A pipeline leakage identification method, device and equipment based on a time sequence adaptive graph and a readable storage medium

Abstract

The application discloses a pipeline leakage identification method based on a timing adaptive graph, and belongs to the technical field of pipeline safety monitoring. The method comprises the following steps: acquiring multi-source timing monitoring samples in the pipeline operation process, and constructing a continuous short-time window sequence; taking the response characteristics of each sensor in the current window as a synchronous flow, taking the response characteristics of adjacent windows with a time offset relationship as a time delay flow, jointly representing the synchronous flow and the time delay flow to obtain a synchronous time delay double-flow input; performing multi-time receptive field feature extraction on the synchronous time delay double-flow input to obtain dynamic characteristics corresponding to multiple time scales, and constructing an adaptive graph structure corresponding to each time scale respectively; inputting the dynamic characteristics of each scale into the adaptive graph structure of the corresponding scale for node information propagation to obtain graph features of each scale; and fusing the graph features of each scale to obtain a multi-scale fusion graph feature, and identifying pipeline leakage based on the multi-scale fusion graph feature. The application can improve the reliability of pipeline leakage identification.

Description

Technical Field

[0001] This application relates to the field of pipeline safety monitoring technology, and in particular to a pipeline leak identification method, apparatus, equipment, and readable storage medium based on a time-series adaptive graph. Background Technology

[0002] With the continuous development of long-distance oil and gas pipeline networks, pipeline leak detection is of great significance for ensuring the safe operation of pipelines. Since pipeline leaks are usually accompanied by changes in multi-source monitoring signals such as pressure, vibration, and acoustics, leak condition identification based on multi-source monitoring data has become an important technical means in the field of pipeline condition monitoring.

[0003] Existing pipeline leak detection methods typically perform feature analysis on collected monitoring signals and combine this with preset criteria or recognition models to determine whether a leak has occurred. However, in actual operation, factors such as the pipeline's environment, the state of the transported medium, and equipment operating conditions can affect the monitoring signals, making leak signals susceptible to noise interference. Furthermore, the performance of monitoring signals varies under different leak scenarios, resulting in insufficient recognition stability of existing methods under complex operating conditions.

[0004] Therefore, how to improve the reliability of pipeline leak identification based on multi-source monitoring data to meet the leak detection needs under different operating conditions has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this application is to provide a pipeline leak identification method, apparatus, device, and readable storage medium based on a time-series adaptive graph, so as to improve the reliability of leak identification.

[0006] To solve the above-mentioned technical problems, this application provides the following technical solution: A pipeline leak identification method based on time-series adaptive graphs includes: Acquire multi-source time-series monitoring samples during pipeline operation, and construct a continuous short-time window sequence based on the multi-source time-series monitoring samples; The response features of each sensor within the current continuous window are taken as the synchronization stream, and the response features of adjacent windows that have a time offset relationship with the synchronization stream are taken as the time delay stream. The synchronization stream and the time delay stream are jointly represented to obtain the synchronous and time delay dual-stream input. Multi-temporal receptive field feature extraction is performed on the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales, and adaptive graph structures corresponding to different time scales are constructed for the dynamic features under different time scales respectively. The dynamic features at each time scale are input into the corresponding adaptive graph structure for node information propagation to obtain graph features at different time scales. Multiple graph features are fused to obtain multi-scale fused graph features, and pipeline leaks are identified based on these multi-scale fused graph features to obtain pipeline leak identification results.

[0007] Preferably, constructing a continuous short-time window sequence based on the multi-source time-series monitoring samples includes: The multi-source time-series monitoring samples are divided according to a preset window length and sliding step size to obtain multiple time-series windows; Multiple time-series windows are combined in chronological order to obtain a continuous short-time window sequence that includes information about the current window and adjacent windows.

[0008] Preferably, the multi-temporal receptive field feature extraction of the synchronous delay dual-stream input includes: The synchronous delay dual-stream input is processed by channel mapping to obtain a fused feature representation that integrates the current window response information and the change information of adjacent windows; Parallel feature extraction is performed on the fused feature representation under multiple different time receptive fields to obtain short-timescale dynamic features, medium-timescale dynamic features, and long-timescale dynamic features.

[0009] Preferably, the step of constructing adaptive graph structures corresponding to different time scales for dynamic features includes: Use at least one of pressure sensors, vibration sensors, and acoustic sensors as graph nodes; For dynamic features at different time scales, corresponding learnable node representations are set, and scale-related mapping is used to perform feature transformation on the learnable node representations to obtain node relationship information at the corresponding time scale. Based on the node relationship information, the relationships between different graph nodes are interacted to determine the degree of association between different graph nodes, and an adaptive adjacency matrix corresponding to different time scales is generated based on the degree of association.

[0010] Preferably, generating adaptive adjacency matrices for different time scales based on the degree of association includes: Generate a node association matrix based on the degree of directional association between any two graph nodes; The node association matrix is ​​nonnegated and then sorted according to the degree of association between nodes; By retaining a preset number of highly associated nodes for each graph node and setting the association relationships between other nodes to zero, a sparse adaptive adjacency matrix is ​​obtained.

[0011] Preferably, the step of inputting the dynamic features at each time scale into the corresponding adaptive graph structure for node information propagation includes: The short-timescale dynamic features, medium-timescale dynamic features, and long-timescale dynamic features are respectively input into their corresponding adaptive graph structures for graph information propagation to obtain the graph propagation features of the corresponding timescales. The graph propagation features corresponding to different time scales are fused, and residual features generated based on the synchronous delay dual-stream input are introduced to obtain multi-scale fused graph features.

[0012] Preferably, pipeline leak identification based on the multi-scale fusion map features includes: Wavelet scattering transform is performed on the window signals corresponding to the multi-source time-series monitoring samples to obtain multi-scale time-frequency texture features; The multi-scale time-frequency texture features are fused with the multi-scale fusion map features to obtain comprehensive recognition features; The pipeline leakage status is determined based on the comprehensive identification features, and the pipeline leakage identification result is output.

[0013] A pipeline leak identification device based on a time-series adaptive graph includes: The data acquisition module is used to acquire multi-source time-series monitoring samples corresponding to the pipeline operation process, and to construct a continuous short-time window sequence based on the multi-source time-series monitoring samples; The dual-stream construction module is used to take the response features of each sensor in the current continuous window as a synchronous stream, take the response features of adjacent windows that have a time offset relationship with the synchronous stream as a time delay stream, and jointly represent the synchronous stream and the time delay stream to obtain synchronous and time delay dual-stream input; The dynamic feature extraction module is used to extract multi-temporal receptive field features from the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales. The adaptive graph construction module is used to construct adaptive graph structures for different time scales based on dynamic features. The graph feature fusion module is used to input the dynamic features at each time scale into the corresponding adaptive graph structure for node information propagation, obtain graph features corresponding to different time scales, and fuse multiple graph features to obtain multi-scale fused graph features. The identification module is used to identify pipeline leaks based on the features of the multi-scale fusion map, and obtain pipeline leak identification results.

[0014] An electronic device, comprising: Memory, used to store computer programs; A processor is used to implement the steps of the pipeline leak identification method based on timing adaptive graphs described above when executing the computer program.

[0015] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the pipeline leak identification method based on a time-adaptive graph described above.

