system with the pipe's starting point as the absolute global origin. Through the above coordinate transformation calculation, the three-dimensional physical coordinates of the physical area of ​​the inner wall of the pipe corresponding to the candidate area in the real pipe can be accurately calculated, and the three-dimensional physical coordinates can be used as the basic visual defect location required for subsequent multi-source data fusion.

[0036] S102: Acoustic emission signals under operating conditions are collected by multi-channel acoustic emission sensors arranged on the outer wall of the pipe. The acoustic emission signals are converted into continuous wavelet transform acoustic maps. The continuous wavelet transform acoustic maps are input into a pre-trained temporal convolutional network for recognition, and acoustic anomaly features and their corresponding acoustic anomaly time windows are extracted.

[0037] Single-dimensional internal visual image detection has inherent technical limitations. On the one hand, complex water flow, floating objects, or large areas of sediment deposits inside pipes can easily obstruct the camera's view, causing cracks hidden beneath these obstructions to go undetected. On the other hand, visual detection can only reflect physical deformations that have already occurred and are visible on the inner wall surface, lacking the ability to detect micro-cracks expanding inside the pipe wall, areas of structural stress concentration, or leaks that have penetrated the pipe wall and are ejecting fluid outwards. Therefore, this embodiment introduces an acoustic emission sensing mode as an orthogonal verification dimension. Acoustic emission refers to the physical phenomenon of transient elastic stress waves generated when the internal strain energy of a pipe's metal or polymer material is rapidly released. When a high-pressure fluid leaks or microscopic tears occur in a pipe, it radiates high-frequency acoustic elastic wave signals in all directions. By deploying a multi-channel acoustic emission sensor array at equal intervals by binding or welding it to the outer wall of the pipe, broadband acoustic emission signals of the pipe under pressurized operation can be continuously acquired at extremely high sampling frequencies, typically in the hundreds of kilohertz to megahertz range.

[0038] In one optional implementation, the acoustic emission signal is converted into a continuous wavelet transform acoustic map, and the continuous wavelet transform acoustic map is input into a pre-trained temporal convolutional network for recognition, extracting acoustic anomaly features and their corresponding acoustic anomaly time windows, including: The acoustic emission signal is continuously sampled using a sliding time window to obtain the time-domain amplitude signal within the current sliding time window. A Gaussian filter is used to suppress background noise in the time-domain amplitude signal. The noise-reduced time-domain amplitude signal is then subjected to continuous wavelet transform to generate a continuous wavelet transform acoustic map containing the distribution of energy of multiple channels with time and frequency. The continuous wavelet transform acoustic map is sequentially input into a temporal convolutional network containing spatial convolutional layers and long short-term memory networks according to the time series. The system determines whether there are leakage features within the current sliding time window and uses the leakage features as acoustic anomalies when they are present. If a leakage feature exists within the current sliding time window, and the duration of the leakage feature exceeds a preset abnormal duration threshold within multiple consecutive sliding time windows, an abnormal state is confirmed, and the start and end time periods corresponding to the multiple consecutive sliding time windows are confirmed as acoustic abnormal time windows.

[0039] Specifically, the above steps can be implemented through the following sub-steps S1021 to S1024: S1021. The acoustic emission signal is continuously sampled using a sliding time window to obtain the time-domain amplitude signal within the current sliding time window.

[0040] The raw data output by the multi-channel acoustic emission sensor is a one-dimensional time-series signal that extends infinitely over time, and due to the extremely high sampling frequency, the data volume exhibits a massive growth characteristic. To meet the processing requirements of computing devices for fixed-length data and to accurately capture anomalous waveforms with transient burst characteristics, a fixed-length sliding time window is set on the time axis. The length of the sliding time window can be preset according to the sound velocity of the pipe material and the sensor spacing, for example, set to 500 milliseconds. The continuously flowing acoustic emission data stream is sliced ​​using a set time step and overlap rate, such as 50%. Through this overlapping sliding window slicing mechanism, a fixed-length time-domain amplitude signal sequence within the current sliding time window is obtained, which avoids the fragmentation of anomalous signals crossing the window boundary and transforms the continuous data stream into discrete samples suitable for batch processing by deep learning networks.

[0041] S1022. A Gaussian filter is used to suppress background noise in the time-domain amplitude signal. The noise-reduced time-domain amplitude signal is then subjected to continuous wavelet transform to generate a continuous wavelet transform acoustic map containing the distribution law of multi-channel signal energy with time and frequency.

[0042] In real pipeline installation sites and industrial operating environments, acoustic emission sensors not only capture valid signals caused by damage and leakage, but also collect a large amount of environmental interference noise, such as mechanical friction noise caused by the operation of pumps inside the pipeline, low-frequency turbulence noise generated by normal fluid flow, and broadband background noise transmitted from ground traffic. To eliminate interference, a Gaussian filter is first applied in the time domain to denoise the intercepted time-domain amplitude signal. The Gaussian filter uses a weight function that follows a normal distribution to perform a one-dimensional convolution operation with the original signal, which can smooth high-frequency random spikes in the signal, effectively suppress Gaussian white noise, and at the same time preserve the signal abrupt edges caused by real anomalies.

