A risk prediction method and skid-mounted system for transmission pipelines based on big data
By obtaining the characterization array and abnormal characterization coefficients of the sensor data set of the external transmission pipeline, identifying and extracting key risk data items, the problems of low risk prediction accuracy and misidentification in the existing technology are solved, and the accuracy of risk prediction in the external transmission pipeline is improved. .
Patent Information
- Application Number
- CN202411297813.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-18
AI Technical Summary
In the risk prediction of external transmission pipelines, it is difficult to accurately identify key risk data items, resulting in misidentification of risk data items, reducing the accuracy and efficiency of risk prediction, and lacking detailed classification of the abnormality of data items, which affects the extraction of key points. Effects of risk data items.
By obtaining the characterization array of the sensor data set of the external transmission pipeline, the abnormal characterization coefficient of the pending risk data items is determined, and then risk data items are detected and estimated, key risk data items are extracted, and the accuracy and accuracy of risk prediction are improved.
The risk prediction neural network has improved the ability to extract risk data items, enhanced the correlation identification of the sensor data set of the external transmission pipeline, and ensured the accuracy and accuracy of risk prediction.
Smart Images

Figure CN118821028B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing, and in particular to a method and a skid-mounted system for predicting risks of an external transmission pipeline based on big data. Background Art
[0002] With the rapid development of the oil and gas energy industry, the construction and operation and maintenance of long-distance transmission pipelines are facing unprecedented challenges. These pipelines often span vast areas and cross complex geological environments. The safety and stability of their operating status are directly related to the reliability of energy supply and the quality of environmental protection. Therefore, how to effectively monitor the operating status of transmission pipelines and timely discover and warn of potential risks has become a key issue that the industry needs to solve. Traditional pipeline monitoring methods mainly rely on manual inspections and regular testing. This method is not only inefficient, but also difficult to achieve real-time monitoring of pipeline status and abnormal warnings. With the rise of big data technology and artificial intelligence, risk prediction methods based on sensor data have gradually become a research hotspot. By deploying a large number of sensors on the transmission pipeline, various parameters of the pipeline (such as pressure, temperature, flow, vibration, etc.) are monitored in real time, and these data are used to construct sensor data sets, providing a rich data source for risk prediction.
[0003] However, sensor data sets usually contain massive, high-dimensional, and complex data, and it is not easy to directly extract valuable risk information from them. In the prior art, although some risk prediction methods based on machine learning or neural networks have been proposed, these methods often focus on statistical analysis and pattern recognition of the entire data set, and lack a detailed assessment of the degree of abnormality of a single data item. Therefore, in practical applications, these methods can produce a large number of misidentified risk data items, reducing the accuracy and efficiency of risk prediction. In addition, there are also deficiencies in the identification of key risk data items in the prior art. Key risk data items refer to those data items with a high degree of abnormality that have an important impact on the operating status of the pipeline. Accurately identifying these key data items not only helps to quickly locate the source of risk, but also provides strong support for subsequent emergency response and maintenance decisions. However, due to the lack of detailed division of the degree of abnormality of data items in the existing methods, it is often difficult to extract key risk data items. Summary of the invention
[0004] The purpose of the present invention is to provide a method and a skid-mounted system for predicting risks of external transmission pipelines based on big data. The present invention is implemented as follows:
[0005] In a first aspect, the present invention provides a method for predicting the risk of an external transmission pipeline based on big data, the method comprising: obtaining a characterization array of an external transmission pipeline sensor data set, the characterization array of the external transmission pipeline sensor data set being used to characterize the pipeline operation status of the external transmission pipeline sensor data set; determining one or more pending risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set; determining an abnormal characterization coefficient of each of the pending risk data items, the abnormal characterization coefficient being used to characterize the degree of abnormality of the pending risk data item in the external transmission pipeline sensor data set; determining a risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items; and obtaining risk estimation information of the external transmission pipeline based on the risk data item detection result.
[0006] In a second aspect, the present invention provides a skid-mounted system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.
[0007] The beneficial effects of the present invention include at least the following contents: after the present invention completes the extraction of pending risk data items of the external transmission pipeline sensor data set, the abnormality degree of the pending risk data items is judged, and the extracted risk data items are classified according to the abnormality degree. In this way, in the process of selecting risk data items based on the abnormal characterization coefficients of the pending risk data items, not only can the misidentified risk data items with low abnormality degrees in the pending risk data items be discarded, thereby increasing the accuracy of the risk prediction neural network extraction results, but also the key risk data items with high abnormality degrees in the pending risk data items can be extracted, thereby improving the ability of the risk prediction neural network to extract risk data items, distinguishing the correlation between the risk data items and the external transmission pipeline sensor data set, and being able to characterize the key operating status of the external transmission pipeline sensor data set based on the key risk data items, so as to accurately identify risks and ensure the risk prediction accuracy of the external transmission pipeline. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 It is a flow chart of a method for predicting risk of an external transmission pipeline based on big data provided in an embodiment of the present invention.
[0009] Figure 2 It is a schematic diagram of the composition of a skid-mounted system provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0010] The execution subject of the risk prediction method for transmission pipelines based on big data in the embodiment of the present invention is the data processing system in the skid-mounted system. As a highly integrated data acquisition and processing device, the skid-mounted system is widely used in oil and gas fields, industrial monitoring, environmental monitoring and other fields. It mainly includes: sensor interface module: This module is responsible for connecting various sensors, such as pressure sensors, temperature sensors, flow sensors, etc., to collect real-time data from the site. These sensors are usually installed near the equipment or environment to be monitored, and transmit the data to the skid-mounted system by wired or wireless means. Data acquisition unit: The acquisition unit is responsible for receiving data from the sensor interface module and performing preliminary processing, such as signal amplification, filtering, analog-to-digital conversion, etc., to ensure the accuracy and reliability of the data. The unit usually includes a high-performance analog-to-digital converter (ADC) and necessary signal processing circuits. Data processing and control center: This is the core part of the skid-mounted system, which consists of a high-performance processor, storage device and control software. The processor is responsible for executing complex algorithms and performing in-depth analysis, processing and storage of the collected data. The control software is responsible for the overall operation and management of the system, including the scheduling of data acquisition, the control of processing procedures and communication with external systems. Power supply module: Provides stable and reliable power support for the entire skid-mounted system, usually including AC / DC converters, battery packs and power management systems to ensure that the system can work properly in various environments. Communication interface module: This module supports multiple communication protocols, such as Ethernet, Wi-Fi, 4G / 5G, satellite communications, etc., enabling the skid-mounted system to exchange data with a remote monitoring center, other equipment or a cloud platform to achieve remote monitoring and fault diagnosis. User interaction interface: Provides an intuitive operation interface that allows on-site operators to view real-time data, set system parameters, monitor system status, etc. The interface may include input and output devices such as display screens, keyboards, and touch screens.
[0011] The embodiment of the present invention provides a method for predicting risks of transmission pipelines based on big data. Figure 1 As shown, the method includes:
[0012] Step S100: Acquire a characterization array of an external transmission pipeline sensor data set, wherein the characterization array of the external transmission pipeline sensor data set is used to characterize a pipeline operation state of the external transmission pipeline sensor data set.
[0013] In an application scenario example, such as an external pipeline system in an oil or gas transmission network, the system uses a series of sensors installed on the pipeline to monitor key parameters in real time, such as pressure, temperature, flow, and possible vibration or sound. These sensor data are centrally collected and processed by a skid-mounted system (an integrated data acquisition and processing device), forming a huge sensor data set.
[0014] During the execution of step S100, first, the skid-mounted system collects data from sensors distributed throughout the pipeline at regular intervals. These data include real-time pressure values (such as P1=120bar, P2=118bar...), temperature readings (such as T1=30°C, T2=32°C...), flow rates (such as Q1=500m³ / h, Q2=480m³ / h...), etc. These data are recorded in the form of time series, and the data at each time point constitutes a data point. The collected raw data may contain noise, missing values or outliers, so the computer system first cleans and preprocesses these data. For example, missing values are filled by interpolation methods, noise is removed by filtering techniques, and thresholds are set to identify and process outliers. After preprocessing, the computer system begins to extract features from the cleaned data and organizes these features into characterization arrays. In this scenario, the features may include the average, maximum, minimum, and standard deviation of pressure over a period of time, the rate of change of temperature, and the correlation coefficient between flow and pressure. Each feature corresponds to an element in the array, and all features together constitute a multidimensional vector or matrix, namely the representation array.
[0015] For example, for the data of a certain section of pipeline in a specific time period, its characterization array may be as follows (simplified in vector form): [average pressure, maximum pressure, minimum pressure, pressure standard deviation, temperature change rate, flow pressure correlation coefficient] = [119.5bar, 122bar, 117bar, 2.3bar, 0.05h, 0.92]. This feature vector is the characterization array of the sensor data set of the internal and external transmission pipeline in this time period, which comprehensively and concisely describes the operating status of the pipeline. The constructed characterization array is stored in the database of the computer system and is ready for the subsequent risk prediction step. These characterization arrays not only contain a snapshot of the current pipeline operating status, but also can provide training samples for the machine learning model together with historical data to continuously optimize the accuracy of risk prediction.
[0016] Step S200: determining one or more pending risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set.
[0017] In step S200, the computer system uses the characterization array of the external transmission pipeline sensor data set to identify and determine one or more data items that may have risks, and these data items are called pending risk data items.
[0018] For example, in a large oil and gas field's external transmission pipeline system, the system monitors the pipeline's operating status in real time through a dense network of sensors. After the sensor data is processed, it forms a representation array containing multiple feature dimensions, which fully reflects the changes in key parameters such as pressure, temperature, and flow of the pipeline.
[0019] First, the computer system loads the characterization arrays obtained through step S100 from the database. These arrays contain the operating status information of the pipeline at different time points and are the basis for risk prediction. In order to determine the pending risk data items, the computer system may use one or more machine learning models (such as random forests, support vector machines, neural networks, etc.) to analyze the patterns in the characterization arrays. These models are trained to identify feature combinations or data patterns associated with known risk events.
[0020] For example, if historical data shows that when the pressure in a certain section of the pipeline rises sharply and the temperature drops abnormally, it often indicates a possible leakage risk, then the trained machine learning model will learn the mapping relationship between this feature combination and risk events.
[0021] Using the trained model, the computer system traverses and analyzes the current representation array to identify data items that match the risk pattern. These data items are considered to be pending risk data items because they may indicate potential risk events. In actual operation, the model can output a risk score or probability value for each data item, indicating the likelihood that the data item belongs to the risk category. The computer system can set thresholds based on these scores and determine data items with scores above the threshold as pending risk data items.
[0022] In addition to the analysis of a single feature, the computer system can also consider the combination of features and the location of data items in context. For example, multiple consecutive pressure anomaly data points may be more likely to represent a real risk event than a single isolated data point. Finally, the computer system outputs and records the determined pending risk data items for more detailed analysis and processing in subsequent steps. These pending risk data items may include detailed information such as specific time periods, pipeline segments, and related feature values.
[0023] Step S300: determining an abnormal characterization coefficient of each of the pending risk data items, wherein the abnormal characterization coefficient is used to characterize the abnormal degree of the pending risk data item in the external transmission pipeline sensor data set.
[0024] In step S300, the computer system deeply analyzes each pending risk data item to assign it an abnormal characterization coefficient, which quantitatively reflects the degree of abnormality of the data item in the external transmission pipeline sensor data set, and helps to filter out possible false alarms or low-risk events.
[0025] First, the computer system extracts key features from the corresponding representation array for each pending risk data item. These features may include the value of the data item itself (such as pressure value, temperature value), the difference with adjacent data items (such as pressure change rate), and the degree of deviation from the historical average level.
[0026] For example, for a pending risk data item, its feature vector may include the current pressure value P=160bar, the pressure change ΔP=5bar from the previous time point, and the deviation of the pressure value from the average pressure value P_avg=120bar in the same time period of the past week |P-P_avg|=40bar. Next, the computer system uses one or more algorithms to calculate the abnormal characterization coefficient of each pending risk data item. These algorithms may be based on statistical principles (such as Z-score, IQR outlier detection), distance metrics (such as Euclidean distance, Manhattan distance to the center of the data set), or more complex machine learning models (such as isolation forest, autoencoder reconstruction error). Taking the isolation forest algorithm as an example, the algorithm isolates data points by constructing multiple decision trees. Abnormal points are often isolated earlier because they are significantly different from most data points. For each pending risk data item, the isolation forest algorithm will output an abnormal score, which is its abnormal characterization coefficient. The higher the score, the more abnormal the data item.