[0016] The method provided in this application includes: acquiring multi-source time-series monitoring samples during pipeline operation, and constructing a continuous short-time window sequence based on the multi-source time-series monitoring samples; taking the response features corresponding to each sensor within the current continuous window as a synchronous flow, taking the response features of adjacent windows that have a time offset relationship with the synchronous flow as a time-delay flow, and jointly representing the synchronous flow and the time-delay flow to obtain a synchronous-time-delay dual-flow input; extracting multi-time receptive field features from the synchronous-time-delay dual-flow input to obtain dynamic features corresponding to multiple time scales, and constructing an adaptive graph structure corresponding to the time scale for the dynamic features under different time scales; inputting the dynamic features under each time scale into the adaptive graph structure of the corresponding time scale for node information propagation to obtain graph features corresponding to different time scales; fusing multiple graph features to obtain multi-scale fused graph features, and performing pipeline leak identification based on the multi-scale fused graph features to obtain a pipeline leak identification result.

[0017] Because the response process of a leakage event exhibits significant temporal evolution characteristics, pressure changes, vibration propagation, and acoustic responses at different stages may display dynamic patterns at different time scales. A single time scale feature is insufficient to fully describe the complex changes during the leakage process. To adapt to the dynamic changes in sensor correlations at different time stages and response scales during a pipeline leakage event, this application constructs a continuous short-time window sequence based on multi-source time-series monitoring samples when building the input. Then, the response features corresponding to each sensor within the current continuous window are used as a synchronization stream, and the response features of adjacent windows with time offset relationships with the synchronization stream are used as a time-delay stream. The synchronization stream and time-delay stream are jointly represented to obtain a synchronous-time-delay dual-stream input. In other words, the synchronous-time-delay dual-stream input carries not only the response features corresponding to each sensor within the current continuous window but also the response features of adjacent windows with time offset relationships with the synchronization stream. That is, the synchronous-time-delay dual-stream input of this application allows the model input to simultaneously include the synchronization state information at the current moment and the evolutionary trend information between short-time windows, thereby overcoming the limitation of traditional single-window static features in insufficiently representing the dynamic process of leakage and providing richer temporal priors for subsequent multi-scale dynamic modeling.

[0018] Furthermore, considering that a single-scale graph structure can typically only describe node association information at a fixed scale, it is prone to the overlapping of leakage response features at different time scales, reducing the recognition accuracy in complex leakage scenarios. To address this, this application performs parallel feature extraction across multiple temporal receptive fields on the synchronous delay dual-stream input, obtaining dynamic features corresponding to different time scales. An adaptive graph structure is then independently constructed for each time scale, enabling sensor associations at different scales to learn independently within their respective matched graph topologies. Compared to using a single fixed graph or a shared adaptive graph, this approach avoids the overlapping of association patterns at different time scales, allowing features at each scale to be represented within their most suitable association space, thereby improving the accuracy of characterizing multi-scale dynamic changes in leakage signals.

[0019] Then, the dynamic features at each scale are input into the corresponding adaptive graph structure for node information propagation, enabling the directed transmission of sensor correlation information at each scale within the corresponding graph structure. Multi-scale fused graph features are obtained by fusing graph features from multiple scales. This fused feature integrates multi-level information such as instantaneous disturbances, continuous fluctuations, and overall trends, significantly improving the model's ability to distinguish complex operating conditions and weak leakage signals, ultimately enhancing the reliability of pipeline leakage status identification.

[0020] Accordingly, embodiments of this application also provide a pipeline leakage identification device, equipment, and readable storage medium based on a time-adaptive graph, which corresponds to the above-described pipeline leakage identification method based on a time-adaptive graph and has the aforementioned technical effects, which will not be elaborated further here. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an implementation of a pipeline leak identification method based on a time-adaptive graph in this application. Figure 2 It is a fixed diagram; Figure 3 This is a schematic diagram illustrating the construction of an adaptive graph at different time scales in an embodiment of this application; Figure 4 This is a schematic diagram of a short time scale in an embodiment of this application; Figure 5 This is a schematic diagram of a time scale in one embodiment of this application; Figure 6This is a schematic diagram of a long time scale in an embodiment of this application; Figure 7 This is a flowchart illustrating a specific implementation of a pipeline leak identification method based on a time-adaptive graph in this application. Figure 8 This is a block diagram illustrating the main principle of a time-series enhanced multi-scale adaptive graph in an embodiment of this application. Figure 9 This is a schematic diagram of the structure of a pipeline leak identification device based on a time-adaptive graph in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application; Figure 11 This is a schematic diagram of the specific structure of an electronic device in an embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] To facilitate understanding, the relevant technical terms used in this article are explained below.

[0025] Pipeline operation monitoring data refers to continuous time series data collected by pressure sensors, acceleration sensors, acoustic sensors, etc., arranged in the pipeline system. It is used to reflect the pressure fluctuations, vibration response, and acoustic change characteristics of the pipeline under different operating conditions.

[0026] Synchronous-delay dual-stream: This refers to a continuous window feature organization method composed of a synchronization stream and a delay stream. The synchronization stream represents the synchronous response characteristics of different sensors within the current continuous time window, while the delay stream represents the response characteristics of adjacent windows that have a time offset relative to the synchronization stream. The two are jointly represented along the feature channel dimension to enhance the model's ability to express the current response and the delay variations of adjacent windows.

[0027] Multi-temporal receptive field feature extraction: This refers to setting up multiple convolutional branches with different time scales in the main branch to process the synchronous time-delay dual-stream features in parallel, so as to extract dynamic change features at short, medium, and long scales respectively. This process is used to capture instantaneous disturbances, continuous fluctuations, and overall changes within multiple window ranges in leaked signals.

[0028] Multi-scale adaptive graph construction: This refers to the process of constructing adaptive graph structures for short-scale, medium-scale, and long-scale applications by using sensors such as pressure, vibration, and acoustic sensors as graph nodes and combining the dynamic features of the sensors extracted from different time-sensory fields. In its implementation, this application generates adaptive adjacency relationships between sensors at different time scales through learnable node representations, scale-related mappings, and interactive modeling of node relationships. This is used to characterize the sensor association strength corresponding to instantaneous disturbances, continuous fluctuations, and the overall response within multiple window ranges.

[0029] Scale-corresponding graph propagation refers to the process of inputting sensor dynamic features at short, medium, and long scales into corresponding adaptive graph structures for node information propagation. This propagation process still uses sensors such as pressure, vibration, and acoustic sensors as graph nodes, but different scale features correspond to different graph structures. This allows short-term disturbance features, medium-term fluctuation features, and long-term trend features to propagate in their respective matched sensor correlation graphs, thus avoiding the problem of overlapping correlations at different scales caused by using a single graph structure.

[0030] Multi-scale graph feature fusion refers to the fusion of graph propagation results at short, medium, and long scales to form a comprehensive graph feature of the main branches. This fusion process can introduce residual branches and integrate sensor-related features at different scales using methods such as gated fusion, weighted fusion, or feature stitching.

[0031] Wavelet scattering enhancement branch: This refers to a parallel feature enhancement branch set outside the main trunk of the time-series enhanced multi-scale adaptive graph. It is used to directly perform wavelet scattering transformation on the preprocessed original window signal to extract stable multi-scale time-frequency texture features, so as to make up for the problem that statistical time-frequency features are insufficient in expressing local fluctuations, non-stationary textures and multi-scale details.

[0032] Main scattering dual-path fusion: refers to the joint representation of multi-scale map features extracted from the multi-scale adaptive map main branch and multi-scale time-frequency texture features extracted from the wavelet scattering enhancement branch, and the identification of normal pipeline status and different leakage types based on the fused features.

