Pipeline risk monitoring method and system based on artificial intelligence

Abstract

The invention relates to the technical field of pipeline risk monitoring, and discloses a pipeline risk monitoring method and system based on artificial intelligence, and the method comprises the steps: collecting multi-source sensing data of a pipeline operation environment, and carrying out the preprocessing of the data, and obtaining a multi-scale time sequence feature set; and constructing a pipe network diagram model, and embedding the multi-scale time sequence feature set into the pipe network diagram model. And learning the pipe network diagram model by using a space-time diagram attention network to obtain a target prediction result. And constructing an expected economic loss function, and determining an optimal risk threshold based on the expected economic loss function. And comparing the optimal risk threshold with a corrosion event probability prediction value to obtain a risk level, and determining a maintenance priority sequence of each risk point according to the risk level, a historical maintenance record and the importance of a pipeline section. And generating a maintenance work order based on the maintenance priority sequence. According to the invention, dynamic and prospective risk prediction is realized, and the economy and operability of pipeline risk monitoring management are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pipeline risk monitoring, in particular to a pipeline risk monitoring method and system based on artificial intelligence. BACKGROUND

[0002] As an important infrastructure for national energy transportation, oil pipelines play a vital role in the energy and chemical industry. With the development of long-distance oil pipeline technology, its deployment in complex environments, such as mountainous areas, deserts, oceans, and areas with high lightning frequency, has become increasingly common. In these environments, the safety risks of pipeline operation increase, especially lightning, electromagnetic interference, and soil corrosion, which can lead to static electricity accumulation, partial discharge, and electrochemical corrosion on the pipeline surface, seriously threatening the long-term stable operation of the pipeline.

[0003] To address the above risks, CN119151305A proposes a method for monitoring the risk of an oil pipeline region based on parameters such as the number of lightning strikes, lightning current amplitude, and lightning current action time in the region where the pipeline is located, to construct a regional risk factor and quantitatively evaluate the lightning risk of the oil pipeline region. This scheme realizes regionalization and quantification based on traditional lightning risk qualitative description. However, the above scheme does not fully introduce multi-dimensional operation data such as electrochemical corrosion, coating state, and cathodic protection potential, making it difficult to fully reflect the comprehensive risk of the pipeline in complex environments. By statistically analyzing lightning parameters to obtain risk levels, this method belongs to post-event or static evaluation, and cannot predict the probability of future corrosion events based on time series data, lacking foresight. In risk determination, uncertainties caused by data collection errors and model calculation biases are not considered, which may lead to misjudgment in boundary conditions. No direct mapping relationship is established between actual maintenance priority ranking and resource scheduling, making it difficult to support resource allocation and work order scheduling for actual pipeline management departments.

[0004] Therefore, it is necessary to design a pipeline risk monitoring method and system based on artificial intelligence to solve the problems existing in the current technology. SUMMARY

[0005] In view of this, the present application proposes a pipeline risk monitoring method and system based on artificial intelligence, aiming to solve the problems of lagging risk identification, lack of combination with probability prediction, and lack of linkage between regional risk assessment results and resource scheduling and other operation and maintenance decisions in the current technology.

[0006] In one aspect, the present application proposes a pipeline risk monitoring method based on artificial intelligence, comprising: Collecting multi-source sensing data of the pipeline operating environment, and preprocessing the multi-source sensing data to obtain a multi-scale time series feature set; The station field node, the test stake node and the pipe segment edge are constructed into a pipe network graph model based on a pipe network topology, and the multi-scale time sequence feature set is embedded into the pipe network graph model; a space-time graph attention network is used to learn the pipe network graph model to obtain a target prediction result, the target prediction result including a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time; Based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, an expected economic loss function is constructed by combining a shutdown loss cost, a false alarm cost and a maintenance cost, and an optimal risk threshold is determined based on the expected economic loss function. The optimal risk threshold is compared with the corrosion event probability prediction value to obtain a risk level, and a maintenance priority order of each risk point is determined according to the risk level, historical maintenance records and the importance of the pipe section. Based on the maintenance priority order, a maintenance work order is generated, and resource scheduling and task allocation are performed according to personnel resources, material resources and time constraints.

[0007] Further, when the multi-source sensing data is preprocessed to obtain a multi-scale time sequence feature set, it includes: The multi-source sensing data is subjected to time synchronization processing, device fingerprint identification, unit conversion and normalization, data quality detection and drift correction and resampling, missing sample interpolation, and finally feature construction to form a multi-scale time sequence feature set; The multi-source sensing data includes electric field sensing data, soil conductivity data, electrochemical corrosion data, cathodic protection potential data, alternating current interference data, coating detection data, fluid temperature, pressure and flow data, and weather grid data; the feature construction includes generating time domain feature items, frequency domain feature items and cross feature items on the first time sequence window, the second time sequence window and the third time sequence window, and adding business feature items and geographic embedding feature items; the feature vectors corresponding to each time sequence window are spliced to form the multi-scale time sequence feature set.

[0008] Further, when the pipe network graph model is constructed and the multi-scale time sequence feature set is embedded into the pipe network graph model, it includes: The station yard, test pile and pipe section are constructed into a directed graph according to the pipeline design drawing and GIS geographic information, the nodes of the directed graph are station yard nodes and test pile nodes, the edges of the directed graph are pipe section edges, and the burial depth of the pipe section edge is recorded when the edge is generated; for each node, the time sequence feature vector collected at the corresponding position of the node is bound, and the time sequence feature vector is a subset in the multi-scale time sequence feature set; for each pipe section edge, the time sequence feature vectors of the two end nodes of the pipe section edge are weighted and averaged according to the edge weight coefficient, and then spliced with the business feature item of the pipe section edge to form an edge feature vector; the edge weight coefficient is calculated according to the weighted combination of the pipe section length, the soil conductivity difference and the pipe diameter; in the constructed pipe network graph model, the node feature matrix is composed of the time sequence feature vectors of all nodes, the edge feature matrix is composed of the weighted feature vectors of all pipe section edges, and the node feature matrix and the edge feature matrix are combined to generate a graph time sequence input sample, so that the multi-scale time sequence feature set is embedded in the pipe network graph model.

[0009] Further, when learning the pipe network graph model by using a spatio-temporal graph attention network to obtain a target prediction result, the method comprises: The graph time sequence input sample is input into a spatio-temporal graph attention network, the spatio-temporal graph attention network comprises a time sequence encoding layer, a spatial attention layer and an output prediction layer; in the time sequence encoding layer, the node feature matrix and the edge feature matrix of each time sequence window are time-dimensionally encoded based on a multi-layer one-dimensional convolution to obtain a time sequence representation; in the spatial attention layer, the edge weight coefficient is taken as an initial adjacency weight, the time sequence representation of each node and adjacent nodes are weighted and aggregated by graph attention to obtain a spatial correlation representation, and the business feature item is added in the aggregation process to correct the weight contribution of the edge feature; in the output prediction layer, the time sequence representation and the spatial correlation representation are fused to generate the target prediction result, the target prediction result comprises a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time, wherein the uncertainty estimation value is obtained by adding a Monte Carlo dropout mechanism in the output prediction layer.

[0010] Further, when constructing an expected economic loss function based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, combining a shutdown loss cost, a false alarm cost and a maintenance cost, the method comprises: Taking the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value in the target prediction result as input parameters, a shutdown loss cost, a false alarm cost and a maintenance cost are determined as weight coefficients; based on the corrosion event probability prediction value and the uncertainty estimation value, a missed alarm risk probability is calculated, and a missed alarm loss term is generated with the shutdown loss cost as a coefficient; based on the corrosion event probability prediction value and the uncertainty estimation value, a false alarm risk probability is calculated, and a false alarm loss term is generated with the false alarm cost as a coefficient; based on the corrosion rate estimation value, a maintenance requirement intensity is determined, and a maintenance loss term is generated with the maintenance cost as a coefficient; the missed alarm loss term, the false alarm loss term and the maintenance loss term are weighted and summed to form the expected economic loss function.

[0011] Further, when determining the optimal risk threshold based on the expected economic loss function, the method comprises: A candidate threshold set is constructed in a preset step length in the interval [0, 1]; in a historical verification period, for each candidate threshold, the corrosion event probability prediction value in the target prediction result is subjected to probability calibration after temperature scaling, and a confidence weighted probability is calculated in combination with the uncertainty estimation value, and is substituted into the expected economic loss function to obtain a corresponding expected economic loss; the expected economic losses of various pipeline sections are weighted and summarized according to the importance of the pipeline sections, and the average expected economic loss is obtained through time sequence cross verification; a threshold that makes the average expected economic loss minimum and satisfies a preset false alarm rate upper limit and a missed alarm rate upper limit is selected from the candidate threshold set as the optimal risk threshold; and the optimal risk threshold is subjected to exponential smoothing in a sliding time window and a hysteresis interval is set to obtain a running threshold for risk level comparison.

[0012] Further, when comparing the optimal risk threshold with the corrosion event probability prediction value to obtain a risk level, the method comprises: When the corrosion event probability prediction value is less than 50% of the optimal risk threshold, it is determined as low risk; when the corrosion event probability prediction value is greater than or equal to 50% of the optimal risk threshold and less than the optimal risk threshold, it is determined as medium risk; when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold and less than the optimal risk threshold plus the uncertainty estimation value, it is determined as high risk; and when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold plus the uncertainty estimation value, it is determined as extremely high risk.

[0013] Further, when determining the maintenance priority order of each risk point according to the risk level, historical maintenance records and the importance of the pipeline section, the method comprises: convert the risk level into a numerical score, wherein low risk corresponds to 1 point, medium risk corresponds to 2 points, high risk corresponds to 3 points, and extremely high risk corresponds to 4 points; statistics are performed on the historical maintenance records, and a historical score is calculated according to the number of maintenance times, the average maintenance interval, and the defect recurrence rate; the importance of the pipeline section is evaluated, and an importance score is calculated according to whether it is a main section, the size of the oil transmission volume, and the sensitivity of the geographical environment; the risk level score, the historical score, and the importance score are weighted and summed based on weighted calculation to obtain a maintenance priority score, and each risk point is sorted in descending order of the maintenance priority score to form the maintenance priority order.