[0027] Suppose the system calculates that the abnormal characterization coefficient of a pending risk data item is 0.9 (a relatively high value in the context of the isolation forest algorithm), while the abnormal characterization coefficients of some other data items are lower, such as 0.1 or 0.3. Based on this result, the computer system can infer that the pending risk data item has a high degree of abnormality and may require immediate attention and corresponding measures, such as increasing the monitoring frequency, dispatching maintenance personnel to check, etc. On the contrary, data items with lower abnormal characterization coefficients may be regarded as low risk or false alarms, and the system can adopt a more relaxed monitoring strategy.
[0028] Step S400: Determine the risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items.
[0029] In step S400, the computer system uses the abnormal characterization coefficients of each pending risk data item calculated in step S300 to further analyze and determine the risk data item detection results of the external transmission pipeline sensor data set, and converts the quantified abnormality degree into a specific risk assessment conclusion.
[0030] First, the computer system sets one or more thresholds based on business needs and historical experience to determine whether the abnormality of the pending risk data item has reached a level that requires attention. These thresholds can be fixed values or dynamically adjusted based on the characteristics of the data set. For example, the system may set a threshold for the abnormal characterization coefficient to 0.7, and consider data items above this threshold to have a higher risk potential.
[0031] In addition to simple threshold judgment, the computer system can also classify data items into different risk levels based on the size of the abnormal characterization coefficient. For example, data items with an abnormal characterization coefficient between 0.5 and 0.7 are considered low risk, those between 0.7 and 0.9 are considered medium risk, and those above 0.9 are considered high risk.
[0032] In order to prioritize the most urgent risk events, the computer system can also sort the pending risk data items according to the abnormal characterization coefficient. After sorting, the system can choose to focus on the top-ranked data items, or set a processing upper limit (such as the top 10% of data items) based on resource allocation.
[0033] Assume that in step S300, the computer system calculates that the abnormal characterization coefficient of a pending risk data item is 0.85, which is higher than the set threshold of 0.7. According to the risk level classification rules, this data item is considered to be medium-high risk. At the same time, because this data item is located at a key node of the pipeline network and the surrounding environmental factors (such as geological conditions and climatic conditions) also increase the possibility of risk, the system ranks it as one of the risk items that need to be paid attention to at present. Finally, the risk data item detection results will be output, and it is recommended to immediately dispatch professionals to the site for inspection and maintenance.
[0034] Step S500: obtaining risk estimation information of the external transmission pipeline based on the risk data item detection result.
[0035] In step S500, the computer system integrates the risk data item detection results generated in step S400, further refines and generates risk assessment information about the export pipeline. This process converts the specific risk data item analysis results into a comprehensive assessment of the risk status of the entire pipeline or a specific section, providing decision support for the operation and maintenance team and management. For example, a large oil and gas export pipeline system spans multiple provinces and undertakes important energy transportation tasks. In order to ensure the safe and stable operation of the pipeline, the computer system continuously monitors and analyzes massive data from sensors along the line. After the previous steps, the system has identified and evaluated multiple potential risk data items. Now it is necessary to integrate these specific risk points to form a comprehensive estimate of the risk status of the entire pipeline.
[0036] First, the computer system summarizes all data items that are determined to be high risk, medium risk, or low risk, including their location, abnormality type, degree of abnormality, and other information. This information constitutes the basic data for risk estimation. Next, the system uses geographic information system (GIS) technology to mark the risk data items on the pipeline map and analyze their spatial distribution patterns. By observing the aggregation of risk points, the trend of changes along the pipeline, etc., the system can preliminarily determine which sections of the pipeline have higher risk hazards. In order to obtain more accurate risk estimation information, the computer system can apply risk assessment models (such as Bayesian networks, fault tree analysis, etc.) to conduct a comprehensive analysis of the summarized risk data items. These models can consider the interaction between multiple risk factors and their impact on the overall risk level of the pipeline.
[0037] For example, the Bayesian network model can calculate the probability of failure of a pipeline under specific conditions by constructing a conditional probability relationship between risk factors. The system can treat each risk data item as an evidence node, update the state of the entire network according to their abnormal characterization coefficients and prior probabilities, and finally obtain the estimated risk probability of the pipeline. Based on the output results of the risk assessment model, the computer system divides the risk status of the pipeline into different levels (such as low, medium, and high) and sets corresponding warning thresholds. Once the estimated risk level of the pipeline reaches or exceeds the warning threshold, the system will automatically trigger the warning mechanism and notify the operation and maintenance team to take necessary measures. Finally, the computer system organizes the results of the risk assessment into a report, including a detailed list of risk data items, pipeline risk maps, risk assessment model parameters, risk level classification, and warning information. This report will serve as an important basis for the operation and maintenance team to formulate maintenance plans and emergency plans.
[0038] As an implementation manner, the step S300 of determining the abnormal characterization coefficient of each of the pending risk data items may specifically include:
[0039] Step S310: obtaining a data item array representation of each of the pending risk data items;
[0040] Step S320: determining the abnormal characterization coefficient of each of the pending risk data items based on the data item array representation of each of the pending risk data items and the characterization array of the external transmission pipeline sensor data set.
[0041] In step S310, the computer system extracts the specific information of each pending risk data item and converts it into a form that is convenient for subsequent analysis, namely, a data item array representation. This representation is usually a multidimensional array or vector, which contains a series of values describing the characteristics of the data item.
[0042] For example, an oil and gas export pipeline system monitors the key parameters of the pipeline such as pressure, temperature, flow rate, etc. in real time through a dense sensor network. In step S200, the system has identified a batch of pending risk data items based on the sensor data set. These data items may indicate that a part of the pipeline is in an abnormal state, such as abnormal pressure increase, temperature drop, etc. For each pending risk data item, the computer system needs to obtain its detailed data record. For example, a pending risk data item may correspond to the pressure monitoring data of a certain point in the pipeline within a specific time period. These data exist in the form of time series, but in order to perform efficient analysis in step S300, the system needs to convert these time series data into a fixed-length data item array representation. First, the system selects a series of key features based on business needs and data characteristics. For pressure data items, these features may include maximum value, minimum value, average value, standard deviation, change trend (such as slope), etc. Since the dimensions and numerical ranges of different features may vary greatly, direct comparison will lead to deviations. Therefore, the system needs to normalize each feature and convert it to the same numerical range (such as between 0 and 1). After feature selection and normalization, the system arranges these feature values into a vector or matrix in a certain order to form a data item array representation. This data item array representation not only captures the main features of the data item, but also eliminates the dimensional differences between different features, providing a basis for the subsequent calculation of the abnormal characterization coefficient.
[0043] In step S320, the computer system uses the data item array representation of each pending risk data item, combined with the global characterization array of the external transmission pipeline sensor data set, to determine the abnormal characterization coefficient of each data item through a complex calculation process. This process usually involves a variety of mathematical and statistical methods, as well as possible machine learning techniques.
[0044] The calculation of the abnormal characterization coefficient depends on how to define the abnormality. In the scenario of risk prediction of external transmission pipelines, abnormalities usually refer to data items that deviate significantly from the normal operating state. Therefore, the calculation of the abnormal characterization coefficient requires comprehensive consideration of the characteristics of the data item itself and its position in the global data set.
[0045] Method 1: Measurement based on statistical distance. This method uses statistical distance to measure the distance between each undetermined risk data item and the center of the global data set (such as the mean vector). The greater the distance, the more abnormal the data item. First, the system calculates the mean vector of the characterization array of the external transmission pipeline sensor data set, which represents the overall central trend of the data set.
[0046] ;
[0047] in, is the mean vector, N is the number of data items in the data set, xi is an array representation of each data item (converted to the same dimension).
[0048] Then, for each pending risk data item, the system calculates the Euclidean distance (or other distance measure) between its data item array representation and the mean vector.
[0049] ;
[0050] Where j is the Euclidean distance, D is the dimension of the data item array representation, and x jk is the value of the jth undetermined risk data item in the kth dimension, μ k is the value of the mean vector in the kth dimension.
[0051] Finally, the system normalizes all calculated distances (eg, by dividing by the maximum distance) to obtain an abnormal characterization coefficient for each pending risk data item.
[0052] Method 2: Isolation Forest-based Anomaly Detection. This method uses the Isolation Forest algorithm to calculate the anomaly characterization coefficient. Isolation Forest is a tree-based anomaly detection method that isolates anomalies by recursively and randomly splitting the data space.
[0053] First, the system uses the representation array of the external pipeline sensor data set (or a portion of the samples extracted from it) to train the isolation forest model. During the training process, the model learns the intrinsic structure of the data set and constructs multiple isolation trees. For each pending risk data item, the system inputs its data item array representation into the trained isolation forest model and calculates the path length of the data item on each tree. The shorter the path length, the easier it is to isolate the data item, and thus the more likely it is to be an outlier. Based on the average path length of the data item on all trees, the system calculates an anomaly score as the anomaly characterization coefficient of the data item. The calculation of the anomaly score typically involves comparing the actual path length with the path length expected in a random situation.
[0054] ;
[0055] Among them, E(h(x j )) is the data item x j The average path length over all trees. c(n) is a factor (usually a constant) that normalizes the path length for a given number of samples n.
[0056] It can be understood that the formula here is a simplified representation. The calculation of anomaly scores in the actual isolation forest algorithm may be more complicated, involving multiple factors such as logarithmic functions and comparison of path length and tree height.
[0057] Method 3: Comprehensive evaluation combining global and local information. In some cases, in order to more accurately assess the degree of abnormality of the pending risk data item, the computer system can combine the global representation array and the local neighborhood information for comprehensive evaluation. For example, the system can first use the global representation array to determine the approximate location of the data item, and then analyze its degree of deviation in the local neighborhood.
[0058] This comprehensive evaluation method may require the design of more complex algorithms or models, such as density-based clustering algorithms (such as DBSCAN), local outlier factor (LOF) algorithms, etc. These algorithms can consider the relationship between data items and their neighboring points, so as to more accurately identify those outliers that are far away from the global center and appear isolated in the local area.
[0059] As an implementation mode, the characterization array of the external transmission pipeline sensor data set includes a first characterization array, and the first characterization array is used to characterize the global pipeline operation status of the external transmission pipeline sensor data set; then, at this time, the step S320 determines the abnormal characterization coefficient of each of the pending risk data items based on the data item array representation of each of the pending risk data items and the characterization array of the external transmission pipeline sensor data set, which may specifically include: determining the abnormal characterization coefficient of each of the pending risk data items based on the commonality measurement result between the data item array representation of each of the pending risk data items and the first characterization array.
[0060] In the specific implementation of step S320, the computer system uses the commonality measurement results between the first characterization array (global characterization) of the external transmission pipeline sensor data set and the data item array representation of each pending risk data item to determine the abnormal characterization coefficient of each pending risk data item. This process quantitatively evaluates the degree of abnormality of the data item by comparing the similarity between the local (single data item) and the global (entire data set). For example, an oil and gas external transmission pipeline system spanning multiple provinces monitors the operating status of the pipeline in real time through a dense sensor network, including multiple key parameters such as pressure, temperature, and flow. After processing, these sensor data form a characterization array containing rich feature information, in which the first characterization array is used as a global characterization, which comprehensively reflects the overall operating status of the entire pipeline system within a certain period of time.
[0061] In order to determine the abnormal characterization coefficient of the pending risk data items, the computer system needs an effective commonality measurement method to evaluate the similarity or difference between each data item and the global characterization. The following are some possible commonality measurement methods and their application scenarios:
[0062] Method 1: Cosine similarity. This is an indicator that measures the degree of similarity between two vectors in direction. It does not consider the size of the vectors, but only focuses on the difference in direction. In step S320, the computer system can regard the data item array representation of each pending risk data item as a vector and perform cosine similarity calculation with the first representation array vector.
[0063] ;
[0064] Where A is the data item array representation vector of the undetermined risk data item, B is the first representation array vector, · represents the dot product operation, Represents the magnitude of a vector.
[0065] In the anomaly detection scenario, we often pay more attention to the difference between the data item and the global representation rather than the similarity. Therefore, the computer system can obtain a difference index by subtracting the cosine similarity from 1, which is then used as part of the anomaly representation coefficient.