[0033] Leakage status identification: refers to the process of determining whether a pipeline is in a normal state based on pipeline operation monitoring data, and identifying different types of leaks or fault states.

[0034] Please refer to Figure 1 , Figure 1 This is a flowchart of a pipeline leak identification method based on a time-adaptive graph according to an embodiment of this application. The method includes the following steps.

[0035] S101. Obtain multi-source time-series monitoring samples during pipeline operation, and construct a continuous short-time window sequence based on the multi-source time-series monitoring samples.

[0036] Pipelines can refer to pipelines within a pipeline system, which is a critical infrastructure in oil and gas transportation, underwater production systems, and urban water supply systems. Pipeline leaks can cause not only media loss and equipment damage but also environmental pollution and safety incidents.

[0037] The acquisition process of multi-source time-series monitoring samples may include collecting time-series data on pressure, vibration, and acoustic dimensions under normal pipeline operation and different leakage conditions to form multi-source pipeline operation monitoring data. This data is then preprocessed to obtain multi-source time-series monitoring samples. Pressure data reflects internal pipeline pressure fluctuations, vibration data reflects the pipe wall structure response, and acoustic data reflects acoustic changes caused by leakage disturbances. Optionally, the multi-source pipeline operation monitoring data may also include temperature data and facility status data, which are related to pipeline leakage.

[0038] The process involves preprocessing the acquired multi-source monitoring signals, including reading the raw monitoring data from each sensor and extracting the effective signal sequences representing the actual monitoring values. The data from the pressure and acceleration sensors are read as continuous numerical sequences, while the acoustic sensor data is converted into a computable time-series signal according to its original storage format. Mean reduction processing is performed on each channel signal to weaken the impact of DC bias and zero-point drift on subsequent feature extraction. Pressure, vibration, and acoustic signals from different sampling frequencies are resampled to the same target sampling frequency, ensuring that all sensor signals are on the same time scale. Finally, the resampled channel signals are truncated to the shortest effective length to maintain consistency in the time dimension, resulting in a multi-source monitoring sequence that is time-synchronized, of consistent length, and uniform in format.

[0039] After obtaining multi-source time-series monitoring samples, a continuous short-time window sequence can be constructed.

[0040] In one specific embodiment of this application, constructing a continuous short-time window sequence based on multi-source time-series monitoring samples includes: The multi-source time-series monitoring samples are divided according to the preset window length and sliding step size to obtain multiple time-series windows; Multiple time-series windows are combined in chronological order to obtain a continuous short-time window sequence that includes information about the current window and adjacent windows.

[0041] The time-synchronized multi-source monitoring sequences of pressure, vibration, and acoustics can be divided into multiple basic window samples according to a fixed window length and sliding step size. Each basic window corresponds to a continuous operating state and is used to characterize pressure fluctuations, pipe wall vibration response, and acoustic disturbance changes within that time period.

[0042] S102. Take the response features of each sensor in the current continuous window as the synchronization stream, take the response features of adjacent windows that have a time offset relationship with the synchronization stream as the time delay stream, and jointly represent the synchronization stream and the time delay stream to obtain the synchronous and time delay dual-stream input.

[0043] To avoid ambiguity, the synchronization stream in this application embodiment can also be called the synchronization response stream, which is used to characterize the instantaneous response characteristics of each sensor within the current continuous window; the time delay stream can also be called the time shift evolution stream, which is used to characterize the response characteristics of adjacent windows that have a time offset relationship with the synchronization stream, so as to represent the dynamic evolution relationship between adjacent time windows.

[0044] Based on the basic window partitioning, multiple consecutive basic windows are combined into a short-time window sequence sample. This sample retains the synchronous responses of multiple sensors within the current time period and provides adjacent window information for subsequent construction of the time-delay stream. Therefore, the model input is no longer limited to the static features of a single window, but can include the current response and time-delay variation information within consecutive windows.

[0045] That is, after obtaining multiple time-series windows, statistical time-frequency features are extracted for each time-series window, and the statistical time-frequency features are used as the response features of each sensor in the corresponding time-series window. The statistical time-frequency features include at least one of the following: mean, standard deviation, root mean square, peak value, maximum absolute value, skewness, kurtosis, peak factor, zero-crossing rate, spectral correlation features, and frequency band energy distribution, which are used to characterize the amplitude variation, frequency distribution, and energy evolution features of multi-source monitoring signals.

[0046] Statistical time-frequency features are extracted from each base window in the short-time window sequence to characterize the variation characteristics of pressure, vibration, and acoustic signals in the time and frequency domains. The extracted features include mean, standard deviation, root mean square, peak value, maximum absolute value, skewness, kurtosis, peak factor, zero-crossing rate, as well as spectral correlation features and frequency band energy distribution, used to characterize the amplitude variation, frequency distribution, and energy evolution of multi-source sensor signals.

[0047] Based on this, a synchronous time-delay dual-stream feature representation structure is constructed. The synchronous response stream feature Sk and the time-shifted evolution stream feature Dk are concatenated along the feature channel dimension to form the synchronous response and the time-shifted evolution dual-stream concatenated feature Hk: , where Concat( The ) indicates concatenation along the feature channel dimension. This concatenation operation only performs joint organization of the two types of features, without changing the number of graph nodes or the time series length, so that the model input simultaneously includes the synchronization state information at the current moment and the evolution trend information between short time windows. In practical applications, the two-stream concatenated features can also be weighted and combined through a channel mapping layer with a convolution kernel size of 1×1, and then subjected to nonlinear activation processing to obtain the two-stream fused features.

[0048] S103. Perform multi-temporal receptive field feature extraction on the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales, and construct adaptive graph structures for the corresponding time scales for the dynamic features under different time scales.

[0049] Considering that a single-scale graph structure can typically only describe node association information at a fixed scale, it is prone to the overlapping of leakage response features at different time scales, reducing the recognition accuracy in complex leakage scenarios. Therefore, this application's embodiments perform parallel feature extraction across multiple time receptive fields on synchronous delay dual-stream inputs to obtain dynamic features corresponding to different time scales. An adaptive graph structure is independently constructed for each time scale, enabling sensor association relationships at different scales to be learned independently within their respective matched graph topologies. Compared to using a single fixed graph or a shared adaptive graph, this avoids the overlapping of association patterns at different time scales, allowing features at each scale to be represented within their most suitable association space, thereby improving the accuracy of characterizing multi-scale dynamic changes in leakage signals.

[0050] In other words, learnable channel mapping can be used to weight and combine different channel features in the synchronous and delayed streams at each sensor node and time position, and then undergo nonlinear transformation to obtain a new fused feature Ek for the synchronous and delayed dual-stream channels. This feature retains the original sensor node arrangement and time order, but each of its output channels now contains both the current window response and the delayed window response, providing a unified input for subsequent dynamic feature extraction at different time scales.

[0051] In one specific embodiment of this application, multi-temporal receptive field feature extraction is performed on synchronous delay dual-stream input, including: Channel mapping processing is performed on the synchronous delay dual-stream input to obtain a fused feature representation that integrates the current window response information and the change information of adjacent windows; Parallel feature extraction is performed on the fused feature representation under multiple different temporal receptive fields to obtain short-term dynamic features, medium-term dynamic features, and long-term dynamic features.

