Multi-leak point identification method and device based on space-time feature map and convolutional neural network

Abstract

The application provides a multi-leak-point identification method and device based on a space-time feature map and a convolutional neural network. The method continuously collects pressure signals of each sensor along the axial direction of a pipeline, obtains a real-time space-time feature matrix with spatial positions as rows and time sequences as columns, performs sliding time window interception on the real-time space-time matrix to obtain a plurality of window matrices, generates a multi-channel space-time feature map based on each window matrix and the real-time space-time feature matrix, and identifies the space-time feature map to obtain a position set of the multi-leak-point, which includes pipeline segments corresponding to each leak point and a leak hole diameter grade. The method realizes accurate identification of the position of the multi-leak-point and the leak hole diameter grade of the pipeline based on the pressure space-time feature map and the convolutional neural network.

Description

Technical Field

[0001] This application relates to the field of pipeline safety monitoring technology, and more specifically, to a method for identifying multiple leaks based on spatiotemporal feature maps and convolutional neural networks. Background Technology

[0002] Carbon capture, utilization, and storage (CCUS) technology is a key pathway to achieving carbon neutrality, with the safe long-distance transport of carbon dioxide relying on high-pressure pipelines. Pipeline leak detection algorithms (such as negative pressure wave methods, mass balance methods, and statistical analysis methods) are core means to ensure operational safety, but they must undergo thorough testing and verification before being put into practical application. However, conducting large-scale leak experiments on real-world pipelines, especially destructive tests involving multiple locations and apertures, is virtually impossible due to extremely high costs, significant safety risks, and severe environmental damage.

[0003] Current technologies primarily rely on simulation systems for preliminary verification of leak detection algorithms. These simulation systems are typically based on a single leak point assumption, meaning a leak location and orifice diameter are pre-defined on the pipeline model. Transient response data such as pressure and flow rate are generated by solving fluid dynamics equations. During algorithm verification, the simulation data is input into detection modules such as the mass balance method to evaluate its accuracy in locating a single leak event and its alarm response time. Some advanced systems also allow changing the parameters of the single leak point to traverse different operating conditions.

[0004] However, existing single-point simulation systems cannot generate superimposed signals for multi-point leaks, leading to the following specific problems: The mass balance method calculates the total leakage based on the difference in mass flow rate between the pipe inlet and outlet. In a single-point model, this difference can uniquely map to the leak location. However, in multi-point leaks, the total flow rate difference is the sum of the contributions from all leak points, making it impossible to distinguish the flow rate contributed by each point individually. This results in an underdetermined location equation and a non-unique solution. Therefore, existing simulation systems cannot comprehensively and realistically evaluate the response characteristics of leak detection algorithms when facing multi-location, multi-aperture leaks, which has become a bottleneck restricting the development of highly robust pipeline inspection technology. Summary of the Invention

[0005] The purpose of this application is to provide a method and apparatus for identifying multiple leaks based on spatiotemporal feature maps and convolutional neural networks, so as to achieve accurate identification of the location and leakage orifice level of multiple leaks in pipelines based on pressure spatiotemporal feature maps and convolutional neural networks.

[0006] Firstly, a method for identifying multiple leak points based on spatiotemporal feature maps and convolutional neural networks is provided, which may include: Pressure signals from each sensor along the pipeline axis are continuously collected to obtain a real-time spatiotemporal feature matrix with spatial position as rows and time as columns; The real-time spatiotemporal matrix is ​​subjected to a sliding time window to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. Based on each window matrix and the real-time spatiotemporal feature matrix, a multi-channel spatiotemporal feature map is generated; The spatiotemporal feature map is identified to obtain a set of multiple leak locations, which includes the pipe section corresponding to each leak and its leakage orifice size level.

[0007] In one possible implementation, the real-time spatiotemporal feature matrix is ​​obtained, with spatial location as rows and time as columns, including: Arrange the pressure values ​​corresponding to the pressure signals collected by each sensor at different times in chronological order; A two-dimensional real-time spatiotemporal feature matrix is ​​constructed using the spatial position sequence of the sensors along the pipeline axis as the row index and the time sequence as the column index.

[0008] In one possible implementation, a multi-channel spatiotemporal feature map is generated based on each window matrix and the real-time spatiotemporal feature matrix, including: Based on each window matrix, the pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are determined; The pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are stacked with the pressure waveform data of the corresponding time window to generate a three-channel spatiotemporal feature map.