[0066] Method 2: Mahalanobis distance. Mahalanobis distance is a distance measurement method that takes into account the correlation between data. It calculates the distance between two data points based on the covariance matrix of the data. In the scenario of external transmission pipeline risk prediction, if there is a complex correlation structure between the first characterization array and the data item array representation, it may be more appropriate to use the Mahalanobis distance. Direct calculation of the Mahalanobis distance may be computationally intensive and requires knowledge of the covariance matrix of the global data. In practical applications, the computer system can use approximate methods or pre-calculated covariance matrices to simplify the calculation process.
[0067] Method 3: Commonality measurement based on custom features. In addition to directly using the similarity or distance between vectors, the computer system can also design custom features to measure the commonality between the undetermined risk data items and the global representation based on the characteristics of the external pipeline sensor data. For example, features can be designed to compare the difference between the data item and the global mean, median, mode and other statistics, or to compare the consistency of the change trend of the data item in a specific dimension with the global trend.
[0068] In actual operation, step S320 may include: ensuring that the data item array representation of the pending risk data item is consistent with the first characterization array in dimension for subsequent calculations. Selecting an appropriate commonality measurement method (such as cosine similarity, Mahalanobis distance or custom feature measurement) to calculate the commonality measurement results between each pending risk data item and the first characterization array. Based on the commonality measurement results, determine the abnormal characterization coefficient of each pending risk data item through certain conversion or mapping rules (such as negation, normalization, etc.). The higher the abnormal characterization coefficient, the greater the difference between the data item and the global characterization, and the higher the degree of abnormality.
[0069] For example, suppose that in the sensor data set of the external transmission pipeline, the first characterization array comprehensively reflects the key parameters such as the average pressure, temperature and flow rate of the pipeline system in a certain period of time. For a certain pending risk data item, its data item array represents the record of these parameter values of a certain section of the pipeline at a specific time point. The computer system uses cosine similarity as a commonality measurement method to calculate the similarity between the data item and the first characterization array, and subtracts the similarity from 1 to obtain a difference index. The system then normalizes the difference index to between 0 and 1 as the abnormal characterization coefficient of the data item. If the abnormal characterization coefficient is close to 1, it means that the data item is significantly different from the global characterization, which may indicate potential risks.
[0070] As an implementation mode, the step S400 determines the risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items, and can specifically include: determining the pending risk data items whose abnormal characterization coefficients meet the first evaluation requirements as the risk data items included in the external transmission pipeline sensor data set.
[0071] The first evaluation requirement may be one or more of the following conditions:
[0072] Condition 1: The abnormal characterization coefficient is not less than the first reference value;
[0073] Condition 2: The first s abnormal characterization coefficients are arranged in descending order according to the size of the abnormal characterization coefficients, where s≥1.
[0074] In step S400, the computer system sets the first evaluation requirement to screen out the data items that truly represent risks based on the abnormal characterization coefficient of each pending risk data item, thereby determining the risk data item detection results of the external transmission pipeline sensor data set. This process directly determines which data items will be regarded as potential risk sources, thereby providing a basis for subsequent operation and maintenance decisions. For example, a complex oil and gas external transmission pipeline system monitors various key parameters such as pressure, temperature, flow, etc. in real time through sensors throughout the pipeline. In the previous steps, the system has identified a batch of pending risk data items and calculated abnormal characterization coefficients for them. Now, the system needs to determine which data items truly represent risks based on these abnormal characterization coefficients.
[0075] For condition 1, in order to filter out data items with a high degree of abnormality, the computer system may set a first reference value as a threshold. Any data item whose abnormal characterization coefficient is not less than the threshold will be regarded as a risky data item.
[0076] For condition 2, in addition to setting the threshold, the computer system can also sort all pending risk data items according to the size of the abnormal characterization coefficient and select the top several items as risk data items. Here, s is a preset positive integer, indicating the number of risk data items that need attention.
[0077] In practical applications, the computer system can use both condition 1 and condition 2 to comprehensively evaluate the pending risk data items. For example, the system can first set a relatively high first reference value (such as 0.9) to filter out data items with extremely high abnormality; then, for the remaining data items, sort them according to the abnormal characterization coefficient and select the top several items as supplementary risk data items.
[0078] After the assessment is completed, the computer system will output the detection results of the risk data items, which usually include the specific location of the data item (such as pipeline segment number, sensor number), abnormal characterization coefficient, possible risk type and other information. This information will be used in subsequent operation and maintenance decisions, such as increasing the monitoring frequency, dispatching maintenance personnel to check, adjusting operating parameters, etc. In this way, the computer system can efficiently and accurately filter out the data items that truly represent the risk from a large number of pending risk data items, providing strong support for the safe and stable operation of the external transmission pipeline system.
[0079] As another implementation, the step S400, based on the abnormal characterization coefficient of each of the pending risk data items, determines the risk data item detection result of the external transmission pipeline sensor data set, and may include: determining the pending risk data items whose abnormal characterization coefficients meet the second evaluation requirements as key risk data items included in the external transmission pipeline sensor data set.
[0080] The second evaluation requirement may be one or more of the following conditions:
[0081] Condition 1: The abnormal characterization coefficient is not less than the second reference value;
[0082] Condition 2: The first r abnormal characterization coefficients are arranged in descending order according to the size of the abnormal characterization coefficients, where r≥1.
[0083] In the above another embodiment of step S400, the computer system not only focuses on screening out risky data items, but also further focuses on identifying key risk data items with extremely high abnormality and having a significant impact on the overall risk prediction results. This process is achieved by setting a second evaluation requirement, which is more stringent than the first evaluation requirement, to ensure that the selected data items can truly represent the high-risk points in the external transmission pipeline sensor data set.
[0084] For example, a long-distance oil and gas transmission pipeline system faces multiple potential risks due to crossing different geological environments and climatic conditions. In order to ensure the safe and stable operation of the pipeline, the computer system needs to accurately predict and prioritize the key risk points that may lead to serious consequences. In the previous steps, the system has calculated the abnormal characterization coefficients of each pending risk data item, and now it will determine the key risk data items based on these coefficients.
[0085] For condition 1, in order to filter out data items with extremely high abnormality, the computer system sets a second reference value, which is usually higher than the first reference value. Any data item with an abnormal characterization coefficient not less than the threshold will be regarded as a key risk data item.
[0086] For condition 2, in addition to setting the threshold, the computer system can also sort all pending risk data items according to the size of the abnormal characterization coefficient and select the top r data items as key risk data items. Here, r is a positive integer set according to actual needs, which is usually much smaller than the total number of data items to ensure that the selected data items are highly representative and important.
[0087] In practical applications, the computer system can comprehensively evaluate the pending risk data items by combining conditions 1 and 2. For example, the system can first set a higher second reference value to screen out candidate data items with extremely high abnormality; then, the candidate data items are further sorted according to the abnormal characterization coefficients, and the top r data items are selected as the final key risk data items.
[0088] In this way, the computer system can accurately identify key risk points in the external pipeline sensor data set, provide strong support for the operation and maintenance team, and ensure the safe and stable operation of the pipeline system.
[0089] As an implementation manner, the step S200, based on the characterization array of the external transmission pipeline sensor data set, determines one or more undetermined risk data items in the external transmission pipeline sensor data set, which may specifically include:
[0090] Step S210: determining a pipeline operation state type feature of the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set, wherein the pipeline operation state type feature of the external transmission pipeline sensor data set is used to characterize the pipeline operation state corresponding to the external transmission pipeline sensor data set;
[0091] Step S220: determining one or more pending risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set and the pipeline operation status type characteristics of the external transmission pipeline sensor data set.
[0092] In step S210, the primary task of the computer system is to analyze the characterization array of the external transmission pipeline sensor data set to extract type features that can characterize the pipeline operation status. These features usually reflect the behavior pattern of the pipeline under different conditions and are the basis for subsequent risk identification.
[0093] For example, a complex oil and gas transmission pipeline system uses a dense network of sensors to monitor various parameters of the pipeline in real time, such as pressure, temperature, flow, vibration, etc. After preprocessing and feature extraction, these sensor data form a representation array that comprehensively records the operating status of the pipeline. In order to identify potential risks, the computer system needs to first understand the patterns behind these states.
[0094] In the absence of clear labels, computer systems can use unsupervised learning methods (such as cluster analysis, principal component analysis, etc.) to discover the inherent structure in the data. For example, by clustering the representation array using the K-means clustering algorithm, similar pipeline operation states can be divided into the same category, and the center or representative features of each category can be used as pipeline operation state type features.
[0095] If there are some labeled pipeline operation status data in the system, the computer system can use this data to train classification models (such as support vector machines, random forests, neural networks, etc.). The model can automatically extract features that are useful for classification tasks by learning the mapping relationship between input features (i.e., representation arrays) and output labels (i.e., pipeline operation status types). These features will play an important role in the subsequent risk identification process.
[0096] In some cases, the computer system may also need to manually design features in combination with domain knowledge. For example, based on the physical principles of oil and gas transportation, features such as pressure gradient, temperature change rate, and flow volatility can be designed. These features can more intuitively reflect the operating status of the pipeline.
[0097] Assume that after feature extraction, the computer system obtains the following pipeline operation status type features: average pressure gradient, temperature fluctuation range, flow anomaly index. These features together constitute the pipeline operation status type feature set.
[0098] In step S220, the computer system uses the pipeline operation status type features extracted in step S210 and the original characterization array to identify the pending risk data items in the external transmission pipeline sensor data set. These data items may indicate that the pipeline is about to or has already entered an abnormal state, requiring further attention and analysis.
[0099] For example, let's continue with the oil and gas transmission pipeline system. After determining the pipeline operation status type characteristics, the computer system needs to filter out those data items that may be at risk based on these characteristics and characterization arrays. These data items may correspond to abnormal pressure increases, temperature drops, sudden changes in flow, etc. in a certain part of the pipeline.
[0100] The computer system can set one or more thresholds for each feature. When the value of a data item on a certain feature exceeds or falls below the set threshold, the data item is considered to be at risk. For example, a pressure gradient threshold can be set. When the pressure gradient of a section of pipeline exceeds the threshold, the section of pipeline is considered to be at risk of pressure anomaly. In addition to simple threshold judgments, computer systems can also use more complex anomaly detection algorithms (such as isolation forests, local anomaly factors LOF, etc.) to identify abnormal data items. These algorithms evaluate the abnormality of data items by analyzing the degree of deviation from global or local data sets.
[0101] In some cases, the computer system can also use a known risk pattern library to match the current data item. If the feature combination of a data item matches a pattern in the risk pattern library, the data item is considered to be risky. In practical applications, the computer system can combine multiple methods to comprehensively evaluate the risk of data items. For example, preliminary risk candidates can be screened out through threshold judgment, and then the candidates can be further screened and sorted using anomaly detection algorithms. Finally, the sorted data items can be manually reviewed or automatically decided based on domain knowledge.
[0102] Through the implementation of the above steps S210 and S220, the computer system can efficiently and accurately identify potential risk data items from the external transmission pipeline sensor data set, providing strong support for subsequent risk prediction and response.
[0103] As an implementation manner, the step S210, determining the pipeline operation status type feature of the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set, may include:
[0104] Step S211: determining the confidences corresponding to K pipeline operation status types respectively based on the characterization array of the external transmission pipeline sensor data set; wherein the confidence corresponding to the x-th pipeline operation status type among the K pipeline operation status types is used to characterize the probability that the external transmission pipeline sensor data set belongs to the x-th pipeline operation status type, 1≤x≤K;
[0105] Step S212: based on the confidences respectively corresponding to the K pipeline operation status types, determining u pipeline operation status types with the largest confidences from the K pipeline operation status types, and obtaining the u target pipeline operation status types, where u≥1;
[0106] Step S213: Based on the confidences corresponding to the u target pipeline operation status types, the representation arrays corresponding to the u target pipeline operation status types are fused to obtain a fused representation array; wherein the pipeline operation status type features of the external transmission pipeline sensor data set include the fused representation array.
[0107] In step S211, the core task of the computer system is to analyze the characterization array of the external transmission pipeline sensor data set to determine the confidence that the data set belongs to each of the K predefined pipeline operation status types. The confidence here is essentially a probability estimate, which indicates the possibility that the data set belongs to a specific operation state. For example, an oil and gas transmission pipeline system spanning multiple provinces monitors the key parameters of the pipeline in real time through a dense sensor network, such as pressure, temperature, flow, etc. After preprocessing, these sensor data form a characterization array that comprehensively records the operation status of the pipeline. In order to accurately predict risks, the system needs to first determine the type of operation state the pipeline is currently in.