[0052] In this embodiment, three relative temporal receptive fields—short-time, medium-time, and long-time—can be set (other different temporal receptive fields can also be set in practical applications, which will not be listed here). These fields are used to perform parallel temporal modeling of the new fusion feature Ek of the dual-channel streams. The short-time receptive field focuses on extracting local transient disturbances, the medium-time receptive field focuses on characterizing fluctuations within a certain time range, and the long-time receptive field focuses on depicting trend evolution over a longer time range. This yields the dynamic features corresponding to the short-time, medium-time, and long-time time scales, respectively.

[0053] The dynamic features obtained at each time scale maintain the correspondence between sensor nodes and are used for adaptive graph structure modeling at the corresponding scales, thereby achieving the matching and propagation of multi-scale dynamic features and scale-corresponding graph structures.

[0054] In one specific embodiment of this application, adaptive graph structures corresponding to different time scales are constructed for dynamic features at different time scales, including: Use at least one of pressure sensors, vibration sensors, and acoustic sensors as graph nodes; For dynamic features at different time scales, corresponding learnable node representations are set, and scale-related mapping is used to perform feature transformation on the learnable node representations to obtain node relationship information at the corresponding time scale. Based on node relationship information, the relationships between different graph nodes are interacted to determine the degree of association between different graph nodes, and an adaptive adjacency matrix corresponding to different time scales is generated based on the degree of association.

[0055] Among them, the generation of adaptive adjacency matrices for different time scales based on the degree of association includes: Generate a node association matrix based on the degree of directional association between any two graph nodes; The node association matrix is ​​nonnegated and then sorted according to the degree of association between nodes; By retaining a preset number of highly associated nodes for each graph node and setting the association relationships between other nodes to zero, a sparse adaptive adjacency matrix is ​​obtained.

[0056] The following uses sensor measurement points or sensor channels, such as pressure, vibration, and acoustic sensors, as nodes in the graph. The edges between nodes represent the information correlation between different sensor-monitored features, and the edge weights represent the degree of correlation between the corresponding nodes during the graph feature propagation process.

[0057] In a fixed diagram, the node connections are typically predetermined based on the sensor installation location, the distance between measuring points, or the pipeline topology. For example... Figure 2 As shown, assuming nodes v1 to v6 are sequentially set along the pipeline, according to the principle of physical proximity (e.g. Figure 2 (See the on-site installation diagram above). Connections such as v1 to v2, v2 to v3, v3 to v4, v3 to v5, and v5 to v6 can be pre-established to form a fixed diagram (e.g., ...). Figure 2The diagram below shows the pre-determined node connections based on pipe segment, topological adjacency, and relative distance. This diagram is fixed before model training, and the connections between nodes remain unchanged during training. While a fixed diagram can represent known spatial adjacencies between sensors, physical proximity does not necessarily indicate the strongest correlation of their monitoring characteristics during leak identification; sensors that are not physically adjacent may also exhibit strong characteristic correlations due to joint responses to the same leak disturbance. Therefore, pre-setting node connections solely based on physical location makes it difficult to adjust the diagram structure according to monitoring data and leak identification targets.

[0058] To reduce reliance on manually constructed graph rules, node representations that can be updated during model training can be set for all sensor nodes. The directed association degree between any two nodes is calculated, and the top k connections with the highest association degree for each node are retained, thus forming a single adaptive graph. For example, the association degree between node v1 and other nodes is calculated separately; if v1 has a high association degree with v3 and v4, the corresponding connection is retained, and the remaining connections with lower association degrees are deleted. After processing all nodes in the same way, an adaptive adjacency matrix jointly determined by the training data and the leakage classification target can be obtained.

[0059] While a single adaptive graph can learn node connectivity and edge weights during training, directly applying it to multiple temporal receptive fields requires sharing the same adjacency matrix for short-term, medium-term, and long-term dynamic features. In this case, the same node has the same neighboring nodes and propagation weights across different temporal receptive fields, making it difficult to represent potential sensor correlation differences across different timeframes.

[0060] To this end, this application embodiment sets two sets of shared learnable basic node representations for all sensor nodes, and sets mutually independent scale mapping parameters for different temporal receptive fields, thereby obtaining the node representations corresponding to each temporal receptive field, and constructing corresponding adaptive adjacency matrices. Let the set of temporal receptive field scales be R, and the two sets of shared learnable basic node representations be E1 and E2, respectively. For any temporal receptive field r∈R, the mapping function corresponding to that scale is used respectively. and Transforming the basic node representation yields: ; The corresponding adaptive adjacency matrix is ​​expressed as: Among them, E1 and E2 are two sets of learnable basic node representations shared by each temporal receptive field; and The node representation matrix obtained after scaling; matrix interaction terms. It is used to calculate the directional association between any two nodes; α is the association degree adjustment coefficient; ReLU is used to remove non-positive associations; Top-k is used to sort the association degree of each node and retain only the top k connections with the highest association degree. This is the final adaptive adjacency matrix of the corresponding time receptive field.

[0061] Specifically, such as Figure 3 As shown, the scale-adaptive graph is constructed according to the process of learnable node representation, scale-related mapping, node relationship interaction modeling, and sparsification of the associated structure.

[0062] First, each sensor measurement point or sensor channel is treated as a graph node, and a node representation that can be updated during model training is set for all nodes. This node representation does not rely on manually preset node connections, but is updated under the supervision of the leakage state identification task, and is used to characterize the potential correlation features between different sensor nodes.

[0063] Subsequently, for three time receptive fields—short-term, medium-term, and long-term—independent scale mapping parameters were used to transform the two sets of shared basic node representations, resulting in two sets of node representations for each of the short-term, medium-term, and long-term scales. Since different mapping parameters are used for different time receptive fields, the same sensor node can form different relational representations under different time receptive fields, providing a foundation for subsequent learning of scale-differentiated node connection relationships.

[0064] Within each time receptive field, interactive operations are performed on the relationship representations of any two nodes to calculate the directional association degree between the nodes, and the association degrees of all node pairs are used to form the corresponding association matrix. The greater the association degree, the stronger the information propagation relationship between the corresponding nodes within that time receptive field. Since the association degree between node vi and node vj can be different from the association degree between node vj and node vi, the resulting association structure is a directed structure.

[0065] The association matrix under each time receptive field is nonnegated, and the non-positive association degree is set to zero. Then, taking each node as the starting node, the non-negative directed association degree from that node to other nodes is sorted. At most k non-zero connections with high association degree are retained by Top-k selection, and the remaining connections are set to zero. In this way, the original dense node relationship is transformed into a sparse association structure, where the retained connections are directed edges in the graph, and the corresponding association degree is the edge weight.

[0066] Following the above process, short-time adaptive correlation structures, medium-time adaptive correlation structures, and long-time adaptive correlation structures are obtained respectively. To illustrate the differences between the three correlation structures, we still use six sensor nodes v1. Taking v6 as an example, the physical arrangement of the six nodes can be compared with... Figure 2 Maintain consistency, but do not use adaptive graphs. Figure 2 Instead of using pre-set fixed connections, this method calculates the correlation between any two nodes at short-term, medium-term, and long-term scales, and performs Top-k filtering on each node.