[0009] In one possible implementation, the pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are stacked with the pressure waveform data of the corresponding time window to generate a three-channel spatiotemporal feature map, including: For each window matrix, perform the following steps: Use this window matrix as the first feature channel; Calculate the pressure signal difference between two adjacent sensors along the pipe axial direction in the window matrix to form a sequence of pressure differences between adjacent sensors, and use it as the second feature channel; Calculate the pressure change rate of each sensor on the time axis in the window matrix to form a pressure change rate sequence of each sensor, and use it as the third feature channel. The first, second, and third feature channels are stacked along the channel dimension to generate a three-channel spatiotemporal feature map.

[0010] In one possible implementation, the spatiotemporal feature map is identified to obtain a set of multiple leak location points, the corresponding pipe segment and its leakage aperture level for each leak point, including: Using a configured deep convolutional neural network, the spatiotemporal feature map is processed to obtain a multi-label classification vector with the same dimension as the number of pipeline segments. Each label represents whether there is a leak in the corresponding pipeline segment and the level of the leak aperture. Based on this multi-label classification vector, the set of multiple leak locations is determined, and the corresponding pipeline segment and leakage aperture level of each leak point are identified.

[0011] In one possible implementation, based on the multi-label classification vector, the set of multiple leak location locations is determined, including: Non-maximum suppression is performed on the multi-label classification vector to suppress pipe segment alarms where the spatially adjacent pipes have a non-maximum leakage probability, and the final set of multiple leak location is output.

[0012] In one possible implementation, nonmaximum suppression is performed on the multi-label classification vector, including: Extract the leakage probability corresponding to each pipeline segment from the multi-label classification vector output by the deep convolutional neural network; Pipe segments with a leakage probability greater than a preset first threshold are extracted as candidate segments and sorted from high to low leakage probability. Traverse the sorted candidate segments. For the current candidate segment, find its spatially adjacent candidate segments along the pipe axis. Spatially adjacent is defined as the index difference between two segments not exceeding 1 or the physical distance between them on the pipe is less than one segment length. If the leakage probability of an adjacent candidate segment is less than the leakage probability of the current candidate segment, then suppress the alarm output of that adjacent segment; otherwise, retain the one with the higher probability. The unsuppressed candidate segments are output as the final set of multiple leak locations.

[0013] Secondly, a multi-leak point identification device based on spatiotemporal feature maps and convolutional neural networks is provided, the device including: The acquisition unit is used to continuously collect pressure signals from each sensor along the axial direction of the pipeline and obtain a real-time spatiotemporal feature matrix with spatial position as rows and time order as columns. The interception unit is used to perform sliding time window interception on the real-time spatiotemporal matrix to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. The generation unit is used to generate multi-channel spatiotemporal feature maps based on each window matrix and the real-time spatiotemporal feature matrix. The identification unit is used to identify the spatiotemporal feature map to obtain a set of locations of multiple leaks, which includes the pipe section corresponding to each leak and its leakage orifice size level.

[0014] Thirdly, an electronic device is provided, which includes 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; When a processor executes a program stored in memory, it implements any of the steps described in the first aspect above.

[0015] Fourthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the steps of any of the methods described in the first aspect above.

[0016] The multi-leakage identification method and apparatus based on spatiotemporal feature maps and convolutional neural networks provided in this application converts one-dimensional time-series signals into image data with spatiotemporal correlation characteristics by constructing a three-channel spatiotemporal feature map that includes pressure waveform, pressure difference between adjacent sensors, and pressure change rate. This fully utilizes the spatial correlation and temporal dynamic characteristics of pressure signals. By using a deep convolutional neural network to extract features and perform multi-label classification on the spatiotemporal feature map, multiple leakage features superimposed in the signal can be effectively identified, achieving accurate location of multiple leaks. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a multi-leakage identification method based on spatiotemporal feature maps and convolutional neural networks provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a multi-leakage identification device based on spatiotemporal feature maps and convolutional neural networks provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The words "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are only used to distinguish different components. The words "comprising" or "including," etc., mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but do not exclude other elements or objects. The words "connected," "coupled," or "connected," etc., are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0021] Figure 1 This is a flowchart illustrating a multi-leakage point identification method based on spatiotemporal feature maps and convolutional neural networks, provided as an embodiment of this application. Figure 1 As shown, this method transforms one-dimensional pressure time-series data into a two-dimensional feature map containing spatiotemporal correlation characteristics, and then processes it using a deep learning model, thereby achieving accurate location and hierarchical identification of multiple leak points on a pipeline. This method may include: Step S100: Continuously collect pressure signals from each sensor along the axial direction of the pipeline to obtain a real-time spatiotemporal feature matrix with spatial position as rows and time as columns.