[0014] Further, based on the maintenance priority order, a maintenance work order is generated, and when resource scheduling and task allocation are performed according to personnel resources, material resources, and time constraints, it includes: Maintenance work orders are generated in the order of the maintenance priority order, and each maintenance work order includes the location of the risk point, the corresponding risk level, the type of maintenance task, the estimated maintenance time, and the required material list; after generating the maintenance work order, the number of personnel resources, the inventory status of material resources, and the time constraints are used as scheduling constraints, and the minimum shutdown time and the minimum total maintenance cost are used as optimization objectives to generate a task allocation scheme and determine the personnel allocation, material allocation, and time allocation results corresponding to each maintenance work order.

[0015] Compared with the prior art, the beneficial effects of the present application are that by introducing multi-source sensing and artificial intelligence, the static regional risk assessment mode of the prior art which only relies on lightning statistical parameters is broken through, multi-dimensional time series data such as electric field sensing, soil conductivity, electrochemical corrosion, cathodic protection potential, alternating current interference, coating detection, fluid temperature and pressure flow, and weather grid can be integrated in a unified pipeline network graph model, and related features across spatial nodes and time windows are extracted using a spatiotemporal graph attention network, so as to generate corrosion event probability prediction values, corrosion rate estimation values, and uncertainty estimation values at future time, realize dynamic and forward-looking risk prediction; on this basis, an expected economic loss function is constructed by combining shutdown loss cost, false alarm cost, and maintenance cost, the organic unification of technical indicators and economic indicators is realized through the determination of an optimal risk threshold, and overhauling or hidden danger false reporting caused by pure technical indicators is effectively avoided; through multi-factor fusion sorting of the risk level, historical maintenance records, and pipeline section importance, a reasonable maintenance priority order is formed, and further maintenance work orders are generated and personnel, material, and time resources are scheduled, which breaks through the whole process from risk prediction to actual maintenance decision execution, and improves the economic efficiency and operability of pipeline risk management.

[0016] On the other hand, the present application also provides a pipeline risk monitoring system based on artificial intelligence, which is used to apply the above-mentioned pipeline risk monitoring method based on artificial intelligence, and includes: The collection unit is configured to collect multi-source sensing data of a pipeline operation environment, pre-process the multi-source sensing data to obtain a multi-scale time series feature set; The first processing unit is configured to construct a station yard node, a test pile node and a pipe segment edge into a pipe network graph model based on a pipeline network topology, and embed the multi-scale time series feature set into the pipe network graph model; learn the pipe network graph model by using a spatio-temporal graph attention network to obtain a target prediction result, the target prediction result including a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time; The second processing unit is configured to construct an expected economic loss function based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, in combination with a shutdown loss cost, a false alarm cost and a maintenance cost, and determine an optimal risk threshold based on the expected economic loss function; The judgment unit is configured to compare the optimal risk threshold with the corrosion event probability prediction value to obtain a risk level, and determine a maintenance priority order of each risk point according to the risk level, a historical maintenance record and a pipeline section importance; The scheduling unit is configured to generate a maintenance work order based on the maintenance priority order, and perform resource scheduling and task allocation according to personnel resources, material resources and time constraint conditions.

[0017] It can be understood that the above-mentioned pipeline risk monitoring method and system based on artificial intelligence have the same beneficial effects, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0018] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments and are not meant to limit the present application. Furthermore, the same reference numerals in the attached drawings are intended to refer to the same components throughout. In the drawings: Figure 1 A flowchart of a pipeline risk monitoring method based on artificial intelligence provided for an embodiment of the present application; Figure 2 A functional block diagram of a pipeline risk monitoring system based on artificial intelligence provided for an embodiment of the present application. DETAILED DESCRIPTION

[0019] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly understood, and the scope of the present disclosure can be accurately conveyed to those skilled in the art. It should be noted that the embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict. The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0020] In the conventional existing oil pipeline risk monitoring method, static risk assessment is performed by relying on a single dimension of lightning parameters, resulting in a lack of multi-source environmental data fusion, and a dynamic correlation between corrosion events and complex factors such as electromagnetic interference and soil conductivity cannot be established. The system lacks the ability to extract multi-scale features from time series data, making it difficult to capture the nonlinear law of corrosion rate changes with the environment, and ignoring the uncertainty quantification of the prediction model, resulting in a lack of economic constraints for setting risk thresholds, and failing to support dynamic maintenance decisions.

[0021] For example, in a pipeline system deployed in a lightning-prone area, there is a spatiotemporal asynchrony between soil conductivity data and cathodic protection potential data, and the difference in sampling frequency between electric field sensing data and meteorological grid data results in a failure to align features. The existing method extracts statistical features using a fixed time window, which cannot identify the corrosion propagation path between station nodes and test pile nodes in the pipeline network topology, resulting in a lag in local corrosion event prediction. When generating maintenance work orders, only a single risk level is used for division, without considering the defect recurrence rate in historical maintenance records and the weight of pipeline section oil transportation, resulting in an imbalance in the allocation of maintenance resources for high-priority pipeline sections.

[0022] If the above problems are not solved, static risk assessment will result in an increased corrosion event omission rate and unnecessary shutdown losses due to false triggering of maintenance work orders. Prediction results lacking uncertainty quantification can produce false positives under boundary conditions, exacerbating the conflict between maintenance costs and safety risks. Failure to fuse multi-source data features will mask the relevance of key corrosion triggers, such as ignoring the synergistic effect of alternating current interference data and coating defects, which can cause sudden pipe wall perforation accidents. The lack of maintenance priority ranking mechanism will result in delays in maintaining high-risk pipeline sections, increasing the probability of pipeline leaks and environmental remediation costs.

[0023] To this end, the present application proposes a pipeline risk monitoring method based on artificial intelligence, comprising: Figure 1 S100: Collecting multi-source sensing data of the pipeline operating environment, and preprocessing the multi-source sensing data to obtain a multi-scale time series feature set. S100: Collecting multi-source sensing data of the pipeline operating environment, and preprocessing the multi-source sensing data to obtain a multi-scale time series feature set.

[0024] S200: Based on the pipeline network topology, the station field node, the test pile node and the pipe segment edge are constructed into a pipe network graph model, and a multi-scale time sequence feature set is embedded into the pipe network graph model. A spatiotemporal graph attention network is used to learn the pipe network graph model to obtain a target prediction result, which includes a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time.

[0025] S300: Based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, a stop loss cost, a false alarm cost and a maintenance cost are combined to construct an expected economic loss function, and an optimal risk threshold is determined based on the expected economic loss function.

[0026] S400: The optimal risk threshold is compared with the corrosion event probability prediction value to obtain a risk level, and a maintenance priority order of each risk point is determined according to the risk level, historical maintenance records and the importance of the pipeline section.

[0027] S500: Based on the maintenance priority order, a maintenance work order is generated, and resource scheduling and task allocation are performed according to personnel resources, material resources and time constraints.

[0028] Specifically, the multi-source sensing data refers to the electric field, soil conductivity, electrochemical corrosion, cathodic protection potential, alternating current interference, coating detection, fluid temperature and pressure flow, and meteorological data collected from the pipeline operating environment. Specifically, real-time synchronous collection can be achieved by using a distributed sensor network and an Internet of Things terminal to comprehensively cover the physical, chemical, and environmental factors affecting pipeline corrosion. The preprocessing refers to time synchronization, device identification, unit conversion, normalization, quality detection, drift correction, resampling, and missing data interpolation of the original sensing data. Specifically, sliding window filtering, Kalman filtering, or interpolation algorithms can be used to eliminate data noise and abnormalities and ensure the reliability of feature extraction. The pipe network graph model refers to a directed graph structure constructed based on the pipeline network topology, with stations and test piles as nodes and pipe sections as edges. Specifically, a graph database or graph computing framework can be used to dynamically represent the spatial correlation and physical connection relationship of the pipeline system. The spatio-temporal graph attention network refers to a deep learning model combining time dimension convolution and graph attention mechanism. Specifically, one-dimensional convolution layers can be used to extract time sequence features, and the spatial correlation of nodes and edges can be aggregated through a multi-head attention mechanism to capture the spatio-temporal evolution law of pipeline corrosion risk. The expected economic loss function refers to a mathematical optimization objective that integrates corrosion probability, corrosion rate, uncertainty, and shutdown cost, false alarm cost, and maintenance cost. Specifically, a weighted loss function can be constructed using Bayesian decision theory to balance the economy and safety of risk warning. The risk level refers to the classification of low, medium, high, and extremely high risks based on the comparison of corrosion probability prediction value and optimal threshold. Specifically, hysteresis interval and exponential smoothing strategy can be used to avoid frequent fluctuations in the threshold value to improve the stability of risk determination. The repair priority order refers to the ranking result generated by integrating risk level, historical repair records, and pipe section importance. Specifically, the analytic hierarchy process or entropy weight method can be used to calculate the weighted score to guide the efficient allocation of limited repair resources.

[0029] The present application realizes dynamic prediction of pipeline corrosion risk through multi-source sensing data fusion and spatio-temporal graph attention network, optimizes risk determination threshold based on economic cost, and generates repair priority based on multi-dimensional evaluation, forming a closed-loop management from data collection, risk prediction to repair decision.

[0030] The working process and principle of the present application are to collect multi-source sensing data of the pipeline operating environment, including electric field sensing data, soil conductivity data, electrochemical corrosion data, cathodic protection potential data, etc. The data is preprocessed, including time synchronization, unit conversion, data quality detection, etc., to finally form a multi-scale time series feature set.

[0031] Then, a pipe network graph model is constructed based on the pipe network topology. The station nodes and test pile nodes are taken as the nodes of the graph, and the pipe sections are taken as the edges of the graph. Then, the multi-scale time sequence feature set is embedded into the graph model, the node is bound to the time sequence feature vector of the corresponding position, and the feature of the edge is obtained by weighted average of the features of the two end nodes.