[0043] After basic time-domain denoising, a time-frequency domain transformation is performed. Since leakage and crack propagation signals typically exhibit non-stationary broadband high-frequency pulses, traditional Fast Fourier Transform (FFT) methods cannot reflect the abrupt changes in specific frequency components at specific time points. Therefore, Continuous Wavelet Transform (CWT) is employed. A suitable mother wavelet function for analyzing abrupt signals is selected, such as the Morlet wavelet. By performing inner product operations on the denoised time-domain amplitude signal under different scale and time-shift parameters, the one-dimensional time-domain signal is mapped and expanded into a two-dimensional time and scale matrix. The values ​​in this matrix represent the energy amplitude at a specific time point and frequency scale. These energy amplitudes are further converted into visual pixel color values ​​using a pseudo-color mapping algorithm, thereby generating a two-dimensional image reflecting the distribution of signal energy over time and frequency. Furthermore, the two-dimensional images obtained by synchronously acquiring and transforming multiple adjacent acoustic emission sensor channels are stacked and fused along the channel dimension, ultimately generating a continuous wavelet transform acoustic map containing multi-channel signal energy and showcasing both time-frequency distribution and spatial perspective.

[0044] S1023. Input the continuous wavelet transform acoustic map into a temporal convolutional network containing a spatial convolutional layer and a long short-term memory network in a time sequence according to the time sequence, determine whether there is a leakage feature in the current sliding time window, and take the leakage feature as an acoustic anomaly feature when there is a leakage feature.

[0045] After converting the abstract one-dimensional acoustic signal into an intuitive two-dimensional acoustic map, a pre-trained Temporal Convolutional Network (TCN) is used for pattern recognition. This TCN employs a cascaded hybrid architecture. The network's front end contains multiple layers of two-dimensional spatial convolutional layers and max-pooling layers. When the continuous wavelet transform acoustic map is input, the spatial convolutional layers are responsible for scanning the frequency band energy accumulation patches, high-frequency pulse textures, and energy gradient distributions between different channels in the image, extracting the high-level spatial representation vector of the acoustic map.

[0046] Since pipe leaks or persistent tears are dynamic physical processes that evolve over time, exhibiting clear contextual and temporal dependencies, the feature vectors extracted by the spatial convolutional layers are sequentially fed into the Long Short-Term Memory (LSTM) module at the back end of the network. LLSTM, relying on its unique input, forget, and output gate structures, selectively memorizes key features of the current time window and nonlinearly fuses them with implicit state features from past historical time windows, thereby learning the evolution trend of acoustic signals over a longer period. The network's output, through fully connected layers and a normalized exponential function, outputs the probability distribution of the acoustic map belonging to different state categories within the current sliding time window. Based on this, it determines whether typical leak characteristics or other destructive ringing wave characteristics exist within the current sliding time window, and when leak characteristics are confirmed, they are registered as acoustic anomalies captured in the current stage.

[0047] S1024. When a leakage feature exists within the current sliding time window, and the duration of the leakage feature exceeds a preset abnormal duration threshold within multiple consecutive sliding time windows, the abnormal state is confirmed, and the start and end time periods corresponding to the multiple consecutive sliding time windows are confirmed as acoustic abnormal time windows.

[0048] In real outdoor environments, isolated events such as occasional flying stones hitting the outer wall of pipes or single vibrations from heavy construction machinery hammers can generate high-frequency broadband pseudo-anomaly signals that are extremely similar to pipe damage within a very short time. If an alarm is triggered directly based solely on the identification results of a single window, the system's false alarm rate will be too high. To filter out these transient and occasional interferences, a strict time-dimensional stability verification mechanism is introduced at the decision logic layer. A state counter is maintained in memory to continuously monitor the output of the temporal convolutional network. Only when a leakage feature is detected within the current sliding time window, and this leakage feature is stably and continuously identified in multiple subsequent sliding time windows, and the overall physical time span exceeds a pre-defined threshold for the abnormal duration (e.g., 1.5 seconds), is the acoustic anomaly officially confirmed. After confirmation, the start and end times of the absolute physical clocks corresponding to the multiple consecutive sliding time windows that triggered the alarm are extracted. This time span is combined, recorded, and confirmed as the acoustic anomaly time window, providing a time anchor for subsequent spatiotemporal intersection comparisons.

[0049] In one alternative implementation, after acquiring acoustic emission signals during operation using multi-channel acoustic emission sensors arranged on the outer wall of the pipe, and inputting the continuous wavelet transform acoustic map into a pre-trained temporal convolutional network for recognition, the method further includes: High-dimensional spatial features corresponding to each acquisition channel are extracted from the continuous wavelet transform acoustic map. Feature splicing and dimensionality reduction operations are performed on the high-dimensional spatial features of each acquisition channel to obtain the channel fusion feature vector. The channel fusion feature vector is input into a preset leakage aperture regression network to calculate the leakage aperture estimate corresponding to the leakage feature. The estimated leakage aperture is mapped to the initial leakage level parameter, and the leakage level parameter is used as an auxiliary evaluation index to characterize the severity of acoustic anomalies.

[0050] Specifically, the above steps are crucial for further assessing the severity of the anomaly after it has been confirmed, and include the following sub-steps: S1025. Extract the high-dimensional spatial features of each acquisition channel from the continuous wavelet transform acoustic map, and perform feature splicing and dimensionality reduction operations on the high-dimensional spatial features of each acquisition channel to obtain the channel fusion feature vector.

[0051] When a leak occurs due to a breach or crack in the pipe wall, the leak point acts as an acoustic emission source, propagating elastic waves to sensors on both sides. Due to the exponential energy attenuation of sound waves propagating through the metal pipe wall and surrounding soil, and the dispersion effect caused by differences in the propagation speeds of different frequency components, the acoustic maps generated by sensor acquisition channels at different distances from the sound source exhibit significant differences in energy intensity and high-frequency bandwidth. By using a pre-convolutional layer, high-dimensional spatial feature matrices corresponding to each acquisition channel are separated and extracted from the continuous wavelet transform acoustic maps. Subsequently, these feature matrices representing monitoring results from different physical locations are concatenated along the depth channel dimension. To prevent excessively high feature dimensions after concatenation, which could lead to model overfitting or computational explosion, a one-to-one convolutional kernel with cross-channel information interaction capability is used to reduce the dimensionality and smooth the compressed feature tensor, ultimately fusing them into a compact channel fusion feature vector.