[0108] Assume that the system predefines the following K pipeline operation status types: status type 1: normal operation; status type 2: slight leakage; status type 3: severe leakage; ...; status type K: other abnormal status. In order to calculate the confidence of each status type, the computer system can use a variety of machine learning classification algorithms, such as logistic regression, support vector machine (SVM), random forest, neural network, etc. These algorithms can extract useful information from the representation array and learn the mapping relationship between input features and output labels (i.e. pipeline operation status type). Taking the random forest algorithm as an example, the algorithm integrates learning by constructing multiple decision trees. Each decision tree will split based on the features in the representation array until the preset stopping condition is reached (such as the number of samples in the leaf node reaches the threshold). During the training process, the algorithm will use the data set with known labels to optimize the splitting rules of each tree, so that the entire forest can achieve the best classification ability for unknown data.
[0109] When classifying new data (i.e., the representation array of the external transmission pipeline sensor data set), each decision tree will output a prediction result (i.e., the type of state the pipeline may be in) and the corresponding probability estimate (i.e., confidence). The random forest algorithm combines the prediction results of all decision trees through a voting mechanism (for classification problems) or an averaging mechanism (for regression problems) to obtain the final classification result and confidence estimate. Suppose that after the random forest algorithm analyzes the representation array of the external transmission pipeline sensor data set, it obtains the following confidence estimates: state type 1 (normal operation): 0.1; state type 2 (minor leakage): 0.7; state type 3 (serious leakage): 0.15; ...; state type K (other abnormal states): 0.05. These confidence values reflect the probability that the data set belongs to each predefined state type.
[0110] In step S212, the computer system selects u types with the highest confidence from the K predefined pipeline operation status types according to the confidence calculated in step S211. The purpose of this step is to reduce the complexity of subsequent processing while retaining the most likely status type for further analysis.
[0111] The screening process is relatively simple and intuitive. The computer system only needs to sort the confidences obtained in step S211 in descending order, and then select the first u state types after sorting.
[0112] Continuing with the above confidence example, assuming u=2, the computer system will screen out the two state types with the highest confidence: state type 2 (minor leakage) and state type 3 (serious leakage). These two state types will serve as the focus of subsequent analysis.
[0113] In step S213, the computer system fuses the representation arrays corresponding to the u target pipeline operation status types screened out in step S212 to obtain a comprehensive and more representative fused representation array. This step helps to further refine the key information of the pipeline operation status and provide stronger support for subsequent risk prediction. The choice of fusion method depends on the specific application scenario and data characteristics. Common fusion methods include weighted summation, principal component analysis (PCA), autoencoder reconstruction, etc. The following is an example of weighted summation. In the weighted summation method, the computer system first assigns a weight to each target state type, which is usually proportional to its confidence. Then, the system performs weighted processing on the representation array corresponding to each state type, and adds the weighted arrays to obtain a fused representation array. Continuing with the above-screened two target state types (state type 2 and state type 3) as an example, assuming that their confidences are 0.7 and 0.15 respectively, these two confidence values can be used as weights for weighted summation.
[0114] In practical applications, since the representation arrays of different state types may contain different feature sets, it may not be appropriate to directly perform weighted summation. In this case, the representation arrays of each state type can be first subjected to dimensionality reduction processing (such as using PCA), and converted to the same low-dimensional space, and then weighted summation can be performed. The fused representation array is the final output of step S210. It integrates the information of multiple high-confidence state types and can more comprehensively reflect the current operating status of the pipeline. In the subsequent risk prediction step, the computer system will further identify potential risk data items based on this fused representation array.
[0115] Through the implementation of step S210, the computer system successfully extracted the pipeline operation status type characteristics from the characterization array of the external transmission pipeline sensor data set. This process not only relies on advanced machine learning classification algorithms to accurately calculate the confidence of each status type, but also extracts key information through reasonable screening and fusion methods. The final fusion characterization array provides strong support for subsequent risk prediction, enabling the computer system to more accurately identify potential risk data items and take corresponding countermeasures.
[0116] It should be noted that the specific algorithm selection (such as the random forest algorithm), fusion method (such as weighted summation) and parameter setting (such as the value of u) in the above implementation are all flexibly adjusted according to the application scenario and data characteristics. In practical applications, the computer system may need to go through multiple trials and optimizations to find the most suitable solution for the current task.
[0117] As an implementation manner, the characterization array of the external transmission pipeline sensor data set includes characterization arrays corresponding to a plurality of data items in the external transmission pipeline sensor data set; then, the step S220, based on the characterization array of the external transmission pipeline sensor data set and the pipeline operation state type characteristics of the external transmission pipeline sensor data set, determines one or more pending risk data items in the external transmission pipeline sensor data set, including:
[0118] Step S221: combining the representation arrays corresponding to the respective data items in the external transmission pipeline sensor data set with the pipeline operation status types of the external transmission pipeline sensor data set to obtain the combined representation arrays corresponding to the respective data items;
[0119] Step S222: Based on the combined characterization array, one or more undetermined risk data items in the external transmission pipeline sensor data set are determined.
[0120] In step S221, the primary task of the computer system is to combine the representation array corresponding to each data item in the external transmission pipeline sensor data set with the pipeline operation status type characteristics of the entire data set, thereby constructing a unique combined representation array for each data item. This process aims to combine the global pipeline operation status information with the local data item characteristics to provide a more comprehensive information basis for subsequent risk identification.
[0121] For example, first, the computer system extracts the representation array corresponding to each data item from the external pipeline sensor data set. These representation arrays may contain direct measurement parameters such as pressure value, temperature value, flow value, etc., and may also contain derived features obtained after preprocessing and feature engineering, such as pressure gradient, temperature change rate, etc.
[0122] In step S210, the computer system has determined the pipeline operation status type characteristics of the external transmission pipeline sensor data set. These characteristics may exist in the form of a fused representation array, which comprehensively reflects the key information of the pipeline operation status type to which the entire data set belongs. Next, the computer system combines the representation array corresponding to each data item with the pipeline operation status type characteristics. This combination can be a simple splicing operation, that is, connecting the representation array and the fused representation array end to end to form a larger array; it can also be a more complex fusion process, such as combining the two by weighted summation, feature mapping, etc. Regardless of the method used, the purpose is to integrate the global pipeline operation status information with the local data item characteristics.
[0123] Assume that the representation array corresponding to a data item (D_i) in the external transmission pipeline sensor data set is ([P_i, T_i, Q_i]), where (P_i), (T_i), and (Q_i) represent the pressure value, temperature value, and flow value of the data item, respectively. At the same time, assume that the pipeline operation status type feature (fusion representation array) obtained in step S210 is ([F_1, F_2, ..., F_n]). Then, in step S221, the computer system can combine these two arrays to form a combined representation array ([P_i, T_i, Q_i, F_1, F_2, ..., F_n]) of (D_i).
[0124] In step S222, the computer system identifies the pending risk data items in the external transmission pipeline sensor data set based on the combined characterization array constructed in step S221. This process usually involves complex operations such as pattern matching, anomaly detection or classification decision-making, aiming to screen out those items that may indicate risks from a large number of data items.
[0125] Computer systems can use anomaly detection algorithms (such as isolation forests, local outlier factors LOF, etc.) to analyze abnormal patterns in the combined representation array. These algorithms can identify items that are significantly different from most data items as anomalies, which are likely to be potential risk data items.
[0126] If there are known risk data item samples in the system as training sets, the computer system can use these samples to train a classification model (such as support vector machine, random forest, neural network, etc.). Then, the combined representation array is input into the classification model for prediction to obtain the probability of each data item belonging to the risk category. Based on this probability value, a threshold can be set to determine whether the data item is a risk data item.
[0127] In some cases, computer systems can also combine domain knowledge to design a series of rules to identify risky data items. These rules may be based on specific threshold judgments (such as pressure exceeding a certain value), pattern matching (such as a sudden drop in flow followed by a sustained low value), or complex logical combinations (such as multiple conditions being met at the same time).
[0128] Assume that the computer system uses an anomaly detection algorithm to identify risky data items. Based on the combined representation array constructed in step S221, the algorithm first calculates metrics such as the distance or density between each data item and the remaining data items. Then, the data items are divided into normal areas and abnormal areas based on these metrics. Data items in the abnormal area will be regarded as pending risk data items that require further attention and analysis.
[0129] For example, in an oil and gas transmission pipeline system, a data item's combination representation array shows the simultaneous occurrence of abnormally high pressure values and sharply falling temperature values. This pattern matches a pattern in a known risk pattern library or significantly deviates from the distribution range of normal data, so it is identified as an abnormal item by the algorithm and then determined as a pending risk data item.
[0130] Through the implementation of step S220, the computer system successfully combines each data item in the external transmission pipeline sensor data set with the global pipeline operation status information to construct a combination representation array containing rich information. Based on these combination representation arrays, through complex pattern matching, anomaly detection or classification decision-making operations, potential pending risk data items are identified, providing strong support for subsequent risk prediction and response.
[0131] As an implementation mode, the risk data item detection result of the external transmission pipeline sensor data set is determined by a risk prediction neural network, and the risk prediction neural network includes a characterization information extraction component, a risk data item detection component and an abnormal index determination component; wherein, the characterization information extraction component is used to obtain the characterization array of the external transmission pipeline sensor data set; the risk data item detection component is used to determine one or more pending risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set; and the abnormal index determination component is used to determine the abnormal characterization coefficient of each of the pending risk data items.
[0132] The representation information extraction component is responsible for extracting useful feature information from the original external transmission pipeline sensor data set and encoding this information into the form of a representation array. These representation arrays are the basis for subsequent risk identification and anomaly detection. In practical applications, the representation information extraction component may use a variety of technical means to achieve feature extraction. For example, for time series data (such as pressure, temperature, flow and other sensor data), the component can first perform time series analysis to extract statistical features such as trends, seasonality, and periodicity. In addition, the component can also apply feature engineering technology to generate new and more representative features by combining, transforming or deriving original features. These features may include statistics such as maximum, minimum, mean, standard deviation, rate of change, slope, and derived features based on physical principles or domain knowledge.
[0133] The risk data item detection component uses machine learning algorithms or models to analyze patterns in the data based on the representation array output by the representation information extraction component, and identifies data items that may be risky. These pending risk data items will serve as candidates for subsequent calculation of abnormal representation coefficients.
[0134] The implementation of the risk data item detection component relies on advanced machine learning techniques. Common algorithms include support vector machines (SVM), random forests, gradient boosting trees (GBDT), neural networks (such as convolutional neural networks (CNN), recurrent neural networks (RNN) and their variants LSTM / GRU), etc. These algorithms can learn the mapping relationship between data items and risk status from the representation array, and classify or predict new data items accordingly. During the training phase, the component needs to receive a large number of labeled data sets (i.e., data items with known risk status) and minimize the prediction error by adjusting the algorithm parameters. Taking the convolutional neural network (CNN) as an example, the risk data item detection component first builds a neural network model containing multiple convolutional layers, pooling layers, and fully connected layers. During the training process, the model receives a large number of representation arrays labeled as "normal" or "risky" as input, and adjusts the network weights through the back-propagation algorithm to minimize the classification error (such as cross entropy loss).
[0135] The abnormal indicator determination component further analyzes the pending risk data items identified by the risk data item detection component and quantifies the degree of abnormality of these data items in the overall data set by calculating their abnormal characterization coefficients. The higher the abnormal characterization coefficient, the more likely the data item is to be a real risk source.
[0136] The implementation of the anomaly indicator determination component involves a variety of anomaly detection algorithms or technologies. Common methods include isolation forest, local anomaly factor (LOF), density-based clustering algorithms (such as DBSCAN), and deep learning-based autoencoders or adversarial autoencoders. These methods can evaluate the abnormality of data items from different perspectives, such as based on distance, density, reconstruction error, etc.