[0032] The pipe network graph model after embedding features is learned by using a space-time graph attention network. The network includes a time sequence encoding layer, a spatial attention layer and an output prediction layer. The time sequence encoding layer encodes the time sequence features of the nodes and edges, the spatial attention layer captures the spatial correlation between nodes, and the output prediction layer fuses the time sequence and spatial information to generate a prediction result.

[0033] The prediction result includes the corrosion event probability, corrosion rate and uncertainty estimation at the future time. Based on these prediction results, an expected economic loss function is constructed by combining the shutdown loss, false positives and maintenance costs. The optimal risk threshold is determined by minimizing the function.

[0034] The optimal risk threshold is compared with the corrosion event probability prediction value to obtain the risk level. The maintenance priority order of each risk point is determined by combining the historical maintenance records and the importance of the pipe section. Finally, a maintenance work order is generated based on the order, and scheduling and task allocation are performed according to the resource constraints.

[0035] Through multi-source data fusion and graph neural network modeling, dynamic prediction and quantitative evaluation of pipeline corrosion risk are realized, providing data support for maintenance decision-making.

[0036] As a preferred embodiment, the scheme of the present application is implemented as follows: Multi-source sensor data of the pipeline operating environment is collected. It includes electric field intensity data collected by electric field sensors, soil conductivity data collected by soil resistivity measuring instruments, corrosion current density data collected by electrochemical corrosion monitoring systems, pipe-soil potential data collected by cathodic protection systems, alternating current interference voltage data collected by alternating current interference monitoring devices, coating condition data collected by intelligent balls, fluid temperature, pressure and flow data collected by SCADA systems, and meteorological grid data collected by meteorological stations, etc.

[0037] The collected multi-source sensor data is preprocessed. First, time synchronization processing is performed to unify different sources of data to the same time reference. Then, device fingerprint identification is performed to ensure that the data source is traceable. Next, unit conversion and normalization are performed to unify the data scale. Then, data quality detection is performed to perform drift correction and resampling. Missing samples are handled by interpolation. Finally, features are constructed to generate time domain features, frequency domain features, cross features on different time sequence windows, and business features and geographic embedding features are added to form a multi-scale time sequence feature set.

[0038] A pipe network graph model is constructed based on the pipe network topology. The station yard and test pile are taken as nodes of the graph, and the pipe segment is taken as the edge of the graph. The buried depth information of the pipe segment is recorded when the edge is generated. For each node, the time series feature vector collected at this location is bound. For each pipe segment edge, the feature vectors of the two end nodes are weighted and averaged according to the edge weight coefficient, and then spliced with the business characteristics of the pipe segment to form an edge feature vector. The edge weight coefficient is calculated according to the pipe segment length, soil conductivity difference and pipe diameter.

[0039] A space-time graph attention network is used to learn the pipe network graph model. The network includes a time series encoding layer, a spatial attention layer and an output prediction layer. The time series encoding layer uses multi-layer one-dimensional convolution to encode the time series features of the nodes and edges. The spatial attention layer takes the edge weight coefficient as the initial adjacency weight, and uses the graph attention mechanism to weight and aggregate the node features to obtain the spatial correlation representation. The output prediction layer fuses the time series representation and the spatial correlation representation to generate the prediction result.

[0040] An expected economic loss function is constructed based on the prediction result. The corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value are taken as input parameters, and the shutdown loss cost, the false alarm cost and the maintenance cost are taken as weight coefficients. The missed alarm risk probability and the false alarm risk probability are calculated to generate the missed alarm loss term and the false alarm loss term. The maintenance requirement intensity is determined according to the corrosion rate estimation value to generate the maintenance loss term. The loss terms are weighted and summed to form the expected economic loss function.

[0041] An optimal risk threshold is determined by minimizing the expected economic loss function. A candidate threshold set is constructed in the [0, 1] interval, and the expected economic loss is calculated for each candidate threshold. The threshold that minimizes the average expected economic loss and meets the preset false alarm rate and missed alarm rate upper limit is selected as the optimal risk threshold.

[0042] The optimal risk threshold is compared with the corrosion event probability prediction value to obtain the risk level. The repair priority order is determined according to the risk level, the historical repair record and the pipe segment importance. A repair work order is generated based on the order, including the risk point location, the risk level, the repair task type and other information. Finally, a task allocation scheme is generated according to the personnel, material and time constraints.

[0043] Through the above scheme, the application realizes dynamic prediction and quantitative evaluation of pipeline corrosion risk. By fusing multi-source sensing data, the combined effect of environmental changes on pipeline corrosion is captured, overcoming the problem of missing data dimensions caused by a single parameter. By using graph models and spatiotemporal graph attention networks, the dynamic propagation law of corrosion events is effectively modeled, solving the problem of lag in corrosion propagation path identification in traditional methods. The Monte Carlo dropout mechanism is introduced to quantify the uncertainty of the prediction, improving the reliability of the risk assessment. The expected economic loss function is constructed to determine the optimal risk threshold, balancing maintenance cost and safety risk. The multi-factor weighted maintenance priority ranking mechanism optimizes the allocation of maintenance resources. This scheme provides strong data support for pipeline risk monitoring and maintenance decision-making, helping to improve the safety and economy of pipeline operation.

[0044] In some of the above schemes of the application, after collecting multi-source sensing data of the pipeline operating environment, the multi-source sensing data is directly preprocessed to generate a multi-scale time series feature set. However, different sources of sensing data have problems such as time asynchronization, unit inconsistency, and uneven data quality, which makes it difficult to effectively fuse multi-source data in the subsequent feature construction process and accurately reflect the dynamic changes of the pipeline operating environment.

[0045] The application further proposes that when preprocessing multi-source sensing data to obtain a multi-scale time series feature set, the following steps are included: time synchronization processing of multi-source sensing data, device fingerprint identification, unit conversion and normalization, data quality detection and execution of drift correction and resampling, missing sample interpolation, and finally feature construction to form a multi-scale time series feature set. The multi-source sensing data includes electric field sensing data, soil conductivity data, electrochemical corrosion data, cathodic protection potential data, alternating current interference data, coating detection data, fluid temperature, pressure and flow data, and weather grid data. Feature construction includes generating time domain feature items, frequency domain feature items, and cross feature items on the first time series window, the second time series window, and the third time series window, and adding business feature items and geographical embedding feature items. The feature vectors corresponding to each time series window are spliced to form a multi-scale time series feature set. The time domain feature items include extreme value, mean value, variance, skewness, kurtosis, moving average, moving variance, rate of change, and first / second derivative. The frequency domain feature items include power spectral density, dominant frequency, bandwidth, spectral entropy, and wavelet-based multi-scale energy ratio. The cross feature items include cross-correlation coefficient, phase difference estimation, cross-channel difference, and cross-channel ratio. The business feature items include cathodic protection deviation, alternating current interference threshold overrun rate, lightning strike frequency histogram quantization, coating defect count, pressure pulsation intensity, and flow fluctuation intensity. The geographical embedding feature items include soil partition number, buried depth, slope, latitude and longitude grid code, and time period / season code.

[0046] The time synchronization processing aligns the sampling frequency differences of different devices by unifying the timestamps of multi-source sensing data. The device fingerprint identification assigns a unique identifier to each sensor to distinguish the data sources. The unit conversion and normalization converts the sensing data of different dimensions into a unified standard range. The data quality detection corrects data deviation by threshold judgment and outlier rejection. The drift correction and resampling eliminate sensor drift errors and unify the sampling interval using sliding window mean filtering. The missing data interpolation fills in the missing samples using adjacent period data or historical data distribution. In the feature construction process, the first time window is a minute-level window, the second time window is an hour-level window, and the third time window is a day-level window, respectively extracting time domain, frequency domain and cross features of different time granularities. The business feature item quantifies key business indicators based on pipeline operation rules, and the geographic embedding feature item converts spatial attributes into discrete feature vectors through geographic coding.

[0047] Specifically, the multi-source sensing data undergoes time synchronization processing to eliminate time misalignment, device fingerprint identification to ensure traceability of data sources, unit conversion and normalization to eliminate dimensional differences, data quality detection and drift correction to improve data consistency, resampling to unify data frequency, and missing data interpolation to ensure data integrity. Based on different time windows, time domain features are extracted to capture dynamic trends, frequency domain features are extracted to reflect periodic fluctuations, and cross features are extracted to reveal the correlation between multiple sensors. The business feature item incorporates domain knowledge into feature representation, and the geographic embedding feature item encodes spatial and temporal attributes. By concatenating feature vectors of different windows, a multi-scale time series feature set is formed, which retains local details and reflects long-term trends, while integrating business and geographic information to provide high-discrimination input features for subsequent models. For example, the moving variance of cathodic protection deviation is calculated in the minute-level window, combined with the threshold overrun rate of AC interference in the hour-level window, and the coating defect count in the day-level window, to comprehensively evaluate short-term fluctuations and long-term cumulative effects. The soil partition number and burial depth encoding in the geographic embedding feature further associate environmental factors with corrosion risk, improving prediction accuracy.

[0048] As a preferred embodiment, the scheme of the application is implemented as follows: The multi-source sensing data is preprocessed to obtain a multi-scale time series feature set. First, the collected multi-source sensing data is subjected to time synchronization processing to ensure consistency of the timestamps of different sensor data. Then, a unique device fingerprint identification is assigned to each sensing device for subsequent data tracing. Next, data of different units are subjected to unit conversion and normalization processing to make the data comparable.

[0049] Further, the data quality is detected to identify outliers and noise. The detected abnormal data is subjected to drift correction, and resampling is performed to ensure the continuity and consistency of the data. For missing samples, an interpolation algorithm is used for completion.

[0050] Finally, the features are constructed. Time-domain features, frequency-domain features, and cross-domain features are generated on the first time window (e.g., 5 minutes), the second time window (e.g., 1 hour), and the third time window (e.g., 24 hours), respectively. The time-domain features include statistical quantities such as extreme values, mean values, and variances. The frequency-domain features include power spectral density and dominant frequency. The cross-domain features include cross-correlation coefficients and phase differences. In addition, business features such as cathodic protection deviation and AC interference threshold overrun rate are added, as well as geographic embedding features such as soil partition number and burial depth. The feature vectors of each time window are spliced to form the final multi-scale time series feature set.