[0052] S1026. Input the channel fusion feature vector into the preset leakage aperture regression network to calculate the leakage aperture estimate corresponding to the leakage feature.

[0053] The extracted channel fusion feature vector not only encodes the absolute energy intensity at the time of sound source generation but also implicitly contains the attenuation gradient information of the sound wave across multiple sensor nodes. This channel fusion feature vector is then fed into a pre-constructed and trained leakage orifice diameter regression network. This regression network typically consists of multiple fully connected layers containing nonlinear activation functions. During model training, a large amount of sample data covering different pipe diameters, internal operating pressures, and man-made borehole sizes has been used to regress and fit the network weights. During the inference calculation phase, the leakage orifice diameter regression network directly outputs a continuous physical quantity value—the estimated leakage orifice diameter corresponding to the current leakage characteristic—through multiple layers of nonlinear mapping.

[0054] S1027. Map the estimated leakage aperture value to the initial leakage level parameter, and use the leakage level parameter as an auxiliary evaluation index to characterize the severity of acoustic anomalies.

[0055] While direct physical aperture values ​​are accurate, they are not easily calculated directly with other modal macroscopic indicators within a comprehensive evaluation system. Therefore, based on a pre-defined expert rule mapping table, the output continuous leakage aperture estimates are discretized and mapped to specific initial leakage level parameters. For example, aperture estimates between zero and two millimeters are classified as micro-leakage parameters, those between two and ten millimeters as moderate leakage parameters, and those exceeding ten millimeters as structural failure or jet leakage parameters. Subsequently, the determined leakage level parameters are saved and used as a quantitative auxiliary assessment indicator characterizing the severity of the acoustic anomaly damage, serving as a reference in the risk-weighted summation in subsequent steps.

[0056] S103: Obtain vibration signals along the pipeline from distributed optical fiber sensors laid along the pipeline, extract spatiotemporal features from the vibration signals along the pipeline, and identify disturbance events and their corresponding optical fiber positioning results.

[0057] For trunk pipelines with lengths reaching tens or even hundreds of kilometers, traditional point sensors and image acquisition equipment face physical bottlenecks in terms of detection coverage and real-time performance. Therefore, this embodiment introduces distributed fiber optic sensors as a supplementary and extended sensing layer. These distributed fiber optic sensors utilize communication optical cables laid parallel to the pipeline in the same trench or directly wrapped around the outer wall of the pipeline as a continuous sensing medium. They eliminate the need for a power supply along the pipeline route, enabling 24 / 7, blind-spot-free monitoring of geological activities and third-party construction activities around long-distance pipelines.

[0058] In one optional implementation, vibration signals along the pipeline are acquired from distributed optical fiber sensors laid along the pipeline. Spatiotemporal features of the vibration signals are extracted to identify disturbance events and their corresponding optical fiber positioning results, including: It receives the optical phase change signal returned by a distributed optical fiber sensor based on the coherent Rayleigh scattering principle and uses the optical phase change signal as a vibration signal along the line; The energy distribution features of the vibration signal along the line in the spatial dimension and the frequency distribution features in the time dimension are extracted, and spatiotemporal features are spliced ​​to obtain a multidimensional disturbance feature set; The multidimensional disturbance feature set is input into a pre-set event recognition network. The support vector machine classification layer is used to filter environmental noise signals and identify and distinguish third-party construction disturbance events from real damage precursor events. The distinguished events are taken as disturbance events. Based on the fiber optic scattering distance point corresponding to the identified disturbance event, the longitudinal physical distance of the disturbance event along the axial direction of the long-distance pipeline is determined, and the longitudinal physical distance is used as the fiber optic positioning result.

[0059] Specifically, the above steps can be implemented through the following sub-steps S1031 to S1034: S1031. Receive the optical phase change signal returned by the distributed optical fiber sensor based on the coherent Rayleigh scattering principle, and use the optical phase change signal as the vibration signal along the line.

[0060] This embodiment employs optical time-domain reflectometry (OTDR) based on the principle of coherent Rayleigh scattering. The fiber optic interrogator, located at the monitoring center, periodically injects highly coherent and extremely narrow linewidth probe laser pulses into the core of the sensing fiber. As the laser pulse propagates forward in the fiber medium, the non-uniform distribution of microscopic refractive index due to the fiber manufacturing process generates continuous Rayleigh backscattered light. When heavy excavation, manual excavation, or ground fracturing causes rapid deformation of the pipe wall above or around the pipeline, these mechanical vibrations and stress changes are coupled to the sensing fiber through the soil medium, inducing localized minute axial stretching or radial compression. This physical deformation alters the local refractive index of the fiber core through the elasto-optic effect, leading to a dynamic change in the optical phase of the Rayleigh backscattered light generated at that location. The receiver of the fiber optic interrogator uses interferometric demodulation techniques, such as homodyne detection or a heterodyne demodulation scheme based on a 90-degree phase shifter, combined with a phase dewinding algorithm, to calculate the optical phase change signal at different times for each spatial sampling point along the fiber. Since the optical phase shift has a highly linear mapping relationship with the externally applied vibration or strain, the demodulated optical phase change signal is directly received and used as a vibration signal along the pipeline that reflects the real-time stress state along the pipeline.

[0061] S1032. Extract the energy distribution features of the vibration signal along the line in the spatial dimension and the frequency domain distribution features in the time dimension, and perform spatiotemporal feature splicing to obtain a multidimensional disturbance feature set.