[0022] Specifically, in actual monitoring scenarios of long-distance pipelines, a pressure sensor is typically placed at regular intervals (e.g., 1 kilometer or 500 meters) along the pipeline axis. These sensors are numbered S1, S2, ..., Sn in sequence from the beginning to the end of the pipeline. Each sensor continuously collects pressure data at a preset sampling frequency (e.g., 100Hz).

[0023] In this step, the system receives pressure values ​​uploaded by each sensor in real time and organizes them into a two-dimensional matrix. In this matrix, the row index corresponds to the spatial location order of the sensors; for example, the first row corresponds to the data from sensor S1, the second row to the data from sensor S2, and so on. The column index corresponds to the temporal order; for example, the first column corresponds to the data collected at time t1, and the second column to the data at time t2. It should be understood that this construction method transforms the originally discrete, independent one-dimensional time-series signals into two-dimensional spatiotemporal data with spatial topological relationships. This data structure transformation is crucial, as it enables subsequent processing algorithms to simultaneously capture the dynamic changes in the pressure signal in the time dimension (such as the propagation of pressure waves) and the correlation characteristics in the spatial dimension (such as the pressure gradient between adjacent sensors), laying the data foundation for image-based feature extraction using convolutional neural networks.

[0024] Furthermore, although this embodiment describes a construction method with spatial location as rows and time as columns, in other embodiments, a transposed matrix (i.e., with time as rows and spatial location as columns) can also be used, as long as it can reflect the correlation of data in the two dimensions of time and space. This application does not impose strict limitations on this.

[0025] Step S200: The real-time spatiotemporal matrix is ​​subjected to a sliding time window to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window.

[0026] Specifically, since pipeline leakage is a continuously evolving process and the real-time data stream is infinitely long, directly processing the entire real-time spatiotemporal matrix is ​​not only computationally intensive but also makes it difficult to focus on leakage characteristics within a specific time period. Therefore, this step introduces a sliding time window mechanism. A fixed-length time window (e.g., a time length containing 512 sampling points) and a sliding step size (e.g., 64 sampling points) are set. This time window slides along the time axis (i.e., the column direction of the matrix), sliding one step at a time, thereby extracting local sub-matrices, i.e., window matrices, from the real-time spatiotemporal matrix. Each window matrix contains pressure waveform data from all sensors within that time segment. This extraction method preserves complete spatial dimension information while limiting the analysis scope of the temporal dimension, enabling the system to continuously and continuously monitor and analyze the pipeline status in a "micro-element" manner, and promptly detect sudden leakage signals.

[0027] Step S300: Generate multi-channel spatiotemporal feature maps based on each window matrix and the real-time spatiotemporal feature matrix.

[0028] Specifically, to improve the accuracy and robustness of subsequent model recognition, raw stress data alone is often insufficient. This step performs further feature engineering on the window matrix to generate a multi-channel spatiotemporal feature map. This spatiotemporal feature map can be understood as a data structure similar to a color image, containing multiple channels, each representing a specific physical feature.

[0029] For example, the first channel can be the raw pressure waveform data, reflecting the absolute change in pressure; the second channel can be the pressure difference data between adjacent sensors, reflecting the spatial gradient characteristics of pressure, which is very sensitive for identifying the location of leaks; the third channel can be the rate of change of pressure over time, reflecting the dynamic characteristics of pressure fluctuations. Through this multi-channel construction method, the spatiotemporal feature map not only contains the original information but also integrates derived higher-order features, greatly enriching the expressive power of the data and solving the problems of poor noise resistance and unclear features of a single signal source under complex operating conditions.

[0030] Step S400: Identify the spatiotemporal feature map to obtain a set of locations of multiple leak points. This set of locations includes the pipe section corresponding to each leak point and its leakage aperture level.