[0051] Through the above technical solutions, the present application realizes effective preprocessing and feature extraction of multi-source heterogeneous sensor data. Thus, the input data quality of the subsequent risk monitoring model is improved, and the model's perception ability of pipeline state changes at different time scales is enhanced. At the same time, by introducing business features and geographic embedding features, the model can better capture the complexity of the pipeline operating environment, thereby improving the accuracy and interpretability of risk prediction.

[0052] In some of the above schemes of the present application, although the feature extraction of pipeline operation data is realized through the multi-scale time series feature set, the influence of geographic topological structure on feature embedding is not fully considered when constructing the pipe network graph model, resulting in that the association between nodes and pipe segments cannot effectively combine spatial position information, and the learning efficiency of the subsequent spatio-temporal graph attention network is reduced.

[0053] The present application further proposes to construct a directed graph according to the pipeline design drawing and GIS geographic information, with station sites and test piles as nodes and pipe segments as edges and recording the burial depth. For each node, the corresponding time series feature vector is bound, and for the pipe segment edge, the time series feature vectors of the two end nodes are weighted and averaged according to the edge weight coefficient, and then spliced with the business feature items to form the edge feature vector. The edge weight coefficient is calculated by the weighted combination of pipe segment length, soil conductivity difference, and pipe diameter. The node feature matrix and the edge feature matrix are merged to generate the graph time series input sample.

[0054] The calculation of the edge weight coefficient adopts the linear weighted combination of the pipe segment length, the soil conductivity difference, and the pipe diameter, with the weight proportion set to 0.5:0.3:0.2. This proportion can be dynamically adjusted according to different geological environments and operating experience to ensure that the graph structure has a higher response to key influencing factors in spatial modeling. The node feature binding adopts a spatial matching algorithm of GIS geographic coordinates and sensor physical positions, with the positioning error controlled within ±1 meter. The edge feature splicing adopts a dimension alignment mechanism, and when the weighted average feature dimension and the business feature dimension are inconsistent, feature mapping is performed through a fully connected layer.

[0055] Specifically, in the process of constructing the pipe network graph model, first, according to the topological connection relationship of the pipeline design drawing and the GIS geographic coordinate data, the physical entity is converted into a directed graph structure containing coordinate attributes. Each test pile node is bound to the time series feature vector collected by the corresponding sensor, and the spatial matching algorithm is used to ensure that the feature vector strictly corresponds to the physical position. For the feature construction of the pipe segment edge, a dynamic weighted fusion mechanism of the features of the two end nodes is adopted, and the edge weight coefficient is calculated by comprehensively considering the pipe segment length, soil conductivity gradient change and pipe diameter parameters, which effectively reflects the environmental differences of different pipe segments. The merging operation of the node feature matrix and the edge feature matrix adopts the tensor splicing method to form a multi-dimensional graph time series sample containing spatial topological relationship. This feature embedding method enables the subsequent spatiotemporal graph attention network to capture the spatial correlation and temporal dynamics of the pipeline network at the same time, providing accurate input data basis for corrosion prediction.

[0056] As a preferred embodiment, the scheme of the application is implemented as follows: The station yard, test pile and pipe segment are constructed into a directed graph according to the pipeline design drawing and GIS geographic information. The nodes of the directed graph are station yard nodes and test pile nodes, and the edges of the directed graph are pipe segment edges. The burial depth of the pipe segment edge is recorded when the edge is generated. For each node, the time series feature vector collected at the corresponding position of the node is bound. The time series feature vector is a subset of the multi-scale time series feature set. For each pipe segment edge, the time series feature vectors of the two end nodes of the pipe segment edge are weighted and averaged according to the edge weight coefficient, and then spliced with the business feature item of the pipe segment edge to form an edge feature vector. The edge weight coefficient is calculated according to the weighted combination of the pipe segment length, soil conductivity difference and pipe diameter.

[0057] Specifically, the NetworkX library in Python programming language can be used to construct the directed graph model. By reading the pipeline design drawing and GIS data, a graph object is created and nodes and edges are added. For each station yard and test pile node, its corresponding geographic coordinates, device ID and other information are taken as node attributes. For each pipe segment edge, its length, burial depth and other information are taken as edge attributes.

[0058] Further, for each node, a feature subset of the node position is extracted from the multi-scale time series feature set, such as including electric field intensity, soil conductivity, cathodic protection potential and other time series data. These feature vectors can be stored using NumPy arrays and bound as node attributes.

[0059] For each pipe segment edge, the weighted average of the feature vectors of the two end nodes is calculated. The edge weight coefficient can be calculated by the following formula: w=αL+βΔC+γ*D wherein w represents an edge weight coefficient, L is a pipe segment length, AC is a difference of soil conductivity at both ends, D is a pipe diameter, and a, b, and g are weight coefficients. These parameters can be adjusted according to actual conditions.

[0060] Thus, the edge feature vector can be expressed as: V_edge = w * (V_node1 + V_node2) / 2 + V_business wherein V_node1 and V_node2 are feature vectors of the two end nodes, and V_business is a business feature item of the pipe segment.

[0061] Finally, the feature vectors of all nodes form a node feature matrix, and the feature vectors of all edges form an edge feature matrix. These two matrices can be stored in SciPy sparse matrix format to improve computational efficiency. The node feature matrix and the edge feature matrix are combined to generate a graph time sequence input sample, completing the embedding of the multi-scale time sequence feature set into the pipe network graph model.

[0062] Through the above technical solutions, the present application realizes the organic combination of complex pipe network topology and multi-source sensing data. By constructing a directed graph model, the spatial relationship of the pipe system is accurately expressed. The feature weighted average is performed using the edge weight coefficient, effectively capturing the mutual influence between the two end nodes of the pipe segment. The business feature item is fused with the time sequence feature, fully reflecting the pipe operation state. This graph-structured data representation method lays a foundation for subsequent graph neural network learning, which is conducive to mining the spatiotemporal correlation in the pipe system and improving the accuracy of risk prediction.

[0063] In some of the above schemes of the present application, the pipe network graph model constructs the topology of the pipe network through node and edge features, but relying only on static feature embedding cannot effectively capture the spatiotemporal dynamic correlation in the pipe operation process, resulting in limited accuracy of corrosion event prediction, inability to quantify the uncertainty of model prediction, and affecting the reliability of subsequent risk decision-making.

[0064] The application further proposes learning a pipe network graph model by using a spatio-temporal graph attention network to obtain target prediction results, including corrosion event probability prediction values, corrosion rate estimation values and uncertainty estimation values at future time. The graph time sequence input sample is input into the spatio-temporal graph attention network, and the spatio-temporal graph attention network includes a time sequence encoding layer, a spatial attention layer and an output prediction layer. In the time sequence encoding layer, the node feature matrix and the edge feature matrix of each time sequence window are encoded in the time dimension based on a multilayer one-dimensional convolution to obtain a time sequence representation. In the spatial attention layer, the edge weight coefficient is taken as the initial adjacency weight, the time sequence representation of each node and the adjacent node are weighted and aggregated by graph attention to obtain a spatial correlation representation, and a business feature item is added in the aggregation process to correct the weight contribution of the edge feature. In the output prediction layer, the time sequence representation and the spatial correlation representation are fused to generate target prediction results, including corrosion event probability prediction values, corrosion rate estimation values and uncertainty estimation values at future time, wherein the uncertainty estimation value is obtained by adding a Monte Carlo dropout mechanism in the output prediction layer.

[0065] In the time sequence encoding layer, the time sequence of nodes and edges is analyzed by a multilayer one-dimensional convolution in a sliding window, the convolution kernel size is set to a combination of 3x1 and 5x1, and local patterns of different time spans are extracted. In the spatial attention layer, when calculating the attention weight between nodes, a business feature item is introduced as a bias item, for example, the cathodic protection deviation is taken as an edge feature correction factor to suppress the interference of abnormal nodes on adjacent regions. In the output prediction layer, a fully connected network is used to map the fused spatio-temporal features to corrosion probability and rate prediction values, and the Monte Carlo dropout mechanism is kept in an activated state in the inference stage, and the prediction variance is generated by multiple forward propagation sampling to serve as an uncertainty estimation value.

[0066] Specifically, the node feature matrix and edge feature matrix in the graph time sequence input sample are input into the time sequence encoding layer, and the trend change features of different time granularities are extracted by sliding a one-dimensional convolution kernel along the time axis. For example, a 3x1 convolution kernel captures short-term fluctuations, and a 5x1 convolution kernel identifies medium-term trends. The outputs of multiple layers of convolution are spliced to form a time sequence representation. The spatial attention layer constructs an initial adjacency matrix based on the edge weight coefficients, and calculates the dynamic correlation weight between nodes through a multi-head attention mechanism. For example, in areas where the soil conductivity difference is large, the attention weight between adjacent nodes is automatically reduced to reduce abnormal propagation. The business feature item is encoded as an attention weight correction item. For example, when the coating defect count exceeds the threshold, the aggregation weight of the corresponding node is suppressed. The output prediction layer inputs the spatio-temporal feature vector into a fully connected network to generate point estimates of corrosion probability and rate, and at the same time, through Monte Carlo dropout, some neurons are randomly shielded, and the distribution variance of the prediction result is obtained by repeated reasoning, which is output as an uncertainty estimate value. This process realizes the hierarchical extraction and dynamic fusion of spatio-temporal features, improves the prediction accuracy, and quantifies the model uncertainty, providing a reliable basis for subsequent risk decision-making.