[0062] The original vibration signal along the pipeline is a two-dimensional data matrix containing a large amount of time and space dimensions. Directly using it for classification and identification would suffer from the curse of dimensionality. Therefore, targeted feature engineering processing is required. In the spatial dimension, the short-time energy variance, spatial zero-crossing rate, and signal crest factor of the vibration signal at multiple adjacent spatial sampling points are calculated. These spatial features can effectively characterize the ripple range and diffusion intensity of the energy source causing the vibration along the pipeline axis. In the time dimension, a fast Fourier transform is performed on the time series signal of each spatial point where vibration occurs to extract its frequency domain components; simultaneously, Mel frequency cepstral coefficients can be further calculated to obtain the energy distribution envelope characteristics of the signal in different frequency bands. By extracting the frequency domain features of the time dimension, the characteristic frequencies of different mechanical equipment can be effectively distinguished. For example, the impact of a manual pickaxe has a wide-band discrete impact spectrum, while the movement of heavy excavator tracks exhibits low-frequency periodic spectral components. Finally, the extracted spatial energy distribution features and the temporal frequency distribution features are spliced ​​and fused to form a multi-dimensional disturbance feature set that can comprehensively describe the current state of the optical fiber.

[0063] S1033. Input the multidimensional disturbance feature set into the preset event recognition network, use the support vector machine classification layer to filter the environmental noise signal, and identify and distinguish between third-party construction disturbance events and real damage precursor events, and take the distinguished events as disturbance events.

[0064] After constructing a multi-dimensional perturbation feature set, it is input into a pre-built event recognition network for logical inference. The event recognition network can be a deep architecture containing a multilayer perceptron, with the core being the Support Vector Machine (SVM) classification layer at the network's end. The SVM classification layer, based on a nonlinear radial basis function kernel, implicitly maps low-dimensional features to a high-dimensional space and searches for the optimal decision hyperplane in the high-dimensional space that maximizes the classification margin. In the pipeline monitoring scenario, the SVM classification layer first performs a filtering task, identifying and eliminating environmental noise signals that are random, long-lasting, and have gradual energy, such as strong winds blowing through vegetation, rainfall impacts, and background vibrations caused by vehicles traveling at constant speeds on normal roads. After eliminating environmental noise, the classification layer performs multi-class classification on the remaining high-energy burst features, focusing on identifying and distinguishing third-party construction disturbance events that pose a direct threat to pipeline safety, such as excavator excavation or pile driver impacts very close to the pipeline, and real pre-damage events induced by severe bending strain caused by soil loss in the trench leading to pipeline suspension. The high-risk events successfully identified and distinguished by the support vector machine classification layer are uniformly identified as perturbation events in the fiber optic mode.

[0065] S1034. Based on the fiber optic scattering distance point corresponding to the identified disturbance event, determine the longitudinal physical distance of the disturbance event along the axial direction of the long-distance pipeline, and use the longitudinal physical distance as the fiber optic positioning result.

[0066] Based on the principle of optical time-domain reflectometry, the time difference of a laser pulse's round-trip flight in an optical fiber has a precise one-to-one correspondence with the location of the disturbance. Assuming the speed of light in a vacuum is... The refractive index of the fiber core is The time difference consumed when the optical fiber interrogation host records a certain characteristic optical phase change and returns it to the receiver is... At that time, the fiber optic scattering distance point that generated the disturbance signal It can be calculated using a formula, the specific formula is as follows:

[0067] In the formula, the physical meaning of each parameter is clear. Represents the constant of the speed of light in a vacuum. This represents the total time difference of flight of the laser pulse from the transmitter to the receiver, during which the backscattered light carrying perturbation phase information returns. This represents the effective refractive index coefficient of the optical fiber medium. Based on the time difference... The specific fiber optic scattering distance point where the abnormal signal occurred was determined. Because the laying of sensing optical cables may involve localized coiling or serpentine laying to adapt to terrain, resulting in increased excess length, geometric compensation calculations are performed by combining optical cable construction drawings and historical calibration data to accurately map the optical fiber cable length into the longitudinal physical distance along the axial direction of the long-distance pipeline. This calculated longitudinal physical distance is then used as the optical fiber positioning result output by the distributed optical fiber sensing layer.

[0068] S104 maps the visual defect location, acoustic anomaly time window, and fiber optic positioning results to a preset unified pipe section coordinate system, performs cross-modal consistency determination, and outputs the final pipe damage level when the visual defect location, acoustic anomaly time window, and fiber optic positioning results meet the preset spatiotemporal consistency conditions, combining candidate defect categories, acoustic anomaly characteristics, and disturbance events.

[0069] The above steps obtained independent identification results from three different physical dimensions: endoscopic vision, high-frequency acoustics of the outer wall, and distributed fiber optic strain along the line. However, due to the limitations of the hardware itself and interference from the complex industrial environment, any single-modal identification result is susceptible to false alarms and false negatives. To ensure the absolute reliability of the final alarm decision, cross-modal consistency judgment must be performed. This requires different sensors to provide mutually corroborating chain evidence in terms of temporal and spatial physical laws. Only after the spatiotemporal logical chain is closed can a final judgment be made.