[0137] Taking the Isolation Forest algorithm as an example, the anomaly indicator determination component first builds multiple isolation trees to isolate data points. During the training process, the algorithm randomly selects features and data points to build trees until each data point is isolated or reaches the preset tree height limit. Since anomalies are usually easier to isolate (i.e., have a shorter path length), the algorithm is able to calculate the anomaly score of each data point based on the path length. For the undetermined risk data item, the component calculates its anomaly score and uses it as the anomaly characterization coefficient. The higher the anomaly score (or the larger the value obtained after appropriate transformation), the more abnormal the data item.
[0138] Based on this network architecture, the method provided by the embodiment of the present invention also includes a training process of the risk prediction neural network, which may specifically include the following steps:
[0139] Step S10: acquiring a characterization array of an external transmission pipeline sensor data set sample based on the characterization information extraction component, wherein the characterization array of the external transmission pipeline sensor data set sample is used to characterize a pipeline operation state of the external transmission pipeline sensor data set sample.
[0140] In step S10, the computer system performs the following operations to obtain a representation array of samples of the external transmission pipeline sensor data set. First, the system loads samples of the external transmission pipeline sensor data set from a storage medium (such as a database, a file system, or a cloud storage). These samples are a subset of historical data and contain sensor data of known pipeline operating conditions (such as normal, slight leakage, severe leakage, etc.). After loading the data, the system preprocesses the data to eliminate noise, fill missing values, standardize or normalize features, etc. This step is crucial to improving the performance of subsequent machine learning models. Next, the system uses a representation information extraction component to extract features from the preprocessed data. These features may be predefined by domain knowledge or dynamically generated by automatic feature selection or feature learning algorithms. For example, for pressure sensor data, the system may extract the maximum value, minimum value, average value, standard deviation of pressure, and the pressure change rate between adjacent time windows as features. The extracted features are encoded in the form of vectors or arrays to form a representation array. Each representation array corresponds to a data item in the data set (such as sensor data within a time window) and contains key information describing the pipeline operating status of the data item.
[0141] Assume that the sample of the external transmission pipeline sensor data set contains a series of pressure sensor readings with a time resolution of once per hour. In step S10, the computer system first loads this data and performs necessary preprocessing (such as removing outliers, filling missing time points, etc.). Then, the system uses the representation information extraction component to extract the pressure features within each time window (such as 24 hours), including the maximum value, minimum value, average value, standard deviation, and the pressure change rate between adjacent time windows. These features are encoded into a five-dimensional vector, such as [maximum value, minimum value, average value, standard deviation, change rate]. Finally, the sensor data of each time window is represented as such a representation array for subsequent machine learning model training.
[0142] Step S20: using the risk data item detection component to determine risk data item estimation information of the external transmission pipeline sensor data set sample based on the characterization array of the external transmission pipeline sensor data set sample, wherein the risk data item estimation information includes one or more pending risk data items in the external transmission pipeline sensor data set sample.
[0143] The risk data item detection component predicts which data items may represent risks based on the representation array. This component usually contains one or more machine learning models (such as classifiers, regressors, or anomaly detectors) that are trained to learn the mapping relationship between data item features and risk status.
[0144] In step S20, the computer system performs the following operations to determine the risk data item estimation information of the external transmission pipeline sensor data set sample: First, the system passes the representation array of the external transmission pipeline sensor data set sample as input to the risk data item detection component. These representation arrays contain key feature information describing the pipeline operation status. Next, the machine learning model in the risk data item detection component analyzes and predicts the input representation array. This process may involve complex calculations, such as feature transformation, nonlinear mapping, decision boundary judgment, etc. The model will output one or more risk prediction results for each representation array based on the learned mapping relationship. Based on the prediction results of the model, the system identifies potential risk data items. These pending risk data items are the focus of subsequent risk assessment and response. The identification process may involve setting a threshold (such as a predicted probability or score), and treating data items above the threshold as risk data items. Finally, the system outputs the identified pending risk data items as risk data item estimation information. This information usually includes the location of the data item (such as a timestamp, sensor number, etc.), the predicted risk type (such as leakage, blockage, abnormal fluctuation, etc.), and the corresponding predicted probability or score.
[0145] Assume that the sample of the sensor data set of the external transmission pipeline contains a series of pressure sensor readings, which are processed into representation arrays, each of which contains the maximum value, minimum value, average value, standard deviation of the pressure, and pressure change rate of adjacent time windows. In step S20, the computer system inputs these representation arrays into the risk data item detection component. The machine learning model (such as the support vector machine classifier) in the component analyzes these arrays and predicts whether the data item corresponding to each array has risks.
[0146] Assume that the prediction result of the model is the probability that each data item belongs to the "risk" category. The system sets a threshold (such as 0.5) and regards data items with a predicted probability higher than the threshold as pending risk data items. For example, the "risk" probability of a pressure data item in a certain time window after the model prediction is 0.6, which is higher than the threshold, and is therefore identified as a pending risk data item.
[0147] Finally, the computer system outputs the estimated information of these pending risk data items, including their timestamps, sensor numbers, predicted risk types (such as "abnormal pressure increase"), and predicted probabilities (such as 0.6). This information will be used in subsequent risk assessments and response decisions.
[0148] Step S30: determining an abnormality characterization coefficient of each of the pending risk data items based on the abnormality index determination component, wherein the abnormality characterization coefficient is used to characterize the abnormality degree of the pending risk data item in the external transmission pipeline sensor data set sample.
[0149] The abnormal indicator determination component is responsible for calculating the abnormal characterization coefficient of the pending risk data item. This component is usually based on a complex anomaly detection algorithm or model, which can comprehensively consider multiple factors such as the degree of deviation of the data item in the global or local data set and the similarity with the known risk pattern, so as to accurately quantify its abnormality.
[0150] In step S30, the computer system performs the following operations to determine the abnormal characterization coefficient of each pending risk data item: First, the system passes the pending risk data items identified in step S20 as input to the abnormal indicator determination component. These data items have passed the preliminary screening of the risk data item detection component and are considered to be potential risk sources. Next, the algorithm or model in the abnormal indicator determination component analyzes the input pending risk data items and calculates their abnormal characterization coefficients. This process may involve a variety of calculation methods and evaluation indicators, such as distance metrics (such as Euclidean distance, Manhattan distance), density estimation (such as local outlier factor LOF), reconstruction error (such as autoencoder), etc. Taking the local outlier factor (LOF) as an example, the algorithm evaluates the degree of abnormality by comparing the local density difference between the data item and its neighboring points. For each pending risk data item, the algorithm first calculates the distance of its k nearest neighbors and estimates its local density based on this. Then, the algorithm compares the local density of the data item with the local density of its k nearest neighbors and calculates the abnormal score as the abnormal characterization coefficient. The higher the abnormal score, the more abnormal the data item. Finally, the system outputs the calculated anomaly characterization coefficients and associates them with the corresponding pending risk data items. These anomaly characterization coefficients will be used in the subsequent risk assessment and ranking process to help the computer system identify the most critical risk data items.
[0151] Assume that in an oil and gas export pipeline system, after screening in step S20, the computer system identifies a series of pending risk data items, which involve different sensors and time points. In step S30, the system uses the local anomaly factor (LOF) algorithm in the anomaly indicator determination component to calculate the anomaly characterization coefficients of these data items. For a pending risk data item (such as the pressure data of a certain section of pipeline at a specific time point), the algorithm first determines its k nearest neighbors (such as k=5) and calculates the distances of these neighbors. Then, the algorithm estimates the local density of the data item and its neighbors and compares their differences. If the local density of the data item is significantly lower than that of its neighbors, the algorithm calculates a higher anomaly score as the anomaly characterization coefficient, indicating that the data item appears abnormal in the data set. Finally, the system outputs each pending risk data item and its corresponding anomaly characterization coefficient. For example, if the anomaly characterization coefficient of a data item is 0.95 (assuming that a higher anomaly score indicates more abnormality), it means that the data item has a high degree of abnormality in the data set and is likely to be a real risk event.
[0152] Step S40: Based on the abnormal characterization coefficient of each of the pending risk data items, determine the key risk data item estimation information of the external transmission pipeline sensor data set sample, and the key risk data item estimation information includes: one or more key risk data items determined from the one or more pending risk data items based on the abnormal characterization coefficient.
[0153] In step S40, the computer system performs the following operations to determine the estimated information of key risk data items: First, the system takes the abnormal characterization coefficients of each pending risk data item calculated in step S30 as input. These abnormal characterization coefficients quantify the degree of abnormality of the data item in the data set and are an important basis for screening key risk data items. Next, the system sorts all pending risk data items according to the size of the abnormal characterization coefficients. The sorting process may involve simple numerical comparisons or more complex sorting algorithms, such as quick sorting, merge sorting, etc. After the sorting is completed, the system sets a screening threshold or selects the top N data items as key risk data items based on actual needs (such as resource constraints, risk tolerance, etc.). Based on the sorting and screening results, the system determines the estimated information of key risk data items. This information usually includes the specific location of the data item (such as timestamp, sensor number, etc.), the abnormal characterization coefficient, and the possible risk type or description. These key risk data items will serve as the focus of subsequent risk assessment and response.
[0154] Assuming that in the oil and gas transmission pipeline system, after the calculation in step S30, the computer system obtains a series of undetermined risk data items and their corresponding abnormal characterization coefficients. These abnormal characterization coefficients are distributed in a wide range, ranging from low abnormal values close to 0 to high abnormal values close to 1.
[0155] In step S40, the system first sorts these abnormal characterization coefficients to obtain a sequence in increasing order of abnormality. Then, the system sets a screening threshold (such as 0.8) according to actual needs, and regards data items with abnormal characterization coefficients higher than the threshold as key risk data items. Assuming that there are 5 data items in the sorted sequence with abnormal characterization coefficients higher than 0.8, these 5 data items are determined to be key risk data items.
[0156] These key risk data items may represent the most urgent and serious risk sources in the pipeline, such as serious leakage, sudden failure, etc. The system outputs the location of these data items, abnormal characterization coefficients, and possible risk types (such as "abnormal pressure increase, suspected leakage") as key risk data item estimation information. This information will be directly used in subsequent operation and maintenance decisions to guide relevant personnel to quickly respond to and handle risk events.
[0157] Step S40 determines the most critical risk data item estimation information from the pending risk data items through a sorting and screening process based on the abnormal characterization coefficient. This process provides important support for quickly locating high-risk areas, optimizing resource allocation, and formulating effective response measures.
[0158] Step S50: Based on the risk data item estimation information, key risk data item estimation information, risk data item priori labels and key risk data item priori labels of the external transmission pipeline sensor data set samples, the risk prediction neural network is tuned.
[0159] For example, in oil and gas transmission pipeline systems, in order to ensure the safe operation of pipelines, it is necessary to monitor and predict potential risks in real time. Risk prediction neural networks can automatically identify and quantify risk data items by analyzing sensor data set samples. However, due to the complexity and diversity of sensor data, the initial prediction results of the neural network may have errors. Therefore, it is necessary to use known risk data item prior labels and key risk data item prior labels to tune the neural network to improve its prediction accuracy.
[0160] In step S50, the computer system performs the following operations to tune the risk prediction neural network: First, the system collects risk data item estimation information, key risk data item estimation information, and their corresponding prior labels of the external transmission pipeline sensor data set samples. Prior labels are known, verified real risk data items and key risk data item labels, which are used to evaluate the prediction performance of the neural network. Next, the system calculates the loss function value between the neural network prediction results and the prior labels. The loss function is a function that measures the prediction error, which quantifies the performance of the neural network based on the difference between the prediction results and the true labels. In risk prediction scenarios, commonly used loss functions include cross entropy loss, mean square error loss, etc.
[0161] Taking cross entropy loss as an example, for a binary classification problem (such as risk data items and non-risk data items), the loss function can be expressed as:
[0162] ;
[0163] Where N is the number of samples, y i is the true label of the ith sample (0 or 1), is the predicted probability of the neural network.
[0164] Based on the loss function value, the system uses the back-propagation algorithm to calculate the gradient of the parameters of each layer in the neural network, and updates these parameters according to the gradient descent method or other optimization algorithms. This process aims to minimize the loss function value, thereby improving the prediction accuracy of the neural network. The system repeats the above steps (data preparation, loss function calculation, back-propagation and parameter update) until the preset number of iterations is reached or other stopping conditions are met (such as the loss function value no longer decreases significantly). Through multiple iterative training, the neural network gradually learns the complex patterns in the data and improves its prediction performance. During or after training, the system can use an independent validation set or test set to evaluate the performance of the neural network. By comparing the prediction results of the neural network with the prior markings on the validation set or test set, indicators such as accuracy, recall, F1 score, etc. can be calculated to comprehensively evaluate the performance of the neural network.