[0067] As a preferred embodiment, the scheme of the application is implemented as follows: The graph time sequence input sample is input into the spatio-temporal graph attention network, which includes a time sequence encoding layer, a spatial attention layer, and an output prediction layer. In the time sequence encoding layer, the node feature matrix and edge feature matrix of each time sequence window are encoded in the time dimension based on multiple one-dimensional convolutions to obtain a time sequence representation. Specifically, a 3-layer one-dimensional convolution network is used, with convolution kernel sizes of 5, 3, and 1, and channel numbers of 64, 128, and 256, respectively. In the spatial attention layer, the edge weight coefficients are used as the initial adjacency weights, and the time sequence representation of each node and the adjacent nodes are weighted and aggregated through graph attention to obtain a spatial correlation representation. In the aggregation process, a business feature item is added to correct the weight contribution of the edge feature. Further, the graph attention mechanism uses a multi-head attention structure with 8 heads. In the output prediction layer, the time sequence representation and the spatial correlation representation are fused to generate a target prediction result, which includes a corrosion event probability prediction value, a corrosion rate estimate value, and an uncertainty estimate value at a future time. The uncertainty estimate value is realized by adding Monte Carlo dropout (Dropout=0.3) in the fully connected layer. Thus, randomness is introduced in the prediction process, and the variance of the prediction result is obtained through multiple sampling to estimate the uncertainty.

[0068] By the technical solution, the application realizes accurate prediction and uncertainty quantification of pipeline corrosion risk. The spatiotemporal graph attention network can effectively capture the temporal pattern and spatial correlation in the pipeline network, improving the prediction accuracy. By introducing the Monte Carlo dropout mechanism, the uncertainty of the prediction result can be estimated, providing more comprehensive information support for risk assessment. The future corrosion event probability and corrosion rate can be predicted, and the reliability of the prediction result can be evaluated, which helps the pipeline management department to make reliable decisions, thereby improving the safety and economy of pipeline operation.

[0069] In some of the above schemes of the application, when constructing the expected economic loss function, only the corrosion event probability prediction value and the corrosion rate estimation value are relied on, and the influence of the uncertainty of the model prediction result on the false alarm and missed alarm risk is not considered, resulting in deviation in economic loss evaluation, and it is difficult to accurately balance the relationship between shutdown loss, false alarm cost and maintenance cost.

[0070] The application further proposes to construct an expected economic loss function based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, combined with the shutdown loss cost, the false alarm cost and the maintenance cost.

[0071] The false alarm loss term calculates the false alarm risk probability through the corrosion event probability prediction value and the uncertainty estimation value, and generates it with the shutdown loss cost as the coefficient. The false alarm loss term calculates the false alarm risk probability through the corrosion event probability prediction value and the uncertainty estimation value, and generates it with the false alarm cost as the coefficient. The maintenance loss term determines the maintenance demand intensity through the corrosion rate estimation value, and generates it with the maintenance cost as the coefficient. The uncertainty estimation value is introduced in the calculation of the false alarm risk probability and the false alarm risk probability, which is used to correct the confidence of the probability prediction value. The maintenance demand intensity is divided into different levels according to the numerical range of the corrosion rate estimation value, for example, the corrosion rate less than 0.1 mm / year corresponds to low maintenance demand, 0.1 to 0.3 mm / year corresponds to medium maintenance demand, and greater than 0.3 mm / year corresponds to high maintenance demand. When weighted summation, the weight coefficients of the shutdown loss cost, the false alarm cost and the maintenance cost are dynamically adjusted according to the historical economic loss data.

[0072] Specifically, the false negative risk probability is calculated by the product of the complement of the probability prediction value and the uncertainty estimation value, for example, when the probability prediction value is 0.8 and the uncertainty estimation value is 0.1, the false negative risk probability is (1-0.8) * 0.1 = 0.02. The false positive risk probability is calculated by the product of the probability prediction value and the uncertainty estimation value, for example, when the probability prediction value is 0.8 and the uncertainty estimation value is 0.1, the false positive risk probability is 0.8 * 0.1 = 0.08. The maintenance demand intensity is mapped to the maintenance cost weight according to the value of the corrosion rate estimation value, for example, when the corrosion rate is 0.2 mm / year, the maintenance cost weight is 0.5. The shutdown loss cost, false positive cost, and maintenance cost are multiplied by the false negative risk probability, false positive risk probability, and maintenance demand intensity respectively, and then summed to form the expected economic loss function. By dynamically adjusting the weight coefficient, the economic loss sensitivity of different pipeline sections can be optimized, for example, the shutdown loss cost weight of a high-throughput main pipeline is set to 0.7, and the maintenance cost weight of a branch pipeline is set to 0.6.

[0073] As a preferred embodiment, the scheme of the present application is implemented as follows: Based on the corrosion event probability prediction value, corrosion rate estimation value, and uncertainty estimation value, the expected economic loss function is constructed in combination with the shutdown loss cost, false positive cost, and maintenance cost. Specifically, the corrosion event probability prediction value, corrosion rate estimation value, and uncertainty estimation value in the target prediction result are taken as input parameters, and the shutdown loss cost, false positive cost, and maintenance cost are determined as weight coefficients. Further, based on the corrosion event probability prediction value and the uncertainty estimation value, the false negative risk probability is calculated, and the false negative loss term is generated with the shutdown loss cost as the coefficient. Thus, based on the corrosion event probability prediction value and the uncertainty estimation value, the false positive risk probability is calculated, and the false positive loss term is generated with the false positive cost as the coefficient. For example, the maintenance demand intensity is determined based on the corrosion rate estimation value, and the maintenance loss term is generated with the maintenance cost as the coefficient. Finally, the false negative loss term, false positive loss term, and maintenance loss term are weighted and summed to form the expected economic loss function.

[0074] Through the above technical scheme, the present application can comprehensively consider the corrosion event probability, corrosion rate, and prediction uncertainty, and construct a loss function that comprehensively reflects the economic impact of pipeline risk in combination with the shutdown, false positive, and maintenance costs. The loss function can quantitatively evaluate the expected economic loss under different risk thresholds, providing a scientific basis for determining the optimal risk threshold, and helping to strike a balance between reducing risk and controlling cost. By combining multi-dimensional prediction results with actual economic factors, the accuracy and practicality of risk assessment are improved, providing more reliable support for pipeline operation decision-making.

[0075] In some of the above schemes of the present application, the construction of the expected economic loss function relies on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, but the importance difference of different pipeline sections in the historical verification period is not fully considered, which may ignore the priority protection demand of key sections in the threshold selection process. At the same time, the fixed threshold is not adaptive enough in the dynamic environment, which may cause frequent threshold switching due to data fluctuations, affecting the stability of risk judgment.

[0076] The present application further proposes to construct a candidate threshold set in the interval according to the preset step size. In the historical verification period, for each candidate threshold, the corrosion event probability prediction value is scaled by temperature to calibrate the probability, and the confidence weighted probability is calculated combined with the uncertainty estimation value, and then substituted into the expected economic loss function to obtain the corresponding expected economic loss. The expected economic losses of each pipeline section are weighted and summarized according to the importance of the pipeline section, and the average expected economic loss is obtained through time series cross-validation. From the candidate threshold set, the threshold that makes the average expected economic loss minimum and meets the preset false alarm rate upper limit and the false alarm rate upper limit is selected as the optimal risk threshold. And the optimal risk threshold is executed exponential smoothing in the sliding time window and set the hysteresis interval to obtain the running threshold for risk level comparison.

[0077] Wherein, the candidate threshold set is constructed with 0.05 as the preset step size, generating 21 candidate values covering the interval of 0 to 1. Temperature scaling calibrates the original probability prediction value by introducing a learnable parameter to match it with the real event distribution. The confidence weighted probability is calculated by linearly weighting the calibrated probability and the uncertainty estimation value, and the weight coefficient is set to 0.7 according to the historical data verification result. The importance of the pipeline section weighted summary adopts a piecewise linear function, and the weight coefficient of the main section is 2.0 and that of the non-main section is 1.0. Time series cross-validation uses a sliding window mechanism, dividing 5 sub-periods to calculate the verification error respectively. The hysteresis interval is set to 0.05 to avoid frequent threshold switching caused by small probability fluctuations. The exponential smoothing coefficient is dynamically adjusted according to the historical threshold change rate, and the initial value is set to 0.2.

[0078] Specifically, the candidate threshold set covers all possible decision boundaries by discretizing the search space, the temperature scaling calibration eliminates the model prediction bias, and the confidence weighted probability fusion improves the decision reliability by incorporating the prediction uncertainty. The importance of the pipeline section weighted ensures that the economic loss of the key section occupies a higher weight in the optimization process, the time series cross-validation reduces the risk of overfitting, the exponential smoothing suppresses the impact of short-term data fluctuations on the threshold, and the hysteresis interval enhances the stability of the threshold. Through multi-dimensional data fusion and dynamic optimization mechanism, the optimal risk threshold can balance the protection needs of different sections and the overall economic loss, and realize the minimization of economic cost under the constraint conditions of false alarm rate and false alarm rate.

[0079] As a preferred embodiment, the scheme of the present application is implemented as follows: A set of candidate threshold values is constructed with a step size of 0.01 in the interval [0, 1]. In the historical validation period, for each candidate threshold value, the probability prediction value of the corrosion event in the target prediction result is subjected to temperature scaling for probability calibration, and the confidence weighted probability is calculated in combination with the uncertainty estimate value, and then substituted into the expected economic loss function to obtain the corresponding expected economic loss. The expected economic losses of each pipeline section are weighted and summarized according to the importance of the pipeline section, and the average expected economic loss is obtained through time series cross-validation. From the set of candidate threshold values, the threshold value that minimizes the average expected economic loss and meets the preset upper limit of false alarm rate 0.1 and the upper limit of missed alarm rate 0.05 is selected as the optimal risk threshold value. And the optimal risk threshold value is subjected to exponential smoothing in a sliding time window of 30 days and a hysteresis interval of ±5% is set to obtain the running threshold value for risk level comparison.

[0080] Through the above technical scheme, the present application can determine the optimal risk threshold value based on historical data and multiple evaluation indexes, and dynamically adjust it through the sliding window and the hysteresis interval, thereby improving the accuracy and stability of the risk threshold value. This method avoids the misjudgment that may be caused by a single fixed threshold value, and takes into account multiple factors such as economic loss, false alarm rate and missed alarm rate, making the risk assessment more comprehensive and reliable. In addition, the dynamic adjustment mechanism can adapt to changes in the pipeline operating environment, improving the real-time performance and adaptability of risk monitoring.