[0070] In one optional implementation, the visual defect location, acoustic anomaly time window, and fiber optic positioning results are mapped to a preset unified pipe segment coordinate system. Cross-modal consistency determination is performed, and the final pipe damage level is output when the visual defect location, acoustic anomaly time window, and fiber optic positioning results meet preset spatiotemporal consistency conditions. This level includes: A three-dimensional pipe network model is established with the fixed starting point of the pipeline as the absolute origin, and the three-dimensional pipe network model is used as a unified pipe segment coordinate system. Based on the preset spatial tolerance parameters, the location of visual defects is mapped to the first spatial coordinate interval in the three-dimensional pipeline network model, and the fiber optic positioning result is mapped to the second spatial coordinate interval in the three-dimensional pipeline network model. The image acquisition device is positioned inside the pipeline when the acoustic anomaly time window occurs, and the third spatial coordinate interval corresponding to the acoustic anomaly time window is calculated based on the position information. Calculate the spatial intersection-union ratio of the first, second, and third spatial coordinate intervals in the three-dimensional pipeline network model, extract the event time intervals where disturbance events occur, and obtain the time overlap between the acoustic anomaly time window and the event time interval. When the spatial intersection-to-union ratio is greater than the preset spatial consistency threshold and the temporal overlap is greater than the preset temporal consistency threshold, the location of the visual defect, the time window of the acoustic anomaly, and the fiber optic positioning result are determined to meet the spatiotemporal consistency condition.

[0071] Specifically, the operation steps for performing cross-modal consistency determination are divided into the following sub-steps S1041 to S1045: S1041. Establish a three-dimensional pipe network model with the fixed starting point of the pipeline as the absolute origin, and use the three-dimensional pipe network model as a unified pipe segment coordinate system.

[0072] Different sensing devices employ different spatial reference benchmarks: vision relies on internal odometers, fiber optics on time-of-flight ranging, and acoustic emission on a topological grid deployed on the outer wall. To enable location comparison under a unified standard, geographic information data or 3D building information model data of the pipeline network are pre-imported. Physical entities such as the control room of the first station of the long-distance pipeline, specific geographic coordinate control piles, or fixed pump station entrances are used as the absolute origin of the spatial coordinate system, i.e., the zero point. Subsequently, based on the actual pipeline route, burial depth, and curvature of the turns, a proportionally mapped 3D cylindrical model of the pipeline network is constructed in digital space. This 3D pipeline network model is then set as the unified pipeline segment coordinate system benchmark for data fusion and interaction throughout the entire system.

[0073] S1042. Combining the preset spatial tolerance parameters, the location of the visual defect is mapped to the first spatial coordinate interval in the three-dimensional pipeline model, and the fiber optic positioning result is mapped to the second spatial coordinate interval in the three-dimensional pipeline model.

[0074] Considering that measurement sensors inevitably have system-level errors in engineering practice, such as the cumulative error of the odometer caused by the wheel of the image acquisition device slipping in a wet pipe, or the physical limitations of the spatial resolution of fiber optic sensors, an absolutely rigid point cannot be used for comparison. Instead, a flexible regional verification mechanism is introduced. When mapping the point-like 3D visual defect location output in step S101 to a unified 3D pipeline model, a pre-set 3D spatial tolerance parameter is used to expand outward around the defect point, for example, by two meters forward and backward along the pipeline axis, forming a first spatial coordinate interval with actual volume in the model. Similarly, for the fiber optic positioning result output in step S103, since it is mainly a one-dimensional longitudinal distance, this one-dimensional distance is projected onto the entire circular cross-section of the pipeline, and the axial interval is scaled down in combination with the tolerance parameter to construct a corresponding second spatial coordinate interval in the 3D pipeline model.

[0075] S1043. Obtain the travel position information of the image acquisition device in the pipeline when the acoustic anomaly time window occurs, and calculate the third spatial coordinate interval corresponding to the acoustic anomaly time window based on the travel position information.

[0076] In acoustic monitoring, acoustic anomalies are recorded in the form of time windows. To achieve spatial uniformity, it is necessary to query the timestamp synchronization logs using a unified internal network clock server. This allows for the precise retrieval of the image acquisition device's location and trajectory within the pipeline at the exact moment the acoustic anomaly occurred within the time window. Since the image acquisition device may be continuously moving within the time window, this trajectory is extracted and, combined with the empirical attenuation transmission distance of acoustically emitted elastic waves in the pipeline's metal wall as the radius of influence, the corresponding third spatial coordinate interval is calculated and generated in the three-dimensional pipeline model using this trajectory as the central axis.

[0077] S1044. Calculate the spatial intersection-union ratio of the first spatial coordinate interval, the second spatial coordinate interval, and the third spatial coordinate interval in the three-dimensional pipeline network model, extract the event time interval of the disturbance event, and obtain the time overlap between the acoustic anomaly time window and the event time interval.

[0078] In the core computational stage of this data convergence, a geometric intersection algorithm is used to calculate the overlapping volume of the first, second, and third spatial coordinate intervals within the 3D pipeline network model, as well as the total covered union volume. The spatial intersection-union ratio is obtained by calculating the ratio of the intersection volume to the union volume. A higher spatial intersection-union ratio indicates a more concentrated physical location for multi-source anomaly alarms. Simultaneously, in the time matching dimension, the exact start and end clocks recorded by the distributed optical fibers when identifying disturbance events are extracted, constituting the event time interval. This event time interval is then aligned and compared with the acoustic anomaly time window obtained from multi-channel acoustic emission calculations on the time axis, calculating the proportion of time overlap between the two time periods within the same time span.

[0079] S1045. When the spatial intersection-to-union ratio is greater than the preset spatial consistency threshold and the temporal overlap is greater than the preset temporal consistency threshold, the location of the visual defect, the time window of the acoustic anomaly, and the fiber optic positioning result are determined to meet the spatiotemporal consistency condition.