[0031] Specifically, this step utilizes a pre-configured deep convolutional neural network (CNN) model to process the aforementioned multi-channel spatiotemporal feature maps. CNN models possess powerful feature extraction capabilities, automatically learning deep features such as texture and edges unique to leakage signals from image-based data. The model's output is not a single classification result, but rather a set of locations for multiple leaks. It is important to emphasize that this set of locations is a multi-dimensional information set, containing not only the pipeline segment number where the leak point is located (i.e., location information), but also the leakage aperture level of that leak point (e.g., micro-leak, small leak, large leak, etc.). This multi-dimensional output characteristic overcomes the shortcomings of traditional methods that can only provide alarms or location information but cannot assess the severity of leaks. It provides pipeline maintenance personnel with a more comprehensive and quantitative basis for decision-making, enabling them to take differentiated emergency repair measures based on different aperture levels, thereby improving maintenance efficiency and reducing safety risks.

[0032] Example 2: Based on step S100 in Example 1, this example details the specific construction process of obtaining a real-time spatiotemporal feature matrix with spatial location order as rows and time order as columns.

[0033] Specifically, the process of obtaining this real-time spatiotemporal feature matrix includes the following steps: First, the pressure values ​​corresponding to the pressure signals collected by each sensor at different times are arranged in chronological order. For example, for the i-th sensor arranged along the pipeline axis, the pressure values ​​collected at times t1, t2, ..., tm are Pi1, Pi2, ..., Pim, respectively. These values ​​constitute a pressure sequence arranged in chronological order.

[0034] Subsequently, a two-dimensional real-time spatiotemporal feature matrix is ​​constructed using the spatial position order of the sensors along the pipeline axis as the row index and the time order as the column index. Assuming there are N sensors arranged on the pipeline and data is collected at M time points, the resulting real-time spatiotemporal feature matrix has a dimension of N×M. In this matrix, the element in the i-th row and j-th column represents the pressure value collected by the sensor at the i-th spatial position at the j-th time.

[0035] The core advantage of this construction method lies in its transformation of discrete one-dimensional time-series data into a two-dimensional data structure with a clear spatial topological relationship. In the field of pipeline leak detection, the pressure waves generated by a leak propagate along the pipeline axis, resulting in a strong spatial correlation between pressure signals from adjacent sensors. By mapping spatial locations to row indices, data in adjacent rows of the matrix are also physically adjacent, thus fully preserving the spatial correlation characteristics of the pressure signals. When subsequent processing is performed using convolutional neural networks, the convolutional kernels can slide along the row direction (spatial direction), efficiently extracting the pressure gradient features between adjacent sensors, which is crucial for identifying the location of leak points. Disrupting this continuity along the row direction would significantly increase the difficulty of the model extracting spatial features.

[0036] It should be understood that the above embodiments are illustrated using "spatial location as rows and time as columns" as an example, but this does not constitute a limitation of this application. In other embodiments, a transposed matrix can also be used, that is, the matrix is ​​constructed using the time order as the row index and the sensor spatial location order as the column index. For this transposed structure, the subsequent processing flow (such as the sliding direction of the convolution kernel) only needs to be adjusted accordingly to achieve the same technical effect, which also falls within the protection scope of this application.

[0037] Example 3: Based on step S300 in Example 1, this example details the specific process of generating a multi-channel spatiotemporal feature map based on each window matrix and the real-time spatiotemporal feature matrix.

[0038] Specifically, while single-dimensional raw pressure data contains basic waveform information, it often faces problems such as significant noise interference and insignificant feature representation in complex industrial environments. To address this issue, this embodiment constructs a three-channel spatiotemporal feature map containing multi-dimensional physical features through feature engineering. This process specifically includes the following steps: First, based on each window matrix, the pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are determined. This step aims to extract implicit spatial correlation and temporal dynamic features from the raw data.

[0039] Subsequently, the pressure difference sequences between adjacent sensors and the pressure change rate sequences of each sensor are stacked with the pressure waveform data of the corresponding time windows to generate a three-channel spatiotemporal feature map. This stacking method is similar to the process of converting a grayscale image into a color image, which greatly enriches the feature dimensions of the data.

[0040] For each window matrix, perform the following steps S31 to S34 respectively: S31: Use this window matrix as the first feature channel. This first feature channel retains the original pressure waveform data, reflecting the changes in absolute pressure values ​​of each sensor along the pipeline axis within the time window. It is the basis for subsequent feature extraction and contains the most original physical information.