[0081] In some of the above schemes of the present application, after determining the optimal risk threshold value, if only a single threshold value is used for risk judgment, the risk level may frequently jump due to fluctuations in the corrosion event probability prediction value or dynamic changes in the uncertainty estimate value, especially when the prediction value is close to the threshold value, the misjudgment risk increases, affecting the stability of the subsequent maintenance priority ranking.

[0082] The present application further proposes that when comparing the optimal risk threshold value with the corrosion event probability prediction value to obtain the risk level, it includes: when the corrosion event probability prediction value is less than 50% of the optimal risk threshold value, it is determined as low risk; when the corrosion event probability prediction value is greater than or equal to 50% of the optimal risk threshold value and less than the optimal risk threshold value, it is determined as medium risk; when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold value and less than the optimal risk threshold value plus the uncertainty estimate value, it is determined as high risk; when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold value plus the uncertainty estimate value, it is determined as extremely high risk.

[0083] The risk level division adopts a multi-level interval judgment mechanism, wherein the low-risk interval is set to be less than 50% of the lower limit of the threshold, and the buffer interval is set to reduce the interference of the slight fluctuation of the prediction value on the low-risk judgment. The medium-risk interval covers 50% of the lower limit of the threshold to the complete threshold range, and is used to identify the area with potential risk but not reaching the triggering condition. The high-risk interval dynamically expands the upper limit of the threshold in combination with the uncertainty estimation value, and reflects the confidence range of the model prediction. The extremely high-risk interval is for the extreme case exceeding the upper limit of the confidence. The boundaries of the intervals are dynamically adjusted by the numerical proportion and the uncertainty, for example, 50% is taken as the starting point of the medium risk, so as to avoid directly taking the complete threshold as the boundary to cause the medium and high risk judgment to be too sensitive.

[0084] Specifically, in the risk judgment process, firstly, the optimal risk threshold is compared with the corrosion event probability prediction value. When the prediction value is less than 50% of the threshold, it is judged as low risk. The interval is set to filter the false positives caused by data noise or short-term fluctuations. When the prediction value reaches 50% of the threshold but does not exceed the complete threshold, it is judged as medium risk. The interval is used to mark the area that needs to be continuously monitored but does not need immediate intervention. When the prediction value exceeds the threshold but does not reach the superimposed value of the threshold and the uncertainty estimation value, it is judged as high risk. At this time, the model prediction has a high confidence, and needs to be processed in priority. When the prediction value exceeds the superimposed value, it is judged as extremely high risk, indicating that there is a high probability of a corrosion event and the model prediction uncertainty is low. By introducing the uncertainty estimation value as a threshold expansion factor, the risk level division can dynamically adapt to the confidence level of the model prediction, for example, automatically relaxing the high-risk judgment range when the uncertainty is high, to avoid missing judgment caused by model bias. At the same time, the 50% buffer interval and the hysteresis interval processing can reduce the level jump caused by the fluctuation of the prediction value near the threshold, and improve the stability of the judgment result.

[0085] As a preferred embodiment, the scheme of the application is implemented as follows: The corrosion event probability prediction value is compared with the optimal risk threshold to obtain the risk level. When the corrosion event probability prediction value is less than 50% of the optimal risk threshold, it is judged as low risk. When the corrosion event probability prediction value is greater than or equal to 50% of the optimal risk threshold and less than the optimal risk threshold, it is judged as medium risk. When the corrosion event probability prediction value is greater than or equal to the optimal risk threshold and less than the optimal risk threshold plus the uncertainty estimation value, it is judged as high risk. When the corrosion event probability prediction value is greater than or equal to the optimal risk threshold plus the uncertainty estimation value, it is judged as extremely high risk.

[0086] Specifically, first, the corrosion event probability prediction value and the optimal risk threshold value are obtained. For example, the corrosion event probability prediction value is 0.65, the optimal risk threshold value is 0.7, and the uncertainty estimation value is 0.1. Then, 0.65 is compared with 50% of 0.7 (i.e., 0.35), and 0.65 is greater than 0.35, so it does not belong to low risk. Further, 0.65 is compared with 0.7, and 0.65 is less than 0.7, so it is determined to be medium risk. Thus, the risk level in this case is medium risk.

[0087] As another case, if the corrosion event probability prediction value is 0.75, then 0.75 is greater than the optimal risk threshold value 0.7, but less than 0.7+0.1=0.8, so it is determined to be high risk. If the corrosion event probability prediction value is 0.85, it is greater than 0.8, and it is determined to be extremely high risk.

[0088] Through the above technical solutions, the present application can divide the risk level into four levels of low, medium, high, and extremely high according to the relationship between the corrosion event probability prediction value and the optimal risk threshold value, thereby achieving fine division of the risk level. By introducing the uncertainty estimation value, a buffer interval can be set between high risk and extremely high risk, thereby avoiding misjudgment of the risk level due to prediction errors. This grading method can provide more accurate risk assessment basis for subsequent maintenance priority sequencing and resource scheduling, and helps to improve the accuracy and effectiveness of pipeline risk management.

[0089] In some of the above schemes of the present application, the dynamic correlation among the real-time risk level, the historical maintenance record, and the importance of the pipeline section is not comprehensively considered when determining the maintenance priority order, resulting in that the maintenance order only depends on a single risk indicator, which cannot balance the defect recurrence risk and the criticality of pipeline operation, and causes low resource allocation efficiency.

[0090] The present application further converts the risk level into a numerical score, wherein low risk corresponds to 1 point, medium risk corresponds to 2 points, high risk corresponds to 3 points, and extremely high risk corresponds to 4 points. The historical maintenance record is counted, and the historical score is calculated according to the number of maintenance times, the average maintenance interval, and the defect recurrence rate. The importance of the pipeline section is evaluated, and the importance score is calculated according to whether it is a main section, the oil transportation volume, and the geographical environmental sensitivity. The risk level score, the historical score, and the importance score are weighted and summed based on weighted calculation to obtain the maintenance priority score, and each risk point is sorted in descending order of the maintenance priority score to form the maintenance priority order.

[0091] In the risk level score, the discrete quantification method is adopted to map different risk levels through fixed score intervals. In the historical score calculation, the maintenance frequency is compressed by a logarithmic function to avoid data skewness, the average maintenance interval is reflected by the inverse form to reflect the time urgency, and the defect recurrence rate is linearly weighted. In the importance score, the trunk section is given a reference weight, the oil transmission volume is converted by a segmented function, and the geographical environmental sensitivity is quantified according to the soil corrosion grade and geological disaster probability. In the weighted calculation process, the weight coefficients of each score item are dynamically adjusted according to the pipeline operation strategy, and the dimension difference is eliminated by normalization processing.

[0092] Specifically, the risk level score is converted into quantifiable values by discretization, providing basic data for subsequent weighted calculation. In the process of statistical maintenance record, the maintenance frequency is converted by logarithm to reduce the influence of abnormal values, the average maintenance interval is processed by inverse to strengthen the weight of short-term repeated defects, and the defect recurrence rate directly reflects the regional maintenance reliability. In the pipeline importance assessment, the trunk section is determined according to the network topology connectivity, the oil transmission volume is classified and mapped according to the actual transmission capacity, and the geographical environmental sensitivity integrates the soil corrosion rate and geological disaster history data. In the weighted summation stage, a dynamic weight distribution mechanism is adopted to adjust the score item weight according to the maintenance resource abundance and pipeline operation state in different periods, ensuring that the priority ranking adapts to the actual operation and maintenance demand. Through multi-dimensional score fusion and dynamic weight adjustment, the objective quantitative ranking of maintenance sequence is realized, effectively coordinating the contradiction between immediate risk, historical maintenance quality and pipeline strategic value.

[0093] As a preferred embodiment, the scheme of the application is implemented as follows: when determining the maintenance priority order, first map the risk level to a numerical score of 1 to 4 points, where low risk corresponds to 1 point, medium risk corresponds to 2 points, high risk corresponds to 3 points, and extremely high risk corresponds to 4 points. The historical maintenance record score is calculated by counting the maintenance frequency, average maintenance interval in days, and defect recurrence rate of each risk point in the past three years, where more than 5 times of maintenance frequency deducts 1 point, less than 180 days of average maintenance interval deducts 1 point, and more than 20% of defect recurrence rate deducts 2 points. The pipeline section importance score is quantified based on three dimensions: 3 points are added if it belongs to the trunk section, 2 points are added if the oil transmission volume exceeds 1 million cubic meters / day, and 1 point is added if it is in a geographically sensitive area of earthquake zone or ecological protection zone. The risk level score, historical score and importance score are weighted and summed according to the weight ratio of 5:3:2 to form the maintenance priority score. Finally, the priority ranking list is generated according to the score from high to low, and the risk points with the same score are arranged in geographical coding order.

[0094] By the technical scheme, the application effectively solves the problem of single maintenance priority judgment dimension in the traditional method, and a multi-dimensional decision model is constructed by fusing dynamic risk prediction, historical operation and maintenance data and infrastructure importance evaluation. The scheme realizes accurate allocation of maintenance resources for high-risk pipe sections, avoids maintenance lag caused by misjudgment of local corrosion events, reduces excessive maintenance cost of non-critical sections, and improves the economy and reliability of the pipeline operation and maintenance system.

[0095] In some schemes of the application, the determination of the maintenance priority order only depends on the risk level and the historical maintenance record, and the importance difference of the pipeline section is not fully considered, which may lead to delayed maintenance of critical sections or unreasonable allocation of resources, and cannot effectively balance risk control and operation and maintenance cost.

[0096] The application further provides a method for generating maintenance work orders based on the maintenance priority order, and scheduling resources and allocating tasks, including: generating maintenance work orders in order of maintenance priority, each maintenance work order containing a risk point position, a corresponding risk level, a maintenance task type, an estimated maintenance time and a required material list. After generating the maintenance work orders, the number of personnel resources, the inventory status of material resources and the time constraint conditions are taken as scheduling constraints, and the minimum shutdown time and the minimum total maintenance cost are taken as optimization objectives to generate a task allocation scheme and determine the personnel allocation, material allocation and time allocation results corresponding to each maintenance work order.