[0067] Let the Top-k filtering parameter k=2, meaning that each node retains at most two connections with the highest correlation. For ease of explanation, assume that node v1 has at least two positive correlations at each time scale. Taking node v1 as an example, if the directed correlation between v1 and v2 / v3 is the highest at a short time scale, then v1-v2 and v1-v3 are retained; if the directed correlation between v1 and v3 / v5 is the highest at a medium time scale, then v1-v3 and v1-v5 are retained; if the directed correlation between v1 and v4 / v6 is the highest at a long time scale, then v1-v4 and v1-v6 are retained. Filtering the remaining nodes in the same way will yield the following results: Figure 4 , Figure 5 and Figure 6 The short-term, medium-term, and long-term adaptive graphs are shown respectively. The above connectivity relationships are only used to illustrate the principle of scaled graph construction; the actual neighbor nodes and their edge weights retained are determined by model training.

[0068] After the graph is constructed, short-term dynamic features are mapped to short-term adaptive association structures, medium-term dynamic features to medium-term adaptive association structures, and long-term dynamic features to long-term adaptive association structures. In the next step, graph feature propagation corresponding to each scale is performed. Thus, short-term, medium-term, and long-term dynamic features no longer share the same graph structure, but instead transmit information according to the node relationships learned under the corresponding time receptive fields.

[0069] S104. Input the dynamic features at each time scale into the corresponding adaptive graph structure for node information propagation to obtain the graph features corresponding to different time scales.

[0070] After obtaining the dynamic features at different time scales and their corresponding adaptive adjacency matrices, the short-term, medium-term, and long-term dynamic features are input into the corresponding adaptive graphs for forward and reverse information propagation, respectively. The propagation results in the two directions are then combined to obtain the short-term, medium-term, and long-term graph propagation features (i.e., graph features).

[0071] In one specific embodiment of this application, dynamic features at each time scale are input into an adaptive graph structure corresponding to that time scale for node information propagation, including: The short-timescale dynamic features, medium-timescale dynamic features, and long-timescale dynamic features are respectively input into their corresponding adaptive graph structures for graph information propagation to obtain the graph propagation features of the corresponding timescales. The graph propagation features corresponding to different time scales are fused, and residual features generated based on synchronous delay dual-stream input are introduced to obtain multi-scale fused graph features.

[0072] In this embodiment, as Figure 7 As shown, a residual branch is set up to preserve the original temporal response information in the synchronous delay dual-stream input. This branch transforms the synchronous delay dual-stream input into residual node features consistent with the propagation feature dimensions of each scale graph through lightweight feature mapping and temporal convergence.

[0073] like Figure 8 As shown, in the backbone feature fusion process, the residual node features and the propagation features of short-term, medium-term, and long-term maps are first concatenated and their node dimensions are aggregated. Then, four gating coefficients are generated based on the aggregation results to adjust the contribution of the four features to the fusion result. Subsequently, the four features are gated and weighted, and then averaged to obtain a unified backbone node representation. Thus, samples from different short-term windows can adaptively adjust the degree of preservation of the original time-series information and the propagation information of different scale maps according to their own response characteristics.

[0074] Furthermore, global node relationship enhancement is performed on the backbone node representation, and the enhancement result is supplemented into the backbone node representation as a residual to extract the overall correlation information between different sensor nodes. Finally, average aggregation and maximum aggregation are performed on the enhanced node representation, and the two aggregation results are concatenated to obtain the backbone graph feature G corresponding to the k-th short-time window sample. k This is used for subsequent multi-branch feature fusion and leakage state identification.

[0075] S105. Fuse multiple graph features to obtain multi-scale fused graph features, and perform pipeline leak identification based on the multi-scale fused graph features to obtain pipeline leak identification results.

[0076] By fusing graph features corresponding to multiple time scales, a multi-scale fused graph feature can be obtained. This multi-scale fused graph feature is then used for pipeline leak identification to obtain the pipeline leak identification result.

[0077] In one specific embodiment of this application, pipeline leak identification based on multi-scale fusion map features includes: Wavelet scattering transform is performed on the window signals corresponding to multi-source time-series monitoring samples to obtain multi-scale time-frequency texture features; Multi-scale time-frequency texture features are fused with multi-scale fusion map features to obtain comprehensive recognition features; The system determines the pipeline leakage status based on comprehensive identification features and outputs the pipeline leakage identification results.

[0078] In this embodiment, in addition to the main temporal-enhanced multi-scale adaptive graph, a wavelet scattering enhancement branch can be set in parallel to extract multi-scale time-frequency texture features from the original window signal. This branch takes the preprocessed original signal corresponding to the k-th short-time window sequence as input and performs multi-scale time-frequency decomposition processing on multi-source sensor signals such as pressure, vibration, and acoustics to obtain local variation information and stable texture structure at different time scales.

[0079] During feature construction, the time-frequency decomposition results are subjected to stability enhancement processing to highlight the non-stationary changes and fine-grained structural features in the signal, thereby forming a robust multi-scale texture representation to characterize the differences in local fluctuations and frequency structure of the original window signal.

[0080] The wavelet scattering enhancement branch does not participate in the adaptive graph structure construction and propagation process. Instead, it acts as an independent texture feature extraction branch, complementing the backbone graph features (i.e., multi-scale fused graph features) in modeling.

[0081] The output features of the wavelet scattering enhancement branch are denoted as follows: , is used to represent the multi-scale time-frequency texture feature representation corresponding to the k-th short-time window sequence, and is used in the subsequent dual-path fusion module.

[0082] In obtaining the trunk plot features G k After combining the wavelet scattering enhancement feature Pk, the two types of features are input into the dual-path attention fusion module. Among them, the backbone map feature mainly represents the dynamic correlation information of multi-source sensors at different time scales, while the wavelet scattering enhancement feature mainly represents the stable multi-scale time-frequency texture information in the original window signal. The two feature charactersize the same short-time window sample from the perspectives of structural correlation and signal texture, respectively.

[0083] In the dual-path fusion process, the importance of the backbone features and wavelet scattering enhancement features to the current sample is first calculated using the feature scoring module. Then, the two scores are normalized to obtain the backbone path weight and the scattering path weight. Subsequently, the two types of features are weighted according to their respective weights, and the weighted features are concatenated to obtain the dual-path integrated representation. ,in, and These represent the fusion weights corresponding to the main path and the scattering path, respectively, and their sum is 1; This represents feature concatenation. Since the fusion weights are adaptively generated based on the two types of features of the current sample, samples from different short-term windows can have different dual-path fusion ratios. Finally, the dual-path is comprehensively represented. The input status recognition module obtains the classification results of whether the current short-term window belongs to the normal state or different leakage types, thereby realizing pipeline leakage status recognition.

[0084] The method provided in this application provides the following steps: Multi-source time-series monitoring samples are acquired during pipeline operation, and a continuous short-time window sequence is constructed based on these samples. The response features of each sensor within the current continuous window are used as a synchronous flow, and the response features of adjacent windows with a time offset relationship to the synchronous flow are used as a time-delay flow. The synchronous flow and the time-delay flow are jointly represented to obtain a synchronous-time-delay dual-flow input. Multi-time receptive field feature extraction is performed on the synchronous-time-delay dual-flow input to obtain dynamic features corresponding to multiple time scales. Adaptive graph structures corresponding to different time scales are constructed for the dynamic features at different time scales. The dynamic features at each time scale are input into the corresponding adaptive graph structures for node information propagation to obtain graph features corresponding to different time scales. Multiple graph features are fused to obtain multi-scale fused graph features, and pipeline leak identification is performed based on these multi-scale fused graph features to obtain pipeline leak identification results.