[0041] S32: Calculate the pressure signal difference between two adjacent sensors along the pipe axial direction in the window matrix, forming an adjacent sensor pressure difference sequence, which is used as the second feature channel. Specifically, assuming the element in the i-th row and j-th column of the window matrix is ​​P(i,j), then the element at the corresponding position in the adjacent sensor pressure difference sequence is P(i+1,j)-P(i,j). This second feature channel reflects the gradient distribution of the pressure signal in the spatial dimension. Physically, when a leak occurs at a certain point in the pipeline, the pressure distribution on both sides of the leak point will show a significant difference, leading to a sudden change in the spatial gradient. Therefore, introducing this channel can significantly enhance the spatial feature representation of the leak point location, making it easier for the model to locate the specific pipe segment where the leak occurs.

[0042] S33: Calculate the pressure change rate of each sensor on the time axis within this window matrix, forming a pressure change rate sequence for each sensor, and use this as the third feature channel. Specifically, the change rate can be approximated by calculating the pressure difference between adjacent moments of the same sensor, i.e., V(i,j) = P(i,j+1) - P(i,j). This third feature channel reflects the dynamic characteristics of the pressure signal in the time dimension. Leakage is usually accompanied by a rapid drop or fluctuation in pressure, and this transient characteristic is particularly evident in the change rate sequence. Compared to the original pressure waveform, the change rate sequence is more sensitive to sudden leaks and can effectively capture the onset time and fluctuation intensity of the leak.

[0043] S34: Stack the first feature channel, the second feature channel, and the third feature channel along the channel dimension to generate a three-channel spatiotemporal feature map.

[0044] In this embodiment, this stacking operation can be implemented using tensor concatenation. For example, assuming the dimension of each window matrix is ​​N×M (N is the number of sensors, and M is the number of time sampling points), the dimension of the generated three-channel spatiotemporal feature map is N×M×3. It should be understood that this data structure perfectly fits the input requirements of deep convolutional neural networks (CNNs). The convolutional kernels of CNNs can automatically learn the weights of different feature channels in the channel dimension, thereby fusing complementary information from the original pressure, spatial gradient, and temporal rate of change. For example, for slow pressure drift, the temporal rate of change channel may have a weak response, but the original pressure channel can provide baseline information; for high-frequency noise interference, the spatial gradient channel can be effectively suppressed through differential operations between adjacent sensors. Through this fusion of multi-dimensional features, the spatiotemporal feature map not only enhances noise resistance but also significantly improves the saliency of leakage features, laying a solid foundation for subsequent high-precision identification. Furthermore, although this embodiment describes the construction method of three channels in detail, in other embodiments, more feature channels can be introduced according to actual needs, such as the frequency domain features of the pressure signal, etc. This application does not strictly limit this.

[0045] Example 4: Based on step S400 in Example 1, this example details the specific process of identifying the spatiotemporal feature map to obtain a set of multiple leak point locations.

[0046] Specifically, this embodiment employs a deep convolutional neural network (CNN) as the core recognition model. This model learns the mapping relationship between leakage features and leakage states from a large number of historical samples through supervised learning. The processing specifically includes: using the configured deep convolutional neural network to process the spatiotemporal feature map, obtaining a multi-label classification vector with the same dimension as the number of pipe segments. Each label represents whether a leakage exists in the corresponding pipe segment and its leakage aperture level.

[0047] In practical implementation, the deep convolutional neural network configured in this way needs to have a specific structure to adapt to the spatiotemporal feature map in this application. As a preferred implementation, the deep convolutional neural network can adopt a variant structure similar to ResNet (Residual Network) or VGGNet. Specifically, the front end of the network typically contains several convolutional layers and pooling layers. The convolutional layers are responsible for extracting local features from the spatiotemporal feature map. For example, 3x3 or 5x5 convolutional kernels are used to slide along the spatial dimension (sensor distribution direction) and temporal dimension of the feature map to capture the texture features of the pressure waveform and the gradient change features between adjacent sensors. The pooling layers are used to reduce the feature dimensionality and enhance the translation invariance of the features, that is, no matter how small the leakage signal is translated on the time axis, the network can recognize the same features. The back end of the network typically connects several fully connected layers to map the extracted high-dimensional features to the output space.