[0080] The calculated spatiotemporal quantification indicators are compared with preset judgment thresholds. If the calculated spatial crossover ratio is greater than the preset spatial consistency threshold, it indicates that the visually identified inner wall damage, the leak sound source captured by the outer wall, and the external construction disturbance located by the fiber optic cable are highly consistent in physical three-dimensional space. Furthermore, if the calculated temporal overlap is greater than the preset temporal consistency threshold, it indicates that these abnormal signals occurred simultaneously or sequentially with a strict causal logic within a very short time window. Only when both of these stringent spatiotemporal cross-validations are passed is it confirmed that the multimodal sensors provide irrefutable corroborating evidence, completely eliminating false alarms caused by a single sensor from flying stones, isolated water flow sounds, or lens dirt, and determining that the system's multimodal data meets the spatiotemporal consistency condition.

[0081] In one alternative implementation, the final pipe damage level is output by combining candidate defect categories, acoustic anomaly features, and disturbance events, including: When determining whether the spatiotemporal consistency condition is met, the corresponding visual risk value, acoustic risk value, and fiber optic risk value are determined based on the candidate defect category, acoustic anomaly characteristics, and disturbance event, respectively. The visual risk value is assigned a first confidence weight, the acoustic risk value is assigned a second confidence weight, and the fiber optic risk value is assigned a third confidence weight. Based on the first confidence weight, the second confidence weight, and the third confidence weight, the visual risk value, acoustic risk value, and fiber optic risk value are weighted and summed to obtain the comprehensive damage risk value. Based on the preset range of the comprehensive damage risk value, the final pipeline damage level is determined, and an automatic alarm command with a maintenance priority label is generated based on the final pipeline damage level.

[0082] Specifically, the above steps, which integrate severity data from multiple sources to output a quantitative rating, can be achieved through the following sub-steps S1046 to S1049: S1046. When determining whether the spatiotemporal consistency condition is met, the corresponding visual risk value, acoustic risk value and fiber optic risk value shall be determined according to the candidate defect category, acoustic anomaly characteristics and disturbance event, respectively.

[0083] Once the cross-modal spatiotemporal consistency condition is successfully triggered, the risk quantification and rating stage begins. An internally pre-built expert knowledge matrix mapping table assigns risk scores based on the qualitative categories analyzed in previous steps. For candidate defect categories identified by image recognition, lower visual risk values ​​are assigned if they belong to minor surface corrosion or small deposits; if they are identified as severe stress deformation or structural tears, extremely high visual risk values ​​are assigned. For acoustic anomalies detected by acoustic emission, the leakage level parameters calculated in step S1027 are directly retrieved and converted into digitally quantified acoustic risk values. For disturbance events detected by distributed optical fibers, scores are assigned based on the specific intrusion type determined by the support vector machine classification layer. For example, the risk value of optical fiber caused by heavy excavator mechanical cutting is much higher than the risk value of manual excavation or vehicle overloading.

[0084] S1047. Assign a first confidence weight to the visual risk value, a second confidence weight to the acoustic risk value, and a third confidence weight to the fiber optic risk value.

[0085] Because different sensors have different detection physical limits under specific operating conditions, the reliability of their output results varies. For example, visual endoscopy may be subject to judgment bias due to the influence of resolution and sediment cover; however, acoustic emission sensors are extremely sensitive to high-pressure gas leaks and have high confidence; fiber optics are accurate in locating external rough construction but are difficult to detect internal minute cracks. Therefore, it is necessary to dynamically or statically assign weights to different dimensions of risk based on the hardware characteristics of different sensors. For example, a first confidence weight can be assigned to visual risk values, a second confidence weight to acoustic risk values, and a third confidence weight to fiber optic risk values. The weight values ​​can be objectively calibrated during the system initialization phase using fuzzy hierarchical analysis, and the sum of all weights is limited to 100%.

[0086] S1048. Based on the first confidence weight, the second confidence weight, and the third confidence weight, the visual risk value, the acoustic risk value, and the fiber optic risk value are weighted and summed to obtain the comprehensive damage risk value.

[0087] After obtaining the risk scores and corresponding confidence weights for each dimension, a multi-dimensional weighted fusion calculation is performed. The visual risk value is multiplied by the first confidence weight, the acoustic risk value by the second confidence weight, and the fiber optic risk value by the third confidence weight. The three products are then summed. This weighted summation process comprehensively considers the internal physical morphology, external acoustic energy, and the degree of damage caused by macroscopic strain along the pipeline. This ensures that the calculated comprehensive damage risk value is no longer biased towards a single data point, but rather provides a system-level pipeline health quantification indicator with a global perspective and objectivity.

[0088] S1049. Based on the preset value range of the comprehensive damage risk value, determine the final pipeline damage level, and generate an automatic alarm command with a maintenance priority label based on the final pipeline damage level.

[0089] The strategy database is divided into preset numerical ranges for safety risks at different gradients. Through comparison and retrieval, the specific numerical range into which the comprehensive damage risk value falls is determined, thus identifying the final pipeline damage level. For example, if it falls into a lower score range, it is defined as a level requiring regular monitoring due to routine wear; if it falls into a medium score range, it is defined as a level with potential leakage; and if it falls into a high score range, it is defined as a level of severe rupture and high-pressure danger. Once the final pipeline damage level is determined, the linkage module is automatically invoked to generate a structured automatic alarm command. This automatic alarm command embeds processing suggestions with differentiated maintenance priority tags and detailed three-dimensional positioning coordinates. Subsequently, the command is pushed in real-time to the large-screen display terminal of the pipeline safety operation and maintenance control center or the handheld mobile terminal of the repair personnel via industrial Ethernet or wireless communication network. This automated closed-loop processing flow, from multi-source front-end heterogeneous data acquisition, cross-modal intelligent recognition, and spatiotemporal cross-verification to scientific hierarchical decision-making and alarm in the back-end, effectively improves the operational detection accuracy and emergency maintenance efficiency of complex long-distance pipeline networks.