[0165] Assume that a batch of sensor data set samples with prior labels have been collected in the oil and gas transmission pipeline system. These samples include normal data items, risk data items, and key risk data items, and are accompanied by corresponding true labels. In step S50, the computer system uses these samples to tune the risk prediction neural network. First, the system calculates the prediction results of the neural network for each sample in the training set, compares them with the true labels, and calculates the loss function value. Then, the system uses the back propagation algorithm and the gradient descent method to update the parameters of the neural network to minimize the loss function value. After multiple iterations of training, the prediction performance of the neural network gradually improves. Finally, the system uses an independent validation set to evaluate the performance of the neural network. By calculating indicators such as accuracy and recall, the system finds that the neural network performs well in identifying risk data items and key risk data items, achieving the expected training effect.
[0166] As an implementation manner, the step S50, based on the risk data item estimation information, key risk data item estimation information, risk data item priori labels and key risk data item priori labels of the external transmission pipeline sensor data set sample, optimizes the risk prediction neural network, which may specifically include:
[0167] Step S51: determining a first tuning error based on the risk data item estimation information and the risk data item priori mark, wherein the first tuning error is used to evaluate the accuracy of the risk data item detection component;
[0168] Step S52: determining a second tuning error based on the key risk data item estimation information and the key risk data item priori label, wherein the second tuning error is used to evaluate the accuracy of the abnormal index determination component;
[0169] Step S53: Based on the first tuning error and the second tuning error, the risk prediction neural network is tuned.
[0170] In step S51, the computer system first obtains risk data item prediction information and risk data item prior labels. The prediction information is a list of risk data items predicted by the risk data item detection component based on the representation array of the sensor data set samples, and the prior labels are the real risk status labels of these sample data items.
[0171] Next, the system determines the first tuning error by comparing the estimated information with the prior label. This comparison process may involve a variety of evaluation indicators, such as accuracy, recall, F1 score, etc. However, for the sake of simplicity, it can be assumed that the system uses an error calculation method based on classification errors.
[0172] Assuming that the prediction result of the risk data item detection component is a binary classification problem (i.e., the data item is a risk item or a non-risk item), the following formula can be used to calculate the first tuning error: First tuning error = number of incorrectly predicted risk data items / total number of risk data items.
[0173] Specifically, the system traverses all risk data item samples with a priori labels. For each sample, if the estimated information is inconsistent with the a priori label (i.e., it is estimated as a risk item but a priori labeled as a non-risk item, or it is estimated as a non-risk item but a priori labeled as a risk item), it is considered an incorrect prediction. Finally, the system divides the number of incorrect predictions by the total number of risk data items to obtain the first tuning error. After identifying potential risk data items, the risk prediction neural network needs to further evaluate the degree of abnormality of these data items, that is, calculate their abnormal characterization coefficients. This process is completed by the abnormal indicator determination component. However, due to the limitations of the algorithm itself or the influence of data noise, the calculated abnormal characterization coefficient may have certain errors. In order to evaluate the accuracy of the abnormal indicator determination component and tune the neural network accordingly, the computer system needs to calculate the second tuning error.
[0174] In step S52, the computer system first obtains the estimated information of key risk data items and the priori marks of key risk data items. The estimated information is a list of abnormal characterization coefficients calculated by the abnormal indicator determination component based on the estimated information of risk data items, and the key risk data items are screened out according to these coefficients; and the priori marks are the real key risk status labels of these sample data items.
[0175] Next, the system determines the second tuning error by comparing the estimated information with the prior mark. Since the anomaly characterization coefficient is a continuous value rather than a simple classification label, the system needs to adopt a method that can quantify the continuous value prediction error. Commonly used methods include mean square error (MSE), mean absolute error (MAE), etc.
[0176] Assuming that the system uses mean square error as the evaluation indicator, the calculation formula for the second tuning error is as follows:
[0177] ;
[0178] Where N is the number of samples of key risk data items, the estimated anomaly representation coefficient i is the anomaly representation coefficient of the i-th sample calculated by the anomaly indicator determination component, and the prior anomaly representation coefficient i is the true anomaly representation coefficient of the sample (usually needs to be obtained indirectly in some way, such as based on the evaluation of domain experts).
[0179] However, in practical applications, the true anomaly characterization coefficient is often difficult to obtain directly. Therefore, a feasible alternative is to use the prior labeling of key risk data items (i.e., whether they are key risk items) to indirectly evaluate the accuracy of the anomaly characterization coefficient. Specifically, the system can regard data items with estimated anomaly characterization coefficients higher than a certain threshold as key risk items, and compare them with the prior labeling to calculate the second tuning error similar to the classification error.
[0180] After determining the first tuning error and the second tuning error, the computer system needs to tune the risk prediction neural network based on these error information. This process aims to reduce the error and improve the prediction accuracy by adjusting the parameters of the neural network, so that the network can better meet the actual needs of risk prediction of external transmission pipelines.
[0181] In step S53, the computer system first combines the first tuning error and the second tuning error to form an overall tuning target. This target may be to minimize the weighted sum of the two errors, maximize the accuracy of a specific indicator, etc. The choice of a specific target depends on the specific needs of the application scenario and the current performance of the neural network. Next, the system uses the back propagation algorithm and an optimization algorithm (such as gradient descent, Adam optimizer, etc.) to adjust the parameters of the neural network. The back propagation algorithm is responsible for calculating the gradient of the loss function with respect to the network parameters, while the optimization algorithm updates the parameter values according to these gradients. Through multiple iterative training, the parameters of the neural network gradually converge to the vicinity of the optimal solution, thereby minimizing the error and improving the prediction accuracy.
[0182] Assume that the system uses the gradient descent method as the optimization algorithm and sets a fixed learning rate. In each iteration, the system first calculates the estimated information of the risk data items and the estimated information of the key risk data items based on the current network parameters and input data (i.e., the representation array of the sensor data set samples). Then, the system calculates the first tuning error and the second tuning error based on these estimated information and the corresponding prior labels. Next, the system uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the network parameters, and updates the parameter values according to the gradient descent method. Finally, the system checks whether the stopping conditions are met (such as reaching the preset number of iterations, the loss function value no longer decreases significantly, etc.), and if not, continues to the next iteration training.
[0183] During the tuning process, the system may also need to adopt some strategies to accelerate convergence and prevent overfitting. For example, momentum terms can be used to accelerate the speed of gradient descent; early stopping methods can be used to stop training in advance when the performance of the validation set begins to decline; regularization terms (such as L1 regularization, L2 regularization) can be used to limit the complexity of network parameters, etc.
[0184] After tuning is completed, the system needs to use an independent test set to evaluate the performance of the risk prediction neural network. The test set should contain sensor data set samples that are not involved in the training process in order to objectively reflect the generalization ability of the network. Evaluation indicators can include accuracy, recall, F1 score, area under the ROC curve (AUC), etc. By comparing the changes in performance indicators before and after tuning, the system can verify the effectiveness and necessity of the tuning process.
[0185] As an implementation mode, the risk data item estimation information includes the estimated confidences respectively corresponding to the respective data items in the external transmission pipeline sensor data set sample, and the estimated confidences include the estimated confidences of the data items for multiple risk status types; the risk data item prior tags include the prior confidences respectively corresponding to the respective data items in the external transmission pipeline sensor data set sample, and the prior confidences include the prior confidences of the data items for multiple risk status types. Based on this, the step S51, determining the first tuning error based on the risk data item estimation information and the risk data item prior tags, may specifically include: determining the first tuning error based on the estimated confidences respectively corresponding to the respective data items in the external transmission pipeline sensor data set sample, and the prior confidences respectively corresponding to the respective data items in the external transmission pipeline sensor data set sample.
[0186] For each data item in the sample of the external transmission pipeline sensor data set, the risk data item detection component will output an estimated confidence vector, such as [P_leak, P_clog, P_corrosion], where P_leak, P_clog, and P_corrosion represent the estimated confidence of the data item corresponding to the leakage, blockage, and corrosion risks, respectively. These estimated confidences are calculated by the neural network based on the feature vectors of the data items, reflecting the neural network's predicted probability of the risk state. Corresponding to the estimated confidence vector, each data item also has a prior confidence vector, such as [T_leak, T_clog, T_corrosion], where T_leak, T_clog, and T_corrosion represent the prior confidence of the data item corresponding to the leakage, blockage, and corrosion risks, respectively. These prior confidences are the confidence of the real risk state obtained based on known information (such as historical records, expert evaluations, etc.), and are used as a benchmark for evaluating the accuracy of neural network predictions.
[0187] In step S51, the computer system calculates the first tuning error by comparing the estimated confidence vector and the prior confidence vector of each data item. Since both the estimated confidence and the prior confidence are multidimensional vectors, it is necessary to adopt an error calculation method that can handle multidimensional outputs. Commonly used methods include cross entropy loss (for classification problems) or mean square error loss (for regression problems). However, here, since both the estimated confidence and the prior confidence represent probability distributions, cross entropy loss is more appropriate.
[0188] Assume that the sample of the external transmission pipeline sensor data set contains N data items, each of which corresponds to M risk status types. For the i-th data item, its estimated confidence vector is =[ , ,..., ]; the prior confidence vector is =[ , ,..., ]. Then the first tuning error (cross entropy loss) can be calculated as:
[0189] ;
[0190] Among them, t ij is the prior confidence of the i-th data item corresponding to the j-th risk status type, is the estimated confidence of the neural network output. The cross entropy loss measures the difference between the estimated probability distribution and the true probability distribution. The smaller the loss value, the more accurate the prediction.
[0191] Assume that there are two data items in the sample of the external transmission pipeline sensor data set, and each data item corresponds to three risk status types (leakage, blockage, corrosion). The prior confidence vector of the first data item is [0.9, 0.1, 0.0] (indicating high leakage risk, low blockage risk, and no corrosion risk), and the estimated confidence vector output by the neural network is [0.85, 0.12, 0.03]. The prior confidence vector of the second data item is [0.0, 0.0, 1.0] (indicating no leakage risk, no blockage risk, and high corrosion risk), and the estimated confidence vector output by the neural network is [0.02, 0.01, 0.97]. According to the calculation formula of the cross entropy loss, the system can calculate the loss item corresponding to the first data item (that is, the cross entropy of a single data item), and then sum and average the loss items of all data items to obtain the first tuning error. This error value will be used in the subsequent tuning process of the risk prediction neural network.
[0192] Through the above implementation, the computer system can accurately evaluate the prediction accuracy of the risk data item detection component and tune the neural network according to the evaluation results, thereby improving the accuracy and reliability of the entire risk prediction system.
[0193] Optionally, the prior mark of the key risk data item includes one or more prior key risk data items in the sample of the external transmission pipeline sensor data set. Then, the step S52 determines the second tuning error based on the key risk data item estimation information and the prior mark of the key risk data item, including: determining the second tuning error based on the abnormal characterization coefficients corresponding to each of the pending risk data items in the sample of the external transmission pipeline sensor data set, and the prior key risk data items in each of the pending risk data items.
[0194] In the implementation of step S52, the computer system determines the second tuning error by comparing the estimated information of the key risk data item with the prior mark of the key risk data item. This process is crucial for evaluating the accuracy of the abnormal indicator determination component in the risk prediction neural network. Since the abnormal characterization coefficient is a continuous value, and the prior mark is usually a discrete key risk data item identifier, the calculation of the second tuning error needs to consider how to effectively combine the two. For example, in the oil and gas transmission pipeline system, after the sensor data set samples are processed by the risk prediction neural network, a series of pending risk data items and their corresponding abnormal characterization coefficients will be generated. These abnormal characterization coefficients quantify the degree of abnormality of the data item in the data set and are an important basis for identifying key risk data items. At the same time, in order to evaluate the performance of the abnormal indicator determination component, the system also holds the prior marks of the key risk data items, which indicate which data items actually represent high-risk states.
[0195] For each pending risk data item in the sample of the external transmission pipeline sensor data set, the anomaly indicator determination component calculates an anomaly characterization coefficient, which is a continuous value that indicates the degree of anomaly of the data item in the data set. For example, a data item may have an anomaly characterization coefficient of 0.95, indicating that its degree of anomaly is very high. A priori markings are a list of key risk data items determined by the system based on historical data, expert assessments, or other reliable sources. These markings directly indicate which data items are actually high-risk, but do not provide specific numerical values for the degree of anomaly. For example, a priori markings may simply list the index or identifier of the data item without giving the corresponding anomaly characterization coefficient.