[0048] It is particularly important to note that the "multi-label classification vector" output in this embodiment has a specific physical meaning. Assuming the pipeline is divided into N segments, the dimension of this vector is N (or N×K, where K is the number of aperture levels). Each element (or each group of elements) in the vector corresponds to the leakage state of a pipeline segment. Unlike traditional single-label classification, this embodiment employs a multi-label classification strategy, meaning that multiple positions in the vector can simultaneously be positive (or have high probability). For example, if leaks occur simultaneously in the 5th and 8th segments of the pipeline, the values ​​corresponding to positions 5 and 8 in the vector will be significantly higher than those at other positions. Furthermore, the magnitude of this value or the label category can be further mapped to the leakage aperture level (e.g., micro-leak, small leak, large leak). This output format perfectly matches the actual working conditions where multiple leaks coexist, enabling the output of the location and severity information of all leak points at once, greatly improving detection efficiency.

[0049] Based on this multi-label classification vector, a set of multiple leak location points is determined, along with the corresponding pipe segment and its leakage aperture level for each leak point. Specifically, the system parses the vector output by the network, extracts the indexes of all segments with probability values ​​exceeding a preset threshold, and combines them with their corresponding aperture level labels to generate the final alarm information set. This process transforms the abstract neural network output into intuitive engineering data, providing direct decision support for operations and maintenance personnel.

[0050] Example 5: Based on the multi-label classification vector output in Example 4, this example details the post-processing steps for determining the set of multiple leak locations. Specifically, determining the set of multiple leak locations based on the multi-label classification vector includes: performing non-maximum suppression processing on the multi-label classification vector to suppress alarms for pipe segments that are spatially adjacent and have a non-maximum leakage probability, and outputting the final set of multiple leak locations.

[0051] It should be understood that in actual pipeline monitoring scenarios, the pressure waves generated by leaks often cover a certain spatial range, causing multiple adjacent sensors near the leak point to detect abnormal signals simultaneously. After processing by a deep convolutional neural network, this spatial signal overlap often manifests as several consecutive adjacent segments in a multi-label classification vector having a high probability of leakage. If all segments with probabilities exceeding a threshold are directly output as alarm points, the same leak point will trigger repeated alarms in multiple adjacent segments, causing confusion for maintenance personnel and potentially leading to misjudgments of the number of leaks. Therefore, introducing non-maximum suppression (NMS) is crucial. Its core logic is "survival of the fittest," meaning that among a set of spatially adjacent candidate segments, only the one with the highest probability is retained as the final alarm point, thereby eliminating redundancy.

[0052] Specifically, non-maximum suppression is performed on the multi-label classification vector, including the following steps: Step S51: Extract the leakage probability corresponding to each pipe segment from the multi-label classification vector output by the deep convolutional neural network. This step transforms the complex vector output by the neural network into a one-dimensional probability sequence, where each element corresponds to the leakage confidence of a pipe segment.

[0053] Step S52: Extract all pipe segments with a leakage probability greater than a preset first threshold as candidate segments, and sort them from highest to lowest leakage probability. The preset first threshold is used to filter out interference items with extremely low confidence, reducing computational load; sorting ensures that subsequent processing always prioritizes the segments with the highest probability.

[0054] Step S53: Traverse the sorted candidate segments. For the current candidate segment, find its spatially adjacent candidate segments along the pipe axis. Spatially adjacent segments are defined as the index difference between two segments not exceeding 1 or the physical distance between them on the pipe is less than one segment length.

[0055] It is important to clarify the definition of "spatial adjacency" here. This embodiment provides two parallel determination criteria to adapt to different engineering needs. The first is based on index difference, that is, if the absolute value of the difference between the index numbers of two segments does not exceed 1 (e.g., segment 5 and segment 6), then they are considered spatially adjacent. This method is simple to calculate and suitable for standard pipeline scenarios with uniform sensor placement and consistent segment lengths. The second is based on physical distance, that is, calculating the actual physical distance between the center points of two segments on the pipeline; if this distance is less than one segment length, then they are considered adjacent. This method is more flexible and can adapt to complex scenarios with uneven sensor placement or large differences in segment lengths, preventing missed suppression due to excessively short segment lengths or excessive suppression due to excessively long segment lengths. Those skilled in the art can choose one or use a combination of them according to the actual pipeline layout.

[0056] Step S54: If the leakage probability of an adjacent candidate segment is less than the leakage probability of the current candidate segment, then suppress the alarm output of that adjacent segment; otherwise, retain the one with the higher probability.

[0057] This step embodies the core suppression logic of NMS. Since candidate segments are sorted from high to low probability, the segment currently being traversed must be the one with the highest probability among the currently unsuppressed segments. Therefore, if its neighboring segments have a lower probability, it means that the alarms of the neighboring segments are likely "side effects" caused by the same leak point and should be suppressed.