[0085] Because the response process of a leakage event exhibits significant temporal evolution characteristics, pressure changes, vibration propagation, and acoustic responses at different stages may display dynamic patterns at different time scales. A single time scale feature is insufficient to fully describe the complex changes during the leakage process. To adapt to the dynamic changes in sensor correlations at different time stages and response scales during a pipeline leakage event, this application constructs a continuous short-time window sequence based on multi-source time-series monitoring samples when building the input. Then, the response features corresponding to each sensor within the current continuous window are used as a synchronization stream, and the response features of adjacent windows with time offset relationships with the synchronization stream are used as a time-delay stream. The synchronization stream and time-delay stream are jointly represented to obtain a synchronous-time-delay dual-stream input. In other words, the synchronous-time-delay dual-stream input carries not only the response features corresponding to each sensor within the current continuous window but also the response features of adjacent windows with time offset relationships with the synchronization stream. That is, the synchronous-time-delay dual-stream input of this application allows the model input to simultaneously include the synchronization state information at the current moment and the evolutionary trend information between short-time windows, thereby overcoming the limitation of traditional single-window static features in insufficiently representing the dynamic process of leakage and providing richer temporal priors for subsequent multi-scale dynamic modeling.

[0086] Furthermore, considering that a single-scale graph structure can typically only describe node association information at a fixed scale, it is prone to the overlapping of leakage response features at different time scales, reducing the recognition accuracy in complex leakage scenarios. To address this, this application performs parallel feature extraction across multiple temporal receptive fields on the synchronous delay dual-stream input, obtaining dynamic features corresponding to different time scales. An adaptive graph structure is then independently constructed for each time scale, enabling sensor associations at different scales to learn independently within their respective matched graph topologies. Compared to using a single fixed graph or a shared adaptive graph, this approach avoids the overlapping of association patterns at different time scales, allowing features at each scale to be represented within their most suitable association space, thereby improving the accuracy of characterizing multi-scale dynamic changes in leakage signals.

[0087] Then, the dynamic features at each scale are input into the corresponding adaptive graph structure for node information propagation, enabling the directed transmission of sensor correlation information at each scale within the corresponding graph structure. Multi-scale fused graph features are obtained by fusing graph features from multiple scales. This fused feature integrates multi-level information such as instantaneous disturbances, continuous fluctuations, and overall trends, significantly improving the model's ability to distinguish complex operating conditions and weak leakage signals, ultimately enhancing the reliability of pipeline leakage status identification.

[0088] As can be seen, in this application, multi-source pipeline operation monitoring signals such as pressure, vibration, and acoustic signals are organized into short-time window samples containing synchronous response and time delay variations. In the main branch, dynamic feature extraction of multi-time sensing fields, adaptive graph construction for different time scales, and graph propagation based on scale correspondence with main branch feature fusion are used to model the dynamic correlation between sensors at different time scales. Simultaneously, a wavelet scattering enhancement branch is set in parallel outside the main branch to supplement the multi-scale time-frequency texture features in the original window signals. The main branch graph features and scattering enhancement features are then fused through a dual-pathway system to ultimately identify the normal pipeline state and different leakage types. This improves the temporal consistency and continuous window representation capability of multi-source monitoring signals, thereby enhancing the reliability of pipeline leak detection.

[0089] Specifically, compared to directly stitching together multi-source signals such as pressure, vibration, and acoustics for modeling, this approach is easily affected by differences in sampling frequency, signal length, and time misalignment, making it difficult for different sensor responses to correspond on the same time scale. This application first performs time synchronization, resampling, and length alignment on the multi-source monitoring signals, and then constructs short-time sliding window samples based on this. Furthermore, this application constructs a synchronization stream from the current continuous windows and a time-delay stream from adjacent windows with time offset relationships, so that the model input simultaneously includes the current synchronization response and the time delay changes of adjacent windows. Therefore, compared to methods that only use single-window features or directly stitch together the original signals, this application can more fully express the dynamic changes of pipeline leakage signals within continuous short-time windows.

[0090] Furthermore, this application enhances the ability to match and model the dynamic characteristics of leakage with the correlation between sensors at different time scales. Pressure fluctuations, pipe wall vibrations, and acoustic disturbances caused by pipeline leakage may manifest as instantaneous disturbances, continuous fluctuations, or overall changes within multiple window ranges. Using only a single-scale feature or a single graph structure is insufficient to simultaneously adapt to leakage response characteristics at different time scales. This application sets up multiple feature extraction structures with different time receptive fields in the main branch to extract sensor dynamic features at short, medium, and long scales respectively; and constructs corresponding adaptive graphs for features at different scales, enabling dynamic features at each scale to propagate within a graph structure that matches their time receptive field. Therefore, this application can learn the correlation between pressure, vibration, and acoustic sensors at different time scales, avoiding the mixing of different scale correlation patterns by a single graph structure.

[0091] This application improves the backbone network's ability to comprehensively utilize temporal information and multi-scale graph features. Specifically, graph propagation results at different scales reflect the sensor correlation patterns of the leakage signal under different temporal receptive fields. Using only features at a single scale can easily lead to the loss of information on short-term disturbances, continuous fluctuations, or overall changes; conversely, performing graph propagation without preserving the original temporal input may weaken the fundamental information in the current response and time delay changes. This application fuses the scale-corresponding graph propagation results at short, medium, and long scales with backbone features and introduces residual branches to retain the original temporal response information in the synchronous time-delay dual-stream input. In this way, the graph features output by the backbone branches simultaneously include the current response, time delay changes, multi-time-scale dynamic features, and multi-source sensor correlation information, thereby forming a more complete leakage state characterization and improving the ability to distinguish between similar leakage types.

[0092] This application enhances the ability to utilize multi-scale time-frequency texture information of the original signal. Specifically, statistical time-frequency features mainly reflect the amplitude, frequency, and energy distribution of the signal, but their representation of local fluctuations, non-stationary textures, and fine-grained time-frequency structures remains limited. While existing wavelet scattering methods can extract multi-scale time-frequency features, they are typically used alone as single-modal features and are not fully integrated with the sensor association modeling process. This application sets up a wavelet scattering enhancement branch outside the main branch, directly acting on the original window signal to extract stable multi-scale time-frequency texture features, and then fuses these features with the multi-scale map features output from the main branch. Therefore, this application can simultaneously utilize the sensor association modeling capability of the main network and the original signal texture representation capability of the wavelet scattering branch, compensating for the insufficient representation of single-path features.

[0093] This application improves the stability and engineering applicability of pipeline leak identification under complex operating conditions. Specifically, under complex conditions such as flow rate changes, pressure fluctuations, or transient operation, pipeline leak signals are easily affected by background disturbances, and identification using single sensor features, single time-scale features, or simple post-fusion methods is prone to instability. This application, through synchronous time-delay dual-flow sample organization, adaptive graph construction for different time scales, graph propagation and backbone feature fusion based on scale correspondence, and wavelet scattering enhancement branching, forms a joint representation from continuous window dynamic changes, multi-time-scale sensor correlations, to the time-frequency texture of the original signal. Therefore, this application can more fully utilize the effective information from multi-source monitoring data such as pressure, vibration, and acoustics, improving the accuracy, stability, and adaptability to complex operating conditions in identifying pipelines in normal conditions and different leak types.

[0094] Furthermore, without altering the core technical processes and the invention's objective, some local modules in this application can be implemented using equivalent techniques. For example, the wavelet scattering enhancement branch is primarily used to extract multi-scale time-frequency texture features from the original window signal. To achieve the same texture enhancement objective, it can be replaced by multi-scale time-frequency analysis methods such as continuous wavelet transform, discrete wavelet transform, short-time Fourier transform, or variational mode decomposition. The multi-temporal receptive field dynamic feature extraction module is primarily used to obtain sensor dynamic features across different time spans. To achieve the same dynamic modeling objective, it can be replaced by structures such as one-dimensional convolution, temporal convolution networks, gated convolution, or dilated convolution at different scales.