[0196] In step S52, the computer system combines the continuous value abnormal characterization coefficient with the discrete key risk data item prior mark to calculate the second tuning error. A common method is to compare the abnormal characterization coefficient with a threshold, regard the data items above the threshold as predicted key risk data items, and then match them with the prior mark.
[0197] First, the system sets a threshold value of the abnormal characterization coefficient (such as 0.8) to classify the pending risk data items into predicted key risk data items and non-key risk data items. For each pending risk data item, if its abnormal characterization coefficient is higher than the threshold, it is regarded as a predicted key risk data item. The predicted key risk data items are matched with the key risk data items in the priori mark. The matching process may involve comparison of indexes, identifiers or other unique identifiers.
[0198] A simple method is to calculate the proportion of misclassification, that is, the proportion of data items predicted as key risk data items but actually are not (false positives) or not predicted as key risk data items but actually are (false negatives) to the total number of key risk data items. However, this method may not be accurate enough because the prior labels may not contain all the actual key risk data items (i.e., there are cases of missing labels).
[0199] A more precise approach is to use ranking metrics such as average precision (AP) or the area under the receiver operating characteristic curve (AUC). For AP, the system can sort all pending risk data items according to the anomaly characterization coefficient and calculate the average precision at different recall levels. For AUC, the system can calculate the true positive rate (TPR) and false positive rate (FPR) at different thresholds, draw the ROC curve, and then calculate the area under the curve.
[0200] For example, suppose there are 100 undetermined risk data items in the sample of the external transmission pipeline sensor data set, and the threshold value of the abnormal characterization coefficient set by the system is 0.8. After calculation by the abnormal indicator determination component, 20 data items have abnormal characterization coefficients higher than the threshold and are regarded as predicted key risk data items. 15 actual key risk data item indexes are listed in the priori mark. Through matching, it is found that 12 of the predicted key risk data items match the priori mark, and 3 are false positive examples; 3 of the priori marks are not predicted as key risk data items and are false negative examples.
[0201] As an implementation manner, the risk prediction neural network further includes a pipeline operation status type identification component, so the method provided by the embodiment of the present invention may further include:
[0202] Step S1: using the pipeline operation status type identification component, based on the characterization array of the external transmission pipeline sensor data set sample, determine the pipeline operation status type estimation information of the external transmission pipeline sensor data set sample, the pipeline operation status type estimation information of the external transmission pipeline sensor data set sample includes the estimation confidence corresponding to K pipeline operation status types respectively; wherein the estimation confidence corresponding to the x-th pipeline operation status type among the K pipeline operation status types is used to characterize the estimation confidence that the external transmission pipeline sensor data set sample belongs to the x-th pipeline operation status type, 1≤x≤K;
[0203] Step S2: Based on the pipeline operation status type estimation information of the external transmission pipeline sensor data set sample, determine the pipeline operation status type feature of the external transmission pipeline sensor data set sample, and the pipeline operation status type feature of the external transmission pipeline sensor data set sample is used to characterize the pipeline operation status corresponding to the external transmission pipeline sensor data set sample; wherein the risk data item detection component is used to determine the risk data item estimation information of the external transmission pipeline sensor data set sample based on the characterization array of the external transmission pipeline sensor data set sample and the pipeline operation status type feature of the external transmission pipeline sensor data set sample;
[0204] Step S3: Based on the pipeline operation state type estimation information and the pipeline operation state type priori mark, determine a third tuning error, and the third tuning error is used to evaluate the accuracy of the pipeline operation state type identification component.
[0205] Based on this, step S53, tuning the risk prediction neural network based on the first tuning error and the second tuning error, may include: tuning the risk prediction neural network based on the first tuning error, the second tuning error and the third tuning error.
[0206] In step S1, the computer system uses the pipeline operation status type recognition component to determine the pipeline operation status type estimation information based on the characterization array of the external pipeline sensor data set samples. This information includes the estimation confidence corresponding to K predefined pipeline operation status types.
[0207] The pipeline operation status type recognition component can be a multi-classifier, such as a support vector machine (SVM), random forest, gradient boosted tree (GBDT), or deep neural network (DNN). These models are trained to extract features from the representation array and learn the mapping relationship between data items and pipeline operation status types.
[0208] The characterization array as input data contains key features that describe the pipeline operation status, such as statistics or derived features of sensor data such as pressure, temperature, flow, vibration, etc. For each pipeline operation status type, the component will output an estimated confidence, indicating the probability that the data set sample belongs to this state type. For example, for a data set sample, the component may output the following estimated confidence vector: [0.1, 0.2, 0.6, 0.05, 0.05], corresponding to the five state types of normal operation, slight leakage, severe leakage, blockage, and corrosion. Among them, 0.6 means that the sample is most likely to be in a severe leakage state.
[0209] Assume that the sample of the external transmission pipeline sensor data set contains a series of pressure sensor readings, which are formed into a representation array after feature extraction. The pipeline operation status type recognition component (taking random forest as an example) receives these representation arrays as input and outputs the estimated confidence of each state type through a voting mechanism of multiple trees. For example, for the sensor data within a certain time window, the component may output an estimated confidence vector of [0.01, 0.1, 0.85, 0.03, 0.01], indicating that the pipeline within the time window is most likely in a serious leakage state.
[0210] After identifying the operating status type of the pipeline, how to use this information to enhance the detection capability of risk data items becomes a key issue.
[0211] In step S2, the computer system further determines the pipeline operation state type features of the external transmission pipeline sensor data set sample based on the pipeline operation state type estimation information. These features not only reflect the current operation state of the pipeline, but also provide important contextual information for the risk data item detection component. The pipeline operation state type features include, for example, the state type identifier with the highest estimated confidence, a combination of multiple high-confidence state types, or a comprehensive feature vector obtained by weighted summation based on the estimated confidence. For example, if the state type with the highest estimated confidence is a serious leak, then the identifier of the state type (such as encoded as "3") can be used as one of the features; at the same time, the estimated confidence of the first few high-confidence state types can also be considered as an element of the feature vector.
[0212] After obtaining the pipeline operation status type features, the risk data item detection component can combine these features and the characterization array to more accurately identify risk data items. For example, if the pipeline is in a serious leakage state, data items related to abnormal parameters such as pressure and flow may be more likely to be considered risk data items.
[0213] Continuing with the example of severe leakage, assume that the estimated confidence vector output by the pipeline operation status type identification component is [0.01, 0.1, 0.85, 0.03, 0.01]. In step S2, the system can select the state type identifier with the highest estimated confidence (i.e., "severe leakage") as one of the pipeline operation status type features; at the same time, the estimated confidence of the first two high-confidence state types (severe leakage and minor leakage) can also be considered as the first two elements of the feature vector, forming a feature vector such as [0.85, 0.1]. These features will be passed to the risk data item detection component to assist in the identification of risk data items.
[0214] In order to evaluate the performance of the pipeline operation status type recognition component and optimize it, it is necessary to calculate the prediction error of the component. This error not only reflects the component's ability to recognize the pipeline operation status type, but also indirectly affects the accuracy of risk data item detection.
[0215] In step S3, the computer system calculates a third tuning error based on the pipeline operation state type estimation information and the pipeline operation state type prior label (ie, the true label). This error is used to evaluate the accuracy of the pipeline operation state type identification component and guide the subsequent tuning process.
[0216] For each sample of the external transmission pipeline sensor data set, the system holds a priori label of the corresponding pipeline operation status type. These labels are based on the actual status information obtained from historical data, expert evaluation, or other reliable sources. The system calculates the tuning error by comparing the estimated information and the prior label. Common error calculation methods include cross entropy loss (for classification problems), mean square error loss (for regression problems), or other custom loss functions. In the scenario of pipeline operation status type identification, since the output is an estimated confidence vector of multiple status types, cross entropy loss is a suitable choice.
[0217] Assume that the sample of the external transmission pipeline sensor data set contains N data items, each of which corresponds to K pipeline operation status types. For the i-th data item, its estimated confidence vector is =[ , ,..., ], the prior label (one-hot encoding of the true state type) is =[ , ,..., ]. Then the third tuning error (cross entropy loss) can be calculated as:
[0218] ;
[0219] in, is the a priori mark (0 or 1) of the jth pipeline operation status type corresponding to the i-th data item, is the estimated confidence of the component output. The cross entropy loss measures the difference between the estimated probability distribution and the true probability distribution. The smaller the loss value, the more accurate the prediction.
[0220] Assume that there are 100 samples of the external transmission pipeline sensor data set, each of which corresponds to 5 pipeline operation status types (normal operation, slight leakage, severe leakage, blockage, corrosion). For a certain sample, its prior label is [0, 0, 1, 0, 0] (indicating that the true state is severe leakage), and the estimated confidence vector output by the pipeline operation status type identification component is [0.02, 0.08, 0.8, 0.05, 0.05]. According to the calculation formula of the cross entropy loss, the loss term corresponding to the sample can be calculated, and the loss terms of all samples are summed and averaged to obtain the third tuning error. This error value will be used to guide the subsequent tuning process of the pipeline operation status type identification component.
[0221] After determining the first tuning error (evaluating the accuracy of the risk data item detection component), the second tuning error (evaluating the accuracy of the abnormal indicator determination component), and the third tuning error (evaluating the accuracy of the pipeline operation status type identification component), the computer system needs to comprehensively consider these errors to perform overall tuning of the risk prediction neural network.
[0222] In step S53, the computer system determines a comprehensive tuning target based on the weighted sum of the first tuning error, the second tuning error and the third tuning error or other combinations thereof. Then, the back propagation algorithm and an optimization algorithm (such as gradient descent method, Adam optimizer, etc.) are used to adjust the parameters of the neural network to minimize the comprehensive tuning error and improve the prediction accuracy of the network.
[0223] The comprehensive tuning goal may be the weighted sum of multiple tuning errors, maximizing the accuracy of a specific indicator (such as F1 score, AUC, etc.), or other customized optimization goals. The choice of weights depends on the importance of each component to the final risk prediction result and the specific needs of the system. In each iteration, the system first calculates the output results and corresponding tuning errors of each component based on the current network parameters and input data (i.e., the representation array of sensor data set samples). Then, the back propagation algorithm is used to calculate the gradient of the loss function with respect to the network parameters, and the parameter values are updated according to the optimization algorithm. Through multiple iterative training, the parameters of the neural network gradually converge to the vicinity of the optimal solution, thereby minimizing the error and improving the prediction accuracy. During the tuning process, the system also needs to monitor the changes in various tuning errors and the convergence status of the network. If it is found that a certain tuning error continues to increase or the network cannot converge, it may be necessary to adjust the parameters of the optimization algorithm, add regularization terms, or take other measures to prevent overfitting or accelerate convergence.
[0224] As an implementation mode, the pipeline operation state type prior mark includes the prior confidences corresponding to the K pipeline operation state types respectively, and the prior confidence corresponding to the x-th pipeline operation state type is used to characterize the prior confidence that the sample of the external transmission pipeline sensor data set belongs to the x-th pipeline operation state type. Based on this, step S3, based on the pipeline operation state type estimation information and the pipeline operation state type prior mark, determines the third tuning error, which may include: determining the third tuning error based on the estimated confidences corresponding to the K pipeline operation state types respectively, and the prior confidences corresponding to the K pipeline operation state types respectively.
[0225] In the implementation of step S3, the computer system determines the third tuning error by comparing the pipeline operation state type estimation information with the pipeline operation state type prior mark, which is crucial for evaluating the accuracy of the pipeline operation state type identification component. Since both the estimation information and the prior mark are about the confidence distribution of K pipeline operation state types, the calculation of the third tuning error needs to be able to handle the matching problem of this multi-dimensional output. In the oil and gas export pipeline system, in order to accurately predict potential risks, it is first necessary to understand the current operation state of the pipeline. The pipeline operation state type identification component outputs the estimated confidence that each sample belongs to K predefined state types based on the characterization array of the sensor data set samples. At the same time, the system also holds the pipeline operation state type prior mark corresponding to each sample, which is based on the real state confidence distribution obtained from historical data, expert evaluation or other reliable sources. By comparing the estimated confidence with the prior confidence, the performance of the identification component can be evaluated and tuned accordingly.