[0058] Step S55: Output the unsuppressed candidate segments as the final set of multiple leak locations.

[0059] To better understand the necessity of this step, consider a specific defensive example: Suppose a leak occurs at kilometer 10 of a pipeline, and the pressure wave generated by this leak strongly affects sensors at kilometers 9, 10, and 11. After CNN identification, the probability of a leak is 0.75 for the 9-kilometer segment, 0.92 for the 10-kilometer segment, and 0.78 for the 11-kilometer segment. Without NMS processing, the system would output three alarms simultaneously, misleading maintenance personnel into believing that three leaks have occurred. However, by applying the NMS logic of this embodiment, the system first selects the 10-kilometer segment with the highest probability, then finds that the 9-kilometer and 11-kilometer segments are spatially adjacent and have lower probabilities, thus suppressing the latter two. Ultimately, the system only outputs the 10-kilometer segment as the leak point, accurately reconstructing the single-point leak fact at the site. This process effectively solves the problem of "false alarm merging" in multi-leak identification, significantly improving the accuracy and usability of the identification results.

[0060] Corresponding to the above method, this application also provides a multi-leakage point identification device based on spatiotemporal feature maps and convolutional neural networks, such as... Figure 2 As shown, the device includes: The acquisition unit 210 is used to continuously acquire pressure signals from each sensor along the axial direction of the pipeline and acquire a real-time spatiotemporal feature matrix with spatial position as rows and time as columns. The interception unit 220 is used to perform sliding time window interception on the real-time spatiotemporal matrix to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. The generation unit 230 is used to generate multi-channel spatiotemporal feature maps based on each window matrix and the real-time spatiotemporal feature matrix. The identification unit 240 is used to identify the spatiotemporal feature map to obtain a set of locations of multiple leaks, which includes the pipe section corresponding to each leak and its leakage aperture level.

[0061] The functions of each functional unit in the multi-leakage identification device based on spatiotemporal feature maps and convolutional neural networks provided in the above embodiments of this application can be implemented through the above-described method steps. Therefore, the specific working process and beneficial effects of each unit in the multi-leakage identification device based on spatiotemporal feature maps and convolutional neural networks provided in the embodiments of this application will not be repeated here.

[0062] This application also provides an electronic device, such as... Figure 3 As shown, it includes a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340.

[0063] Memory 330 is used to store computer programs; When the processor 310 executes the program stored in the memory 330, it performs the following steps: Pressure signals from each sensor along the pipeline axis are continuously collected to obtain a real-time spatiotemporal feature matrix with spatial position as rows and time as columns; The real-time spatiotemporal matrix is ​​subjected to a sliding time window to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. Based on each window matrix and the real-time spatiotemporal feature matrix, a multi-channel spatiotemporal feature map is generated; The spatiotemporal feature map is identified to obtain a set of multiple leak locations, which includes the pipe section corresponding to each leak and its leakage orifice size level.

[0064] The communication bus mentioned above 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 diagram, but this does not mean that there is only one bus or one type of bus.

[0065] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0066] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0067] The processors mentioned above can be general-purpose processors, including central processing units (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.

[0068] The implementation methods and beneficial effects of the various components of the electronic device in the above embodiments for solving the problem can be found in [reference needed]. Figure 1 The steps in the illustrated embodiments are used to implement the electronic device. Therefore, the specific working process and beneficial effects of the electronic device provided in this application will not be repeated here.

[0069] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform any of the multi-leakage point identification methods based on spatiotemporal feature maps and convolutional neural networks described in the above embodiments.

[0070] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the multi-leakage point identification methods based on spatiotemporal feature maps and convolutional neural networks described in the above embodiments.

[0071] Those skilled in the art will understand that the embodiments in this application can be provided as methods, systems, or computer program products. Therefore, the embodiments in this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments in this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. 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 apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, 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.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function 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.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable 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.

[0075] Although preferred embodiments have been described in this application, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of this application.

[0076] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of the embodiments of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims in this application and their equivalents, then this application also intends to include these modifications and variations.