[0090] This invention also provides a pipeline damage detection device, referring to... Figure 3 The diagram shows a functional block diagram of a pipeline damage detection device 300 according to the present invention. The device may include the following modules: The image recognition module 301 is used to acquire images of the inner wall of the pipe through an image acquisition device inside the pipe, and to identify the inner wall images of the pipe using a preset defect hierarchy classification model to obtain candidate defect categories and the visual defect locations corresponding to the candidate defect categories. The acoustic feature extraction module 302 is used to collect acoustic emission signals under operating conditions through multi-channel acoustic emission sensors arranged on the outer wall of the pipe, convert the acoustic emission signals into continuous wavelet transform acoustic maps, and input the continuous wavelet transform acoustic maps into a pre-trained temporal convolutional network for recognition, thereby extracting acoustic anomaly features and their corresponding acoustic anomaly time windows. The fiber optic positioning module 303 is used to acquire vibration signals along the pipeline collected by distributed fiber optic sensors laid along the pipeline, extract spatiotemporal features from the vibration signals along the pipeline, and identify disturbance events and their corresponding fiber optic positioning results. The consistency determination and output module 304 is used to map the visual defect location, acoustic anomaly time window and fiber optic positioning result to a preset unified pipe section coordinate system, perform cross-modal consistency determination, and when the visual defect location, acoustic anomaly time window and fiber optic positioning result meet the preset spatiotemporal consistency conditions, it combines the candidate defect category, acoustic anomaly characteristics and disturbance events to output the final pipe damage level.

[0091] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the pipeline damage detection method of the present invention.

[0092] The communication bus mentioned in the aforementioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the figure, but this does not mean that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include RAM, or it can include NVM, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0093] The processors mentioned above can be general-purpose processors, including CPUs, network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0094] Furthermore, to achieve the above objectives, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the pipeline damage detection method of the embodiments of the present invention.

[0095] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0096] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, electronic devices, apparatuses, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0097] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0099] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0100] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of pipeline break detection, the method comprising: The method includes: Images of the inner wall of the pipe are acquired by an image acquisition device inside the pipe, and the images of the inner wall of the pipe are identified by a preset defect hierarchy classification model to obtain candidate defect categories and the visual defect locations corresponding to the candidate defect categories. Acoustic emission signals during operation are collected by multi-channel acoustic emission sensors arranged on the outer wall of the pipe. The acoustic emission signals are converted into continuous wavelet transform acoustic maps. The continuous wavelet transform acoustic maps are then input into a pre-trained temporal convolutional network for identification, and acoustic anomaly features and their corresponding acoustic anomaly time windows are extracted. The vibration signals along the pipeline are collected by distributed optical fiber sensors laid along the pipeline. The spatiotemporal features of the vibration signals along the pipeline are extracted to identify the disturbance events and their corresponding optical fiber positioning results. The visual defect location, the acoustic anomaly time window, and the fiber optic positioning result are mapped to a preset unified pipe section coordinate system. Cross-modal consistency determination is performed. When the visual defect location, the acoustic anomaly time window, and the fiber optic positioning result meet the preset spatiotemporal consistency conditions, the final pipe damage level is output by combining the candidate defect category, the acoustic anomaly characteristics, and the disturbance event.

2. The pipeline break detection method of claim 1, wherein, The step of using a pre-set defect hierarchy classification model to identify the inner wall image of the pipe to obtain candidate defect categories and the corresponding visual defect locations includes: Visual features of the pipe inner wall image are extracted, and candidate regions in the pipe inner wall image are located using a candidate box detection network; The candidate region is input into the defect hierarchical classification model, and the candidate region is binary classified into normal state or abnormal state by the first hierarchical classifier; The candidate regions in the anomalous state are further subdivided into at least one of cracks, holes, breaches and sediment occlusion by a second-level classifier trained by the hierarchical loss function to obtain the candidate defect categories. The hierarchical loss function introduces a category weight matrix calculated based on a historical defect dataset to compensate for the classification bias of a small number of defect types. The odometer data of the image acquisition device when acquiring the image of the inner wall of the pipe is obtained. The odometer data and the pixel coordinates of the candidate region in the image of the inner wall of the pipe are combined to calculate the three-dimensional physical coordinates of the physical region of the inner wall of the pipe corresponding to the candidate region in the real pipe. The three-dimensional physical coordinates are used as the location of the visual defect.

3. The method of claim 1, wherein The process of converting the acoustic emission signal into a continuous wavelet transform acoustic map and inputting the continuous wavelet transform acoustic map into a pre-trained temporal convolutional network for recognition, extracting acoustic anomaly features and their corresponding acoustic anomaly time windows, includes: The acoustic emission signal is continuously sampled using a sliding time window to obtain the time-domain amplitude signal within the current sliding time window; A Gaussian filter is used to suppress background noise in the time-domain amplitude signal. A continuous wavelet transform is then performed on the noise-reduced time-domain amplitude signal to generate the continuous wavelet transform acoustic map containing the distribution law of multi-channel signal energy with time and frequency. The continuous wavelet transform acoustic map is sequentially input into the temporal convolutional network containing spatial convolutional layers and long short-term memory networks according to the time sequence, and it is determined whether there is a leakage feature within the current sliding time window. If the leakage feature exists, it is used as the acoustic anomaly feature. If the leakage feature exists within the current sliding time window, and the duration of the leakage feature exceeds a preset abnormal duration threshold within multiple consecutive sliding time windows, an abnormal state is confirmed, and the start and end time periods corresponding to the multiple consecutive sliding time windows are confirmed as the acoustic abnormal time window.