[0095] The adaptive graph construction module for different time scales is mainly used to obtain sensor association structures at short, medium, and long scales, respectively. When modeling associations at the same scale, adjacency relationships at the corresponding scale can be generated using learnable node embeddings, attention weights, similarity metrics, or dynamic graph learning. The graph propagation module based on scale correspondences is mainly used to propagate short-scale dynamic features in the short-scale adaptive graph, medium-scale dynamic features in the medium-scale adaptive graph, and long-scale dynamic features in the long-scale adaptive graph. When achieving information exchange between nodes, graph learning methods such as graph convolutional networks, graph attention networks, message passing networks, or dynamic graph neural networks can be used instead.

[0096] The fusion of backbone features and wavelet scattering enhancement features can be achieved through feature splicing, weighted fusion, gated fusion, residual fusion, or attention fusion.

[0097] Corresponding to the above method embodiments, this application also provides a pipeline leakage identification device based on a time-adaptive graph. The pipeline leakage identification device based on a time-adaptive graph described below can be referred to in correspondence with the pipeline leakage identification method based on a time-adaptive graph described above.

[0098] See Figure 9 As shown, the device includes the following modules: The data acquisition module 101 is used to acquire multi-source time-series monitoring samples corresponding to the pipeline operation process, and to construct a continuous short-time window sequence based on the multi-source time-series monitoring samples; The dual-stream construction module 102 is used to take the response features of each sensor in the current continuous window as the synchronous stream, take the response features of adjacent windows that have a time offset relationship with the synchronous stream as the time delay stream, and jointly represent the synchronous stream and the time delay stream to obtain the synchronous and time delay dual-stream input. The dynamic feature extraction module 103 is used to extract multi-temporal receptive field features from the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales. The adaptive graph construction module 104 is used to construct adaptive graph structures for different time scales based on dynamic features at different time scales. The graph feature fusion module 105 is used to input the dynamic features at each time scale into the adaptive graph structure of the corresponding time scale for node information propagation, obtain graph features corresponding to different time scales, and fuse multiple graph features to obtain multi-scale fused graph features. The identification module 106 is used to identify pipeline leaks based on multi-scale fusion map features and obtain pipeline leak identification results.

[0099] Using the apparatus provided in this application, multi-source time-series monitoring samples during pipeline operation are acquired, and a continuous short-time window sequence is constructed based on the multi-source time-series monitoring samples. The response features corresponding to each sensor within the current continuous window are taken as a synchronous flow, and the response features of adjacent windows with a time offset relationship with the synchronous flow are taken as a time-delay flow. The synchronous flow and the time-delay flow are jointly represented to obtain a synchronous-time-delay dual-flow input. Multi-time receptive field feature extraction is performed on the synchronous-time-delay dual-flow input to obtain dynamic features corresponding to multiple time scales. Adaptive graph structures corresponding to different time scales are constructed for the dynamic features under different time scales. The dynamic features under each time scale are input into the adaptive graph structure of the corresponding time scale for node information propagation to obtain graph features corresponding to different time scales. The multiple graph features are fused to obtain multi-scale fused graph features, and pipeline leakage is identified based on the multi-scale fused graph features to obtain pipeline leakage identification results.

[0100] Because the response process of a leakage event exhibits significant temporal evolution characteristics, pressure changes, vibration propagation, and acoustic responses at different stages may display dynamic patterns at different time scales. A single time scale feature is insufficient to fully describe the complex changes during the leakage process. To adapt to the dynamic changes in sensor correlations at different time stages and response scales during a pipeline leakage event, this application constructs a continuous short-time window sequence based on multi-source time-series monitoring samples when building the input. Then, the response features corresponding to each sensor within the current continuous window are used as a synchronization stream, and the response features of adjacent windows with time offset relationships with the synchronization stream are used as a time-delay stream. The synchronization stream and time-delay stream are jointly represented to obtain a synchronous-time-delay dual-stream input. In other words, the synchronous-time-delay dual-stream input carries not only the response features corresponding to each sensor within the current continuous window but also the response features of adjacent windows with time offset relationships with the synchronization stream. That is, the synchronous-time-delay dual-stream input of this application allows the model input to simultaneously include the synchronization state information at the current moment and the evolutionary trend information between short-time windows, thereby overcoming the limitation of traditional single-window static features in insufficiently representing the dynamic process of leakage and providing richer temporal priors for subsequent multi-scale dynamic modeling.

[0101] Furthermore, considering that a single-scale graph structure can typically only describe node association information at a fixed scale, it is prone to the overlapping of leakage response features at different time scales, reducing the recognition accuracy in complex leakage scenarios. To address this, this application performs parallel feature extraction across multiple temporal receptive fields on the synchronous delay dual-stream input, obtaining dynamic features corresponding to different time scales. An adaptive graph structure is then independently constructed for each time scale, enabling sensor associations at different scales to learn independently within their respective matched graph topologies. Compared to using a single fixed graph or a shared adaptive graph, this approach avoids the overlapping of association patterns at different time scales, allowing features at each scale to be represented within their most suitable association space, thereby improving the accuracy of characterizing multi-scale dynamic changes in leakage signals.

[0102] Then, the dynamic features at each scale are input into the corresponding adaptive graph structure for node information propagation, enabling the directed transmission of sensor correlation information at each scale within the corresponding graph structure. Multi-scale fused graph features are obtained by fusing graph features from multiple scales. This fused feature integrates multi-level information such as instantaneous disturbances, continuous fluctuations, and overall trends, significantly improving the model's ability to distinguish complex operating conditions and weak leakage signals, ultimately enhancing the reliability of pipeline leakage status identification.

[0103] In one specific embodiment of this application, the data acquisition module is specifically used to divide the multi-source time-series monitoring samples according to a preset window length and a sliding step size to obtain multiple time-series windows; Multiple time-series windows are combined in chronological order to obtain a continuous short-time window sequence that includes information about the current window and adjacent windows.

[0104] In one specific embodiment of this application, the dynamic feature extraction module is specifically used to perform channel mapping processing on the synchronous delay dual-stream input to obtain a fused feature representation that integrates the current window response information and the adjacent window change information; Parallel feature extraction is performed on the fused feature representation under multiple different temporal receptive fields to obtain short-term dynamic features, medium-term dynamic features, and long-term dynamic features.

[0105] In one specific embodiment of this application, the adaptive graph construction module is specifically used to use at least one of a pressure sensor, a vibration sensor, and an acoustic sensor as graph nodes. For dynamic features at different time scales, corresponding learnable node representations are set, and scale-related mapping is used to perform feature transformation on the learnable node representations to obtain node relationship information at the corresponding time scale. Based on node relationship information, the relationships between different graph nodes are interacted to determine the degree of association between different graph nodes, and an adaptive adjacency matrix corresponding to different time scales is generated based on the degree of association.

[0106] In one specific embodiment of this application, the adaptive graph construction module is specifically used to generate a node association matrix based on the degree of directional association between any two graph nodes; The node association matrix is ​​nonnegated and then sorted according to the degree of association between nodes; By retaining a preset number of highly associated nodes for each graph node and setting the association relationships between other nodes to zero, a sparse adaptive adjacency matrix is ​​obtained.