[0097] In the maintenance work order generation process, the risk point position is matched with the pipeline topology map through geographic coding, the maintenance task type is mapped to different maintenance intensity levels according to the corrosion rate estimate value and the risk level, the estimated maintenance time is estimated based on the average time consumption of similar tasks in the historical maintenance record, and the required material list is generated according to the task type by calling a preset material configuration template. In the scheduling constraints, the number of personnel resources is divided into multiple work groups according to skill levels, the inventory status of material resources is synchronized to the scheduling system in real time, and the time constraint conditions include maintenance window period limitation and task interdependence. The optimization objectives are realized through a multi-objective planning model, a genetic algorithm is used to solve a Pareto optimal solution set, and the final scheme is selected based on a preset shutdown time weight and a cost weight.

[0098] Specifically, after the generation of the repair work order, the scheduling system first checks whether the material inventory meets the needs of the work order. If the inventory is insufficient, the procurement process is triggered or the priority order is adjusted. When assigning personnel, high-skilled work groups are preferentially assigned to high-risk level work orders, while considering the current location of the personnel and the distance to the task location to shorten the response time. The material allocation adopts a dynamic inventory management mechanism, which updates the inventory status in real time and reserves emergency reserves. During the time allocation process, the maintenance window period is dynamically adjusted according to the pipeline operation plan and weather conditions, and the dependency relationship between tasks is modeled through a directed acyclic graph to avoid conflicts. The optimization model converts the shutdown time into economic losses, which together with the maintenance cost constitutes the objective function, and through iterative solution, an allocation scheme that takes into account efficiency and cost is obtained. Thus, the resource scheduling and task allocation process not only meets the real-time constraints, but also maximizes the operation and maintenance benefits.

[0099] As a preferred embodiment, the scheme of the present application is implemented as follows: the repair work order generation module extracts the risk point list from high to low according to the priority score. Each work order contains the risk point latitude and longitude coordinates, corrosion probability prediction value interval, required replacement anticorrosion layer type, estimated 6-hour operation time, and required epoxy resin coating and cathodic protection material list. The resource scheduling engine calls the personnel skill database to match the maintenance team with pipeline welding qualifications, queries the real-time inventory of coating from the inventory management system, and splits the work order into parallel task chains. The constraint solver is based on genetic algorithm, and the daily 8-hour work upper limit of the team, the 3-day coating replenishment period, and the single shutdown window not exceeding 4 hours are input as hard constraint conditions. Iterative calculation is performed to obtain a scheme with the shortest total shutdown time and lower labor cost than the budget. Finally, the corresponding work order execution sequence and material requisition plan for each team are output.

[0100] Through the above technical scheme, the present application realizes accurate matching of repair tasks and dynamic resource constraints, effectively balancing the contradiction between pipeline operation continuity and maintenance cost. Specifically, a distribution scheme that takes into account time efficiency and economy is automatically generated through a multi-objective optimization algorithm, avoiding the common resource conflict and priority inversion problems in manual scheduling. At the same time, relying on real-time inventory data, the material allocation strategy is dynamically adjusted to reduce maintenance delays caused by insufficient spare parts.

[0101] The above embodiment breaks through the existing static regional risk assessment mode which only relies on lightning statistical parameters by introducing multi-source sensing and artificial intelligence, can integrate multi-dimensional time series data such as electric field sensing, soil conductivity, electrochemical corrosion, cathodic protection potential, alternating current interference, coating detection, fluid temperature and pressure flow, and weather grid in a unified pipe network model, and use a spatiotemporal graph attention network to extract related features across spatial nodes and time windows, thereby generating corrosion event probability prediction values, corrosion rate estimates and uncertainty estimates at future time, realizing dynamic and forward-looking risk prediction. On this basis, an expected economic loss function is constructed by combining the shutdown loss cost, false alarm cost and maintenance cost, and the optimal risk threshold is determined to realize the organic unification of technical indicators and economic indicators, effectively avoiding excessive maintenance or hidden danger caused by pure technical indicators. Through the multi-factor fusion and sorting of risk level, historical maintenance records and pipe section importance, a reasonable maintenance priority order is formed, and further maintenance work orders are generated to schedule personnel, materials and time resources, thereby breaking through the whole process from risk prediction to actual maintenance decision execution, and improving the economy and operability of pipeline risk monitoring and management.

[0102] In another preferred mode based on the above embodiment, referring to Figure 2 The present embodiment provides a pipeline risk monitoring system based on artificial intelligence, which is used to apply the above-mentioned pipeline risk monitoring method based on artificial intelligence, and includes: The acquisition unit is configured to acquire multi-source sensing data of the pipeline operating environment, and to obtain a multi-scale time series feature set by preprocessing the multi-source sensing data.

[0103] The first processing unit is configured to construct a pipe network graph model by constructing station nodes, test pile nodes and pipe segment edges based on the pipe network topology, and to embed the multi-scale time series feature set into the pipe network graph model. The spatiotemporal graph attention network is used to learn the pipe network graph model to obtain a target prediction result, which includes corrosion event probability prediction values, corrosion rate estimates and uncertainty estimates at future time.

[0104] The second processing unit is configured to construct an expected economic loss function based on the corrosion event probability prediction values, corrosion rate estimates and uncertainty estimates, and to combine the shutdown loss cost, false alarm cost and maintenance cost, and to determine the optimal risk threshold based on the expected economic loss function.

[0105] The judgment unit is configured to compare the optimal risk threshold with the corrosion event probability prediction values to obtain a risk level, and to determine the maintenance priority order of each risk point according to the risk level, historical maintenance records and pipe section importance.

[0106] The scheduling unit is configured to generate a maintenance work order based on a maintenance priority order, and to perform resource scheduling and task allocation according to personnel resources, material resources and time constraints.

[0107] Specifically, the acquisition unit performs millisecond timestamp alignment on the multi-source sensing data through the edge computing device, adopts an adaptive filtering algorithm to eliminate signal interference, and generates a standardized data packet for transmission to the cloud. After receiving the data packet, the first processing unit calls a pipe network topology analysis engine to automatically build a weighted directed graph data structure, maps the cathode protection potential data and coating detection data to node attributes, and quantifies the soil conductivity difference and pipe segment length as edge weight coefficients. The spatiotemporal graph attention network inference framework performs three-dimensional convolution operations under GPU acceleration, simultaneously extracts corrosion feature patterns at time scales of 24 hours, 72 hours and 168 hours, and outputs probability prediction values while generating Bayesian uncertainty indicators.

[0108] After receiving the prediction results, the second processing unit calls a preset cost parameter matrix, dynamically calculates the hourly shutdown loss coefficient in combination with the oil transportation volume of the pipe segment, generates 1000 probability distribution samples through Monte Carlo simulation, and calculates the threshold candidate set of the expected economic loss function at a 95% confidence interval. The judgment unit extracts the maintenance work order data of the last three years from the historical maintenance record library, uses a sliding window to statistically analyze the defect recurrence frequency, combines the coordinates of the high-sensitivity areas marked by the GIS geographic information system, generates a three-dimensional scoring vector and performs normalization weighting.

[0109] After receiving the maintenance priority list, the scheduling unit calls the inventory quantities of flange gaskets and anticorrosive coatings in the material inventory database, combines the on-duty state data of the personnel positioning system, and builds an integer programming model with constraints. The multi-objective optimization solver takes minimizing the total downtime as the primary goal, and takes the maintenance cost not exceeding the budget limit as a hard constraint, outputs comprehensive scheduling instructions including personnel dispatch routes, material distribution schemes and time window arrangements, and finally pushes the digital work order to the field operation and maintenance team through a mobile terminal.

[0110] The pipeline risk monitoring system based on artificial intelligence comprises a collection unit, a first processing unit, a second processing unit, a judgment unit and a scheduling unit. The collection unit collects multi-source sensing data through electric field sensors, electrochemical sensors, pressure sensors and meteorological monitoring equipment deployed along the pipeline, and performs time synchronization processing, unit normalization, data quality detection and missing data interpolation, and constructs a multi-scale time series feature set including time domain features, frequency domain features, cross features, business features and geographic embedding features. The first processing unit constructs a pipe network graph model based on the pipeline design map and the geographic information system, takes the station and the test pile as the node, takes the pipe section as the edge, embeds the multi-scale time series features into the feature matrix of the node and the edge, uses the spatiotemporal graph attention network to perform spatiotemporal correlation modeling, and outputs the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value at the future time. The second processing unit constructs an expected economic loss function by combining the shutdown loss cost, the false alarm cost and the maintenance cost, and determines the optimal risk threshold value through historical data cross-validation and dynamic threshold optimization. The judgment unit divides the low, medium, high and extremely high risk levels according to the comparison result of the corrosion event probability prediction value and the optimal risk threshold value, and evaluates the maintenance priority score in combination with the historical maintenance frequency, the defect recurrence rate and the pipeline oil transportation volume. The scheduling unit generates a maintenance work order containing the risk point position, the task type and the material demand according to the priority score, generates a task allocation scheme by using a multi-objective optimization algorithm based on the personnel availability, the material inventory and the time window constraint, and realizes resource scheduling and task execution planning.

[0111] Through the technical solution, the application solves the problems of insufficient multi-dimensional data fusion, lack of prediction foresight and lack of quantitative basis for maintenance scheduling in the prior art, realizes dynamic prediction and uncertainty quantification of pipeline corrosion risk, combines the risk level with the economic loss model, optimizes the maintenance priority ranking, and improves the resource scheduling efficiency. Through the spatiotemporal graph attention network and the multi-scale feature embedding, the spatiotemporal correlation characteristics of the pipeline network are effectively captured, the false alarm and the missed alarm probability are reduced, the prediction uncertainty is quantified through the Monte Carlo dropout mechanism, and the robustness of risk judgment is enhanced. Based on the economic cost function and the dynamic threshold adjustment, the operation and maintenance cost in different risk scenarios is balanced, and the problems of excessive maintenance or insufficient maintenance are avoided. The automatic generation of the maintenance work order and the resource scheduling reduces the manual decision deviation, improves the emergency response speed and the rationality of resource allocation.

[0112] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) having computer usable program code embodied therein.