4. The method of claim 3, wherein After acquiring acoustic emission signals during operation through multi-channel acoustic emission sensors arranged on the outer wall of the pipe, and inputting the continuous wavelet transform acoustic map into a pre-trained temporal convolutional network for recognition, the process further includes: High-dimensional spatial features corresponding to each acquisition channel are extracted from the continuous wavelet transform acoustic map. Feature splicing and dimensionality reduction operations are performed on the high-dimensional spatial features of each acquisition channel to obtain the channel fusion feature vector. The channel fusion feature vector is input into a preset leakage aperture regression network to calculate the estimated leakage aperture value corresponding to the leakage feature. The estimated leakage aperture is mapped to an initial leakage level parameter, and the leakage level parameter is used as an auxiliary evaluation index characterizing the severity of the acoustic anomaly.

5. The method of claim 1, wherein The process of acquiring vibration signals along the pipeline from distributed optical fiber sensors laid along the pipeline, extracting spatiotemporal features from the vibration signals, and identifying disturbance events and their corresponding optical fiber positioning results includes: The optical phase change signal returned by the distributed optical fiber sensor based on the coherent Rayleigh scattering principle is received, and the optical phase change signal is used as the vibration signal along the line. The energy distribution features in the spatial dimension and the frequency domain distribution features in the temporal dimension of the vibration signal along the line are extracted, and spatiotemporal features are spliced ​​to obtain a multidimensional disturbance feature set; The multidimensional disturbance feature set is input into a preset event recognition network. The environmental noise signal is filtered by the support vector machine classification layer, and the third-party construction disturbance event and the actual damage precursor event are identified and distinguished. The distinguished event is taken as the disturbance event. Based on the fiber scattering distance point corresponding to the identified disturbance event, the longitudinal physical distance of the disturbance event along the axial direction of the long-distance pipeline is determined, and the longitudinal physical distance is used as the fiber positioning result.

6. The method of claim 1, wherein The process of mapping the visual defect location, the acoustic anomaly time window, and the fiber optic positioning result to a preset unified pipe segment coordinate system, performing cross-modal consistency determination, and outputting the final pipe damage level when the visual defect location, the acoustic anomaly time window, and the fiber optic positioning result meet preset spatiotemporal consistency conditions includes: A three-dimensional pipe network model is established with the fixed starting point of the pipeline as the absolute origin, and the three-dimensional pipe network model is used as the coordinate system of the unified pipe segment; Based on the preset spatial tolerance parameters, the location of the visual defect is mapped to the first spatial coordinate interval in the three-dimensional pipeline model, and the fiber positioning result is mapped to the second spatial coordinate interval in the three-dimensional pipeline model. The image acquisition device is positioned inside the pipe when the acoustic anomaly time window occurs, and the third spatial coordinate interval corresponding to the acoustic anomaly time window is calculated based on the position information. Calculate the spatial intersection-union ratio of the first spatial coordinate interval, the second spatial coordinate interval, and the third spatial coordinate interval in the three-dimensional pipeline network model, extract the event time interval where the disturbance event occurs, and obtain the time overlap between the acoustic anomaly time window and the event time interval; When the spatial intersection-to-union ratio is greater than a preset spatial consistency threshold and the temporal overlap is greater than a preset temporal consistency threshold, it is determined that the location of the visual defect, the time window of the acoustic anomaly, and the fiber optic positioning result satisfy the spatiotemporal consistency condition.

7. The method of claim 6, wherein The final pipeline damage level is output by combining the candidate defect category, the acoustic anomaly features, and the disturbance event, including: When determining that the spatiotemporal consistency condition is met, the corresponding visual risk value, acoustic risk value, and fiber optic risk value are determined according to the candidate defect category, the acoustic anomaly characteristics, and the disturbance event, respectively. The visual risk value is assigned a first confidence weight, the acoustic risk value is assigned a second confidence weight, and the fiber optic risk value is assigned a third confidence weight. Based on the first confidence weight, the second confidence weight, and the third confidence weight, the visual risk value, the acoustic risk value, and the optical fiber risk value are weighted and summed to obtain a comprehensive damage risk value. Based on the preset value range of the comprehensive damage risk value, the final pipeline damage level is determined, and an automatic alarm command with a maintenance priority label is generated based on the final pipeline damage level.

8. A pipe damage detection device, characterized in that, The device includes: The image recognition module is used to acquire images of the inner wall of the pipe through an image acquisition device inside the pipe, and to identify the inner wall images of the pipe using a preset defect hierarchy classification model to obtain candidate defect categories and the visual defect locations corresponding to the candidate defect categories. The acoustic feature extraction module is used to collect acoustic emission signals during operation through multi-channel acoustic emission sensors arranged on the outer wall of the pipe, convert the acoustic emission signals into continuous wavelet transform acoustic maps, and input the continuous wavelet transform acoustic maps into a pre-trained temporal convolutional network for identification, thereby extracting acoustic anomaly features and their corresponding acoustic anomaly time windows. The fiber optic positioning module is used to acquire vibration signals along the pipeline collected by distributed fiber optic sensors laid along the pipeline, extract spatiotemporal features from the vibration signals along the pipeline, and identify disturbance events and their corresponding fiber optic positioning results. The consistency determination and output module is used to map the visual defect location, the acoustic anomaly time window, and the fiber optic positioning result to a preset unified pipe section coordinate system, perform cross-modal consistency determination, and output the final pipe damage level when the visual defect location, the acoustic anomaly time window, and the fiber optic positioning result meet the preset spatiotemporal consistency conditions, in combination with the candidate defect category, the acoustic anomaly characteristics, and the disturbance event.

9. An electronic device, comprising: include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute a computer program stored in the memory, causing the electronic device to implement the pipeline damage detection method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the pipeline damage detection method as described in any one of claims 1 to 7.

Images