[0226] In this embodiment, the pipeline operation status type prior label is no longer a simple status type identification (such as coding), but includes the prior confidence corresponding to K status types. These prior confidences represent the system's belief in the true state of the sample, which is usually based on a large amount of historical data and domain knowledge. For example, for a certain sensor data set sample, its prior confidence distribution may be [0.05, 0.2, 0.65, 0.05, 0.05], corresponding to the five status types of normal operation, slight leakage, severe leakage, blockage, and corrosion. This indicates that the system believes that there is a 65% chance that the sample is in a severe leakage state.
[0227] In step S3, the computer system needs to find a method to compare the estimated confidence of the K pipeline operation status types with the prior confidence to calculate the third tuning error. Since both are probability distributions, cross-entropy loss can be used as an error metric. Cross-entropy loss is a common method for measuring the difference between two probability distributions, and is particularly effective in classification problems.
[0228] Assume that the sample of the external transmission pipeline sensor data set contains N data items, each of which corresponds to K pipeline operation status types. For the i-th data item, its estimated confidence vector is =[ , ,..., ], the prior confidence vector is =[ , ,..., ]. Then the third tuning error (cross entropy loss) can be calculated as:
[0229] ;
[0230] Among them, t ij is the prior confidence of the i-th data item corresponding to the j-th pipeline operation status type, is the estimated confidence of the recognition component output. The smaller the cross entropy loss, the closer the estimated confidence distribution is to the prior confidence distribution, that is, the better the performance of the recognition component.
[0231] The embodiment of the present invention provides a skid-mounted system, such as Figure 2 As shown, the redundant structure of the skid-mounted system is omitted. Specifically, the skid-mounted system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the skid-mounted system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the skid-mounted system 100 does not constitute a limitation on the embodiments of the present invention.
[0232] An embodiment of the present invention provides a skid-mounted system, and the skid-mounted system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above method is implemented.
[0233] The above descriptions are only some embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A method for predicting risk of external transmission pipelines based on big data, characterized in that: The method comprises: Acquire a characterization array of an external transmission pipeline sensor data set, wherein the characterization array of the external transmission pipeline sensor data set is used to characterize a pipeline operation state of the external transmission pipeline sensor data set; Determine one or more undetermined risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set; Determine an abnormal characterization coefficient of each of the pending risk data items, wherein the abnormal characterization coefficient is used to characterize the abnormal degree of the pending risk data item in the external transmission pipeline sensor data set; Determining a risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items; Obtain risk estimation information of the external transmission pipeline based on the risk data item detection result; The determining of the abnormal characterization coefficient of each of the pending risk data items includes: Obtaining a data item array representation of each of the pending risk data items; Determine an abnormal characterization coefficient of each of the pending risk data items based on the data item array representation of each of the pending risk data items and the characterization array of the external transmission pipeline sensor data set; The characterization array of the external transmission pipeline sensor data set includes a first characterization array, and the first characterization array is used to characterize the global pipeline operation state of the external transmission pipeline sensor data set; The determining of the abnormal characterization coefficient of each of the pending risk data items based on the data item array representation of each of the pending risk data items and the characterization array of the external transmission pipeline sensor data set comprises: Based on the commonality measurement results between the data item array representation of each of the pending risk data items and the first characterization array, an abnormal characterization coefficient of each of the pending risk data items is determined.
2. The method according to claim 1, characterized in that The step of determining the risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items includes: Determine the undetermined risk data item whose abnormal characterization coefficient meets the first evaluation requirement as the risk data item included in the external transmission pipeline sensor data set; The first assessment requirement includes: The abnormal characterization coefficient is not less than a first reference value; and / or; According to the magnitude of the abnormal characterization coefficients, after being arranged in descending order, the abnormal characterization coefficients in the first s numbers, where s ≥ 1; The step of determining the risk data item detection result of the external transmission pipeline sensor data set based on the abnormal characterization coefficient of each of the pending risk data items includes: Determine the undetermined risk data item whose abnormal characterization coefficient meets the second evaluation requirement as the key risk data item included in the external transmission pipeline sensor data set; The second assessment requirement includes: The abnormal characterization coefficient is not less than a second reference value; and / or; According to the magnitude of the abnormal characterization coefficients, after being arranged in descending order, the abnormal characterization coefficients in the first r order, where r≥1; The step of determining one or more undetermined risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set comprises: Determine, based on the characterization array of the external transmission pipeline sensor data set, a pipeline operation state type feature of the external transmission pipeline sensor data set, wherein the pipeline operation state type feature of the external transmission pipeline sensor data set is used to characterize the pipeline operation state corresponding to the external transmission pipeline sensor data set; Based on the characterization array of the external transmission pipeline sensor data set and the pipeline operation status type characteristics of the external transmission pipeline sensor data set, one or more pending risk data items in the external transmission pipeline sensor data set are determined.
3. The method according to claim 2, characterized in that The determining of the pipeline operation status type feature of the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set includes: Based on the characterization array of the external transmission pipeline sensor data set, the confidences corresponding to K pipeline operation status types are determined; wherein the confidence corresponding to the x-th pipeline operation status type among the K pipeline operation status types is used to characterize the probability that the external transmission pipeline sensor data set belongs to the x-th pipeline operation status type, 1≤x≤K; Based on the confidences respectively corresponding to the K pipeline operation status types, determining u pipeline operation status types with the largest confidences from the K pipeline operation status types, and obtaining the u target pipeline operation status types, where u≥1; Based on the confidences respectively corresponding to the u target pipeline operation status types, the representation arrays respectively corresponding to the u target pipeline operation status types are fused to obtain a fused representation array; wherein the pipeline operation status type feature of the external transmission pipeline sensor data set includes the fused representation array; The representation array of the external transmission pipeline sensor data set includes representation arrays corresponding to a plurality of data items in the external transmission pipeline sensor data set; The step of determining one or more undetermined risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set and the pipeline operation status type feature of the external transmission pipeline sensor data set includes: Combining the representation arrays corresponding to the respective data items in the external transmission pipeline sensor data set with the pipeline operation status type features of the external transmission pipeline sensor data set, respectively, to obtain the combined representation arrays corresponding to the respective data items; Based on the combined characterization array, one or more undetermined risk data items in the external transmission pipeline sensor data set are determined.
4. The method according to any one of claims 1 to 3, characterized in that The risk data item detection result of the external transmission pipeline sensor data set is determined by a risk prediction neural network, and the risk prediction neural network includes a characterization information extraction component, a risk data item detection component and an abnormal index determination component; wherein the characterization information extraction component is used to obtain the characterization array of the external transmission pipeline sensor data set; the risk data item detection component is used to determine one or more undetermined risk data items in the external transmission pipeline sensor data set based on the characterization array of the external transmission pipeline sensor data set; the abnormal index determination component is used to determine the abnormal characterization coefficient of each of the undetermined risk data items; The method also includes a training process of the risk prediction neural network, including: Acquire a characterization array of an external transmission pipeline sensor data set sample based on the characterization information extraction component, wherein the characterization array of the external transmission pipeline sensor data set sample is used to characterize a pipeline operation state of the external transmission pipeline sensor data set sample; Using the risk data item detection component, based on the characterization array of the external transmission pipeline sensor data set sample, determine the risk data item estimation information of the external transmission pipeline sensor data set sample, wherein the risk data item estimation information includes one or more undetermined risk data items in the external transmission pipeline sensor data set sample; Determine an abnormality characterization coefficient of each of the pending risk data items based on the abnormality index determination component, wherein the abnormality characterization coefficient is used to characterize the abnormality degree of the pending risk data item in the external transmission pipeline sensor data set sample; Based on the abnormal characterization coefficient of each of the pending risk data items, determining the key risk data item estimation information of the external transmission pipeline sensor data set sample, the key risk data item estimation information comprising: one or more key risk data items determined from the one or more pending risk data items based on the abnormal characterization coefficient; The risk prediction neural network is tuned based on the risk data item estimation information, key risk data item estimation information, risk data item priori labels and key risk data item priori labels of the external transmission pipeline sensor data set samples.
5. The method according to claim 4, characterized in that The risk prediction neural network is tuned based on the risk data item estimation information, key risk data item estimation information, risk data item priori labels and key risk data item priori labels of the external transmission pipeline sensor data set samples, including: Determining a first tuning error based on the risk data item estimation information and the risk data item priori label, wherein the first tuning error is used to evaluate the accuracy of the risk data item detection component; Determining a second tuning error based on the key risk data item estimation information and the key risk data item priori label, wherein the second tuning error is used to evaluate the accuracy of the abnormal indicator determination component; The risk prediction neural network is tuned based on the first tuning error and the second tuning error.
6. The method according to claim 5, characterized in that The risk data item estimation information includes the estimated confidences corresponding to the respective data items in the external transmission pipeline sensor data set sample, and the estimated confidences include the estimated confidences of the data items for multiple risk status types; the risk data item priori label includes the priori confidences corresponding to the respective data items in the external transmission pipeline sensor data set sample, and the priori confidences include the priori confidences of the data items for multiple risk status types; The determining a first tuning error based on the risk data item estimation information and the risk data item priori label comprises: Determining the first tuning error based on the estimated confidences respectively corresponding to the respective data items in the sample of the external transmission pipeline sensor data set and the priori confidences respectively corresponding to the respective data items in the sample of the external transmission pipeline sensor data set; The key risk data item priori label includes one or more key risk data items priori in the external transmission pipeline sensor data set sample; The determining a second tuning error based on the key risk data item estimation information and the key risk data item priori label comprises: Determine the second tuning error based on the abnormal characterization coefficients respectively corresponding to each of the pending risk data items in the external transmission pipeline sensor data set sample and the priori key risk data items in each of the pending risk data items; The risk prediction neural network also includes a pipeline operation status type identification component, and the method also includes: The pipeline operation state type identification component is used to determine the pipeline operation state type estimation information of the external transmission pipeline sensor data set sample based on the characterization array of the external transmission pipeline sensor data set sample, wherein the pipeline operation state type estimation information of the external transmission pipeline sensor data set sample includes the estimation confidences corresponding to K pipeline operation state types respectively; wherein the estimation confidence corresponding to the x-th pipeline operation state type among the K pipeline operation state types is used to characterize the estimation confidence that the external transmission pipeline sensor data set sample belongs to the x-th pipeline operation state type, 1≤x≤K; Based on the pipeline operation status type estimation information of the external transmission pipeline sensor data set sample, the pipeline operation status type feature of the external transmission pipeline sensor data set sample is determined, and the pipeline operation status type feature of the external transmission pipeline sensor data set sample is used to characterize the pipeline operation status corresponding to the external transmission pipeline sensor data set sample; wherein the risk data item detection component is used to determine the risk data item estimation information of the external transmission pipeline sensor data set sample based on the characterization array of the external transmission pipeline sensor data set sample and the pipeline operation status type feature of the external transmission pipeline sensor data set sample; Determine a third tuning error based on the pipeline operation state type estimation information and the pipeline operation state type priori mark, wherein the third tuning error is used to evaluate the accuracy of the pipeline operation state type identification component; The step of tuning the risk prediction neural network based on the first tuning error and the second tuning error includes: The risk prediction neural network is tuned based on the first tuning error, the second tuning error, and the third tuning error.
7. The method according to claim 6, characterized in that The pipeline operation state type priori mark includes the priori confidences corresponding to the K pipeline operation state types respectively, and the priori confidence corresponding to the x-th pipeline operation state type, which is used to characterize the priori confidence that the external transmission pipeline sensor data set sample belongs to the x-th pipeline operation state type; The determining a third tuning error based on the pipeline operation state type estimation information and the pipeline operation state type priori mark includes: The third tuning error is determined based on the estimated confidences respectively corresponding to the K pipeline operation status types and the priori confidences respectively corresponding to the K pipeline operation status types.
8. A skid-mounted system, characterized in that: include: one or more processors; Memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 7 is implemented.
Citation Information
Patent Citations
Detection method for health situation of underground pipe network
CN106596710A
Gas pipeline dynamic risk early warning system
CN112036086A
Electric power intelligent anti-electric shock safety method and system based on deep learning
CN117611410A
Network security assessment method and system based on network risk detection
CN118316723A
Cable channel monitoring system based on sensor network
CN118536817A
Cited By
Disaster recovery system based on cloud platform
CN120909846A