[0113] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0114] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0115] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0116] Finally, it should be noted that the above-mentioned embodiments are merely intended for describing the technical solutions of the present application, but not for limiting thereof. Although the present application is described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalent replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement should be covered in the protection scope of the claims of the present application.

Claims

1. An artificial intelligence-based pipeline risk monitoring method, characterized by, The method comprises the following steps: Collecting multi-source sensing data of pipeline operation environment, and pre-processing the multi-source sensing data to obtain a multi-scale time series feature set; Based on the pipeline network topology, a station site node, a test pile node and a pipe segment edge are constructed into a pipe network graph model, and the multi-scale time series feature set is embedded into the pipe network graph model; A space-time graph attention network is used to learn the pipe network graph model to obtain a target prediction result, which includes a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time; Based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, a stop loss cost, a false alarm cost and a maintenance cost are combined to construct an expected economic loss function, and an optimal risk threshold is determined based on the expected economic loss function; The optimal risk threshold is compared with the corrosion event probability prediction value to obtain a risk level, and a maintenance priority order of each risk point is determined according to the risk level, historical maintenance records and the importance of the pipeline section; Based on the maintenance priority order, a maintenance work order is generated, and resource scheduling and task allocation are performed according to personnel resources, material resources and time constraints. 2.The AI-based pipeline risk monitoring method of claim 1, wherein, When pre-processing the multi-source sensing data to obtain a multi-scale time series feature set, the method comprises the following steps: Time synchronization processing, device fingerprint identification, unit conversion and normalization, data quality detection and drift correction and resampling, missing sample interpolation, and finally feature construction to form a multi-scale time series feature set are performed on the multi-source sensing data; The multi-source sensing data includes electric field sensing data, soil conductivity data, electrochemical corrosion data, cathodic protection potential data, alternating current interference data, coating detection data, fluid temperature, pressure and flow data, and weather grid data; the feature construction includes generating time domain feature items, frequency domain feature items and cross feature items on the first time series window, the second time series window and the third time series window, and adding business feature items and geographic embedding feature items; the feature vectors corresponding to each time series window are spliced to form the multi-scale time series feature set. 3.The AI-based pipeline risk monitoring method of claim 2, wherein, When constructing a pipe network graph model and embedding the multi-scale time series feature set into the pipe network graph model, the method comprises the following steps: The station yard, test pile and pipe section are constructed into a directed graph according to the pipeline design drawing and GIS geographic information, the nodes of the directed graph are station yard nodes and test pile nodes, the edges of the directed graph are pipe section edges, and the burial depth of the pipe section edge is recorded when the edge is generated; for each node, the time sequence feature vector collected at the corresponding position of the node is bound, and the time sequence feature vector is a subset in the multi-scale time sequence feature set; for each pipe section edge, the time sequence feature vectors of the two end nodes of the pipe section edge are weighted and averaged according to the edge weight coefficient, and then spliced with the business feature item of the pipe section edge to form an edge feature vector; the edge weight coefficient is calculated according to the weighted combination of the pipe section length, the soil conductivity difference and the pipe diameter; in the constructed pipe network graph model, the node feature matrix is composed of the time sequence feature vectors of all nodes, and the edge feature matrix is composed of the weighted feature vectors of all pipe section edges, and the node feature matrix and the edge feature matrix are combined to generate a graph time sequence input sample, so that the multi-scale time sequence feature set is embedded in the pipe network graph model. 4.The AI-based pipeline risk monitoring method of claim 3, wherein, When learning the pipe network graph model by using a spatio-temporal graph attention network to obtain a target prediction result, the following steps are included: The graph time sequence input sample is input into a spatio-temporal graph attention network, the spatio-temporal graph attention network includes a time sequence encoding layer, a spatial attention layer and an output prediction layer; in the time sequence encoding layer, the node feature matrix and the edge feature matrix of each time sequence window are encoded in the time dimension based on a multi-layer one-dimensional convolution to obtain a time sequence representation; in the spatial attention layer, the edge weight coefficient is taken as an initial adjacency weight, the time sequence representation of each node and adjacent nodes are weighted and aggregated by graph attention to obtain a spatial correlation representation, and the business feature item is added in the aggregation process to correct the weight contribution of the edge feature; in the output prediction layer, the time sequence representation and the spatial correlation representation are fused to generate the target prediction result, the target prediction result includes a corrosion event probability prediction value, a corrosion rate estimation value and an uncertainty estimation value at a future time, wherein the uncertainty estimation value is obtained by adding a Monte Carlo dropout mechanism in the output prediction layer. 5.The AI-based pipeline risk monitoring method of claim 1, wherein, When constructing an expected economic loss function based on the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value, combining the shutdown loss cost, the false alarm cost and the maintenance cost, the following steps are included: Taking the corrosion event probability prediction value, the corrosion rate estimation value and the uncertainty estimation value in the target prediction result as input parameters, a shutdown loss cost, a false alarm cost and a maintenance cost are determined as weight coefficients; based on the corrosion event probability prediction value and the uncertainty estimation value, a missed alarm risk probability is calculated, and a missed alarm loss term is generated with the shutdown loss cost as a coefficient; based on the corrosion event probability prediction value and the uncertainty estimation value, a false alarm risk probability is calculated, and a false alarm loss term is generated with the false alarm cost as a coefficient; based on the corrosion rate estimation value, a maintenance requirement intensity is determined, and a maintenance loss term is generated with the maintenance cost as a coefficient; the missed alarm loss term, the false alarm loss term and the maintenance loss term are weighted and summed to form the expected economic loss function. 6.The pipeline risk monitoring method based on artificial intelligence according to claim 5, wherein, When determining the optimal risk threshold based on the expected economic loss function, the method comprises: A candidate threshold set is constructed in the [0, 1] interval according to a preset step size; in a historical verification period, for each candidate threshold, the corrosion event probability prediction value in the target prediction result is subjected to probability calibration after temperature scaling, and a confidence weighted probability is calculated in combination with the uncertainty estimation value, and the expected economic loss function is substituted to obtain a corresponding expected economic loss; the expected economic losses of each pipeline section are weighted and summarized according to the importance of the pipeline section, and the average expected economic loss is obtained through time series cross-validation; a threshold that makes the average expected economic loss minimum and satisfies the preset upper limit of false alarm rate and the upper limit of missed alarm rate is selected from the candidate threshold set as the optimal risk threshold; and the optimal risk threshold is subjected to exponential smoothing in a sliding time window and a hysteresis interval is set to obtain a running threshold for risk level comparison. 7.The pipeline risk monitoring method based on artificial intelligence according to claim 6, wherein, When comparing the optimal risk threshold with the corrosion event probability prediction value to obtain a risk level, the method comprises: When the corrosion event probability prediction value is less than 50% of the optimal risk threshold, it is determined as low risk; when the corrosion event probability prediction value is greater than or equal to 50% of the optimal risk threshold and less than the optimal risk threshold, it is determined as medium risk; when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold and less than the optimal risk threshold plus the uncertainty estimation value, it is determined as high risk; and when the corrosion event probability prediction value is greater than or equal to the optimal risk threshold plus the uncertainty estimation value, it is determined as extremely high risk. 8.The pipeline risk monitoring method based on artificial intelligence according to claim 7, wherein, When determining the maintenance priority order of each risk point according to the risk level, the historical maintenance record and the importance of the pipeline section, the method comprises: transforming the risk level into a numerical score, wherein low risk corresponds to 1 point, medium risk corresponds to 2 points, high risk corresponds to 3 points, and extremely high risk corresponds to 4 points; performing statistics on the historical maintenance records, calculating a historical score according to the number of maintenance times, the average maintenance interval, and the defect recurrence rate; evaluating the importance of the pipeline section, calculating an importance score according to whether it is a main section, the size of the oil transportation volume, and the sensitivity of the geographical environment; performing weighted summation on the risk level score, the historical score, and the importance score based on weighted calculation, obtaining a maintenance priority score, and ranking each risk point in descending order of the maintenance priority score to form the maintenance priority order. 9.The pipeline risk monitoring method based on artificial intelligence according to claim 8, wherein, Based on the maintenance priority order, a maintenance work order is generated, and resource scheduling and task allocation are performed according to personnel resources, material resources, and time constraints, including: Maintenance work orders are generated in the order of the maintenance priority, each containing the location of the risk point, the corresponding risk level, the type of maintenance task, the estimated maintenance time, and the required material list. After generating the maintenance work order, the number of personnel resources, the inventory status of material resources, and the time constraints are used as scheduling constraints, and the minimum shutdown time and the minimum total maintenance cost are used as optimization objectives to generate a task allocation scheme and determine the personnel allocation, material allocation, and time allocation results corresponding to each maintenance work order.

10. An artificial intelligence-based pipeline risk monitoring system for applying the artificial intelligence-based pipeline risk monitoring method of any one of claims 1-9, comprising: a collection unit configured to collect multi-source sensing data of a pipeline operating environment, and to obtain a multi-scale time series feature set by preprocessing the multi-source sensing data; a first processing unit configured to construct a pipeline network graph model by embedding the multi-scale time series feature set into the pipeline network graph model based on the pipeline network topology, wherein the pipeline network graph model is constructed by constructing station nodes, test pile nodes, and pipeline segment edges based on the pipeline network topology; learning the pipeline network graph model using a spatio-temporal graph attention network to obtain a target prediction result, wherein the target prediction result includes a corrosion event probability prediction value, a corrosion rate estimation value, and an uncertainty estimation value at a future time; a second processing unit configured to construct an expected economic loss function based on the corrosion event probability prediction value, the corrosion rate estimation value, and the uncertainty estimation value, in combination with shutdown loss cost, false alarm cost, and maintenance cost, and to determine an optimal risk threshold based on the expected economic loss function; a judgment unit configured to compare the optimal risk threshold with the corrosion event probability prediction value to obtain a risk level, and to determine a maintenance priority order of each risk point based on the risk level, historical maintenance records, and pipeline section importance; a scheduling unit configured to generate a maintenance work order based on the maintenance priority order, and to perform resource scheduling and task allocation according to personnel resources, material resources, and time constraints.

Images