[0107] In one specific embodiment of this application, the graph feature fusion module is specifically used to input short-timescale dynamic features, medium-timescale dynamic features and long-timescale dynamic features into their corresponding adaptive graph structures for graph information propagation, thereby obtaining graph propagation features at the corresponding time scales. The graph propagation features corresponding to different time scales are fused, and residual features generated based on synchronous delay dual-stream input are introduced to obtain multi-scale fused graph features.

[0108] In one specific embodiment of this application, the identification module is specifically used to perform wavelet scattering transform on the window signal corresponding to the multi-source time-series monitoring sample to obtain multi-scale time-frequency texture features; Multi-scale time-frequency texture features are fused with multi-scale fusion map features to obtain comprehensive recognition features; The system determines the pipeline leakage status based on comprehensive identification features and outputs the pipeline leakage identification results.

[0109] Corresponding to the above method embodiments, this application also provides an electronic device. The electronic device described below can be referred to in conjunction with the pipeline leakage identification method based on time-adaptive graphs described above.

[0110] See Figure 10 As shown, the electronic device includes: Memory 332 is used to store computer programs; The processor 322 is used to implement the steps of the pipeline leak identification method based on the timing adaptive graph in the above method embodiment when executing a computer program.

[0111] For details, please refer to Figure 11 , Figure 11 This is a schematic diagram of the specific structure of an electronic device provided in this embodiment. The electronic device can vary significantly due to differences in configuration or performance. It may include one or more central processing units (CPUs) (e.g., one or more processors) and a memory 332. The memory 332 stores one or more computer programs 342 or data 344. The memory 332 can be temporary or permanent storage. The program stored in the memory 332 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the data processing device. Furthermore, the processor 322 may be configured to communicate with the memory 332 and execute the series of instruction operations stored in the memory 332 on the electronic device 301.

[0112] Electronic device 301 may also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341.

[0113] The steps in the pipeline leak identification method based on time-adaptive graphs described above can be implemented by the structure of electronic devices.

[0114] Corresponding to the above method embodiments, this application also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the pipeline leakage identification method based on time-adaptive graphs described above.

[0115] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the pipeline leak identification method based on a timing adaptive graph as described in the above method embodiments.

[0116] The readable storage medium can specifically be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0117] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0118] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0119] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0120] Finally, it should be noted that in this document, relationships such as "first" and "second" are used merely 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. Furthermore, the terms "include," "contain," or any other variations are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus 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 apparatus.

[0121] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A pipeline leak identification method based on time-series adaptive graphs, characterized in that, include: Acquire multi-source time-series monitoring samples during pipeline operation, and construct a continuous short-time window sequence based on the multi-source time-series monitoring samples; The response features of each sensor within the current continuous window are taken as the synchronization stream, and the response features of adjacent windows that have a time offset relationship with the synchronization stream are taken as the time delay stream. The synchronization stream and the time delay stream are jointly represented to obtain the synchronous and time delay dual-stream input. Multi-temporal receptive field feature extraction is performed on the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales, and adaptive graph structures corresponding to different time scales are constructed for the dynamic features under different time scales respectively. The dynamic features at each time scale are input into the corresponding adaptive graph structure for node information propagation to obtain graph features at different time scales. Multiple graph features are fused to obtain multi-scale fused graph features, and pipeline leaks are identified based on these multi-scale fused graph features to obtain pipeline leak identification results.

2. The method according to claim 1, characterized in that, Constructing a continuous short-time window sequence based on the multi-source time-series monitoring samples includes: The multi-source time-series monitoring samples are divided according to a preset window length and sliding step size to obtain multiple time-series windows; Multiple time-series windows are combined in chronological order to obtain a continuous short-time window sequence that includes information about the current window and adjacent windows.

3. The method according to claim 1, characterized in that, The multi-temporal receptive field feature extraction of the synchronous delay dual-stream input includes: The synchronous delay dual-stream input is processed by channel mapping to obtain a fused feature representation that integrates the current window response information and the change information of adjacent windows; Parallel feature extraction is performed on the fused feature representation under multiple different time receptive fields to obtain short-timescale dynamic features, medium-timescale dynamic features, and long-timescale dynamic features.

4. The method according to claim 1, characterized in that, The construction of adaptive graph structures for different time scales based on dynamic features includes: Use at least one of pressure sensors, vibration sensors, and acoustic sensors as graph nodes; For dynamic features at different time scales, corresponding learnable node representations are set, and scale-related mapping is used to perform feature transformation on the learnable node representations to obtain node relationship information at the corresponding time scale. Based on the node relationship information, the relationships between different graph nodes are interacted to determine the degree of association between different graph nodes, and an adaptive adjacency matrix corresponding to different time scales is generated based on the degree of association.

5. The method according to claim 4, characterized in that, Based on the aforementioned correlation level, adaptive adjacency matrices are generated for different time scales, including: Generate a node association matrix based on the degree of directional association between any two graph nodes; The node association matrix is ​​nonnegated and then sorted according to the degree of association between nodes; By retaining a preset number of highly associated nodes for each graph node and setting the association relationships between other nodes to zero, a sparse adaptive adjacency matrix is ​​obtained.

6. The method according to claim 1, characterized in that, The step of inputting dynamic features at each time scale into the corresponding adaptive graph structure for node information propagation includes: The short-timescale dynamic features, medium-timescale dynamic features, and long-timescale dynamic features are respectively input into their corresponding adaptive graph structures for graph information propagation to obtain the graph propagation features of the corresponding timescales. The graph propagation features corresponding to different time scales are fused, and residual features generated based on the synchronous delay dual-stream input are introduced to obtain multi-scale fused graph features.

7. The method according to claim 1, characterized in that, Pipeline leak identification based on the multi-scale fusion map features includes: Wavelet scattering transform is performed on the window signals corresponding to the multi-source time-series monitoring samples to obtain multi-scale time-frequency texture features; The multi-scale time-frequency texture features are fused with the multi-scale fusion map features to obtain comprehensive recognition features; The pipeline leakage status is determined based on the comprehensive identification features, and the pipeline leakage identification result is output.

8. A pipeline leak identification device based on a time-series adaptive graph, characterized in that, include: The data acquisition module is used to acquire multi-source time-series monitoring samples corresponding to the pipeline operation process, and to construct a continuous short-time window sequence based on the multi-source time-series monitoring samples; The dual-stream construction module is used to take the response features of each sensor in the current continuous window as a synchronous stream, take the response features of adjacent windows that have a time offset relationship with the synchronous stream as a time delay stream, and jointly represent the synchronous stream and the time delay stream to obtain synchronous and time delay dual-stream input; The dynamic feature extraction module is used to extract multi-temporal receptive field features from the synchronous delay dual-stream input to obtain dynamic features corresponding to multiple time scales. The adaptive graph construction module is used to construct adaptive graph structures for different time scales based on dynamic features. The graph feature fusion module is used to input the dynamic features at each time scale into the corresponding adaptive graph structure for node information propagation, obtain graph features corresponding to different time scales, and fuse multiple graph features to obtain multi-scale fused graph features. The identification module is used to identify pipeline leaks based on the features of the multi-scale fusion map, and obtain pipeline leak identification results.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the pipeline leak identification method based on a timing adaptive graph as described in any one of claims 1 to 7 when executing the computer program.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the steps of the pipeline leak identification method based on a time-adaptive graph as described in any one of claims 1 to 7.

Images