Claims

1. A method for identifying multiple leak points based on spatiotemporal feature maps and convolutional neural networks, characterized in that, The method includes: Pressure signals from each sensor along the pipeline axis are continuously collected to obtain a real-time spatiotemporal feature matrix with spatial position as rows and time as columns; The real-time spatiotemporal matrix is ​​subjected to a sliding time window to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. Based on each window matrix and the real-time spatiotemporal feature matrix, a multi-channel spatiotemporal feature map is generated; The spatiotemporal feature map is identified to obtain a set of multiple leak locations, which includes the pipe section corresponding to each leak and its leakage orifice size level.

2. The method as described in claim 1, characterized in that, Obtain the real-time spatiotemporal feature matrix arranged in spatial order (rows) and temporal order (columns), including: Arrange the pressure values ​​corresponding to the pressure signals collected by each sensor at different times in chronological order; A two-dimensional real-time spatiotemporal feature matrix is ​​constructed using the spatial position sequence of the sensors along the pipeline axis as the row index and the time sequence as the column index.

3. The method as described in claim 1, characterized in that, Based on each window matrix and the real-time spatiotemporal feature matrix, a multi-channel spatiotemporal feature map is generated, including: Based on each window matrix, the pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are determined; The pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are stacked with the pressure waveform data of the corresponding time window to generate a three-channel spatiotemporal feature map.

4. The method as described in claim 3, characterized in that, The pressure difference sequence between adjacent sensors and the pressure change rate sequence of each sensor are stacked with the pressure waveform data of the corresponding time window to generate a three-channel spatiotemporal feature map, including: For each window matrix, perform the following steps: Use this window matrix as the first feature channel; Calculate the pressure signal difference between two adjacent sensors along the pipe axial direction in the window matrix to form a sequence of pressure differences between adjacent sensors, and use it as the second feature channel; Calculate the pressure change rate of each sensor on the time axis in the window matrix to form a pressure change rate sequence of each sensor, and use it as the third feature channel. The first, second, and third feature channels are stacked along the channel dimension to generate a three-channel spatiotemporal feature map.

5. The method as described in claim 1, characterized in that, The spatiotemporal feature map is identified to obtain a set of multiple leak location points, the corresponding pipe segment and leakage orifice size level for each leak point, including: Using a configured deep convolutional neural network, the spatiotemporal feature map is processed to obtain a multi-label classification vector with the same dimension as the number of pipeline segments. Each label represents whether there is a leak in the corresponding pipeline segment and the level of the leak aperture. Based on this multi-label classification vector, the set of multiple leak locations is determined, and the corresponding pipeline segment and leakage aperture level of each leak point are identified.

6. The method as described in claim 5, characterized in that, Based on this multi-label classification vector, the set of multiple leak location locations is determined, including: Non-maximum suppression is performed on the multi-label classification vector to suppress pipe segment alarms where the spatially adjacent pipes have a non-maximum leakage probability, and the final set of multiple leak location is output.

7. The method as described in claim 6, characterized in that, Non-maximum suppression is performed on the multi-label classification vector, including: Extract the leakage probability corresponding to each pipeline segment from the multi-label classification vector output by the deep convolutional neural network; Pipe segments with a leakage probability greater than a preset first threshold are extracted as candidate segments and sorted from high to low leakage probability. Traverse the sorted candidate segments. For the current candidate segment, find its spatially adjacent candidate segments along the pipe axis. Spatially adjacent is defined as the index difference between two segments not exceeding 1 or the physical distance between them on the pipe is less than one segment length. If the leakage probability of an adjacent candidate segment is less than the leakage probability of the current candidate segment, then suppress the alarm output of that adjacent segment; otherwise, retain the one with the higher probability. The unsuppressed candidate segments are output as the final set of multiple leak locations.

8. A multi-leakage point identification device based on spatiotemporal feature maps and convolutional neural networks, characterized in that, The device includes: The acquisition unit is used to continuously collect pressure signals from each sensor along the axial direction of the pipeline and obtain a real-time spatiotemporal feature matrix with spatial position as rows and time order as columns. The interception unit is used to perform sliding time window interception on the real-time spatiotemporal matrix to obtain multiple window matrices; each window matrix contains the pressure waveform data of all sensors within the time window. The generation unit is used to generate multi-channel spatiotemporal feature maps based on each window matrix and the real-time spatiotemporal feature matrix. The identification unit is used to identify the spatiotemporal feature map to obtain a set of locations of multiple leaks, which includes the pipe section corresponding to each leak and its leakage orifice size level.

9. An electronic device, characterized in that, The electronic device includes 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; A processor, when executing a program stored in memory, implements the method of any one of claims 1-7.

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

Images