Deep flow pattern feature extraction and state judgment method for multi-source heterogeneous data

Abstract

The invention discloses a deep flow pattern feature extraction and state judgment method for multi-source asynchronous heterogeneous data. The method is based on three main steps of data preprocessing, feature extraction and leakage judgment based on deep flow pattern feature extraction. The method comprises the following steps: firstly, preliminarily screening data such as pressure, flow and concentration from a multi-source data sensor by using an absolute difference median method; and then, performing normalization and whitening processing on the gas pipeline monitoring data subjected to preliminary screening so as to eliminate dimensions of different sensors and remove redundant information in the data. Through normalization, the data is converted into a unified scale, and it is ensured that the data of different sensors have comparability. Next, a flow pattern learning method is used for carrying out dimensionality reduction on the preprocessed high-dimensional gas pipeline monitoring sequence, and potential low-dimensional flow pattern features are extracted; and fusing the extracted feature data through the combination of an uncertainty theory and a weighted average method, and judging whether the gas pipeline leaks or not.

Description

Technical Field

[0001] The present invention relates to the fields of communication and machine learning, and more specifically to a method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data. Background Art

[0002] With the rapid development of information technology, we are in a data-driven era. Data fusion, as an important branch of information technology, is becoming increasingly important. In the process of economic structure transformation to digital economy, the deep integration of industry and information technology has become a new driving force for economic development. The vigorous development of cloud computing, big data and mobile Internet has not only changed people's lifestyles, but also promoted the deep integration of data in different fields, bringing revolutionary changes to all walks of life. The concept of "Internet +" has attracted much attention in recent years. It emphasizes the deep integration of the Internet and traditional industries. This integration trend not only promotes industrial upgrading, but also provides a broader stage for information integration. In this context, information integration has become a key factor in promoting economic transformation and upgrading and improving the level of government management. At the national level, the importance of promoting economic transformation and upgrading by improving the overall level of informatization has been recognized. More importantly, relying on comprehensive informatization to support deep-level reforms. This requires us to break down the barriers between industries and regions and build a unified big data platform that can fully integrate multi-source data from multiple fields to achieve centralized management and efficient use of data. In the field of comprehensive urban governance, facing the large amount of heterogeneous data generated in multiple fields such as transportation, environment, and energy, how to efficiently organize and process these scattered data has become a major challenge in information management. The modernization and intelligence of urban governance cannot be separated from the effective integration and analysis of these data. The development of data fusion technology can not only improve the accuracy and availability of data, but also provide decision makers with more comprehensive and in-depth insights, thereby improving the quality and efficiency of decision-making.

[0003] In order to overcome the problems caused by the heterogeneity and asynchrony of multi-source data in gas pipelines, it is indeed necessary to establish a multi-source heterogeneous data fusion structural model and a data quantification fusion processing method. In gas pipeline monitoring, multiple sensors may be involved, such as pressure sensors, flow meters, and concentration environment monitoring instruments, which have different data types and sensing mechanisms. In addition, the data sources of these sensors may come from the same platform or multiple platforms, resulting in the inability to keep synchronization between data sources. In order to effectively deal with this situation, it is necessary to first clean and convert the data to eliminate noise and inconsistency in the data and convert it into a unified data format and unit. Then, through the data quantification fusion processing method, the multi-source heterogeneous asynchronous data is structured and unified data integration is performed. In multi-source data fusion, the information obtained by each sensor can be regarded as a description of specific feature data in the sensor space. Due to the differences in the physical properties and spatial positions of the sensors, these feature data are in different description spaces and are not easy to be directly fused. Therefore, before fusion processing, these feature data need to be mapped to a common reference description space. This process is called feature mapping, which involves the process of converting the features of products, data, or information from one expression space to another. By mapping the feature data into a common reference description space, a consistent description of the feature data can be achieved, thereby achieving effective fusion processing of multi-source asynchronous heterogeneous data. This fusion processing can improve the accuracy and availability of data, thereby achieving comprehensive monitoring and analysis of targets, events or situations in gas pipelines. Data fusion technology combines data from multiple sensors and information from related data to achieve higher accuracy and more specific inferences than can be achieved using a single sensor alone. The purpose of data fusion is to improve the availability, accuracy and completeness of data and provide support for subsequent analysis and decision-making. Through data fusion, we can obtain more comprehensive and accurate data from different perspectives and provide better business applications and services. The process of data fusion usually involves multiple links such as data cleaning, conversion and integration to cope with differences in format, structure, semantics, etc. of data from different sources, and finally synchronize them to the database. Data fusion methods include traditional ETL (extraction, transformation, loading) processes, data integration technology, machine learning and data mining methods. In data fusion, it is very important to maintain data consistency and integrity, which involves ensuring that the fused data set is logically consistent and complete, and avoiding data redundancy or omission.

[0004] Effectively utilizing the fusion of multi-source heterogeneous data in gas pipelines is an important issue that needs to be solved urgently in this field. This is related to the quality and efficiency of pipeline network monitoring work, as well as the strength of public safety protection. In view of this, it is urgent to establish a scientific and reasonable multi-source data fusion model and process, taking into account the differences in the characteristics of each sensor. It is necessary to develop appropriate data normalization processing methods to eliminate the inconsistency of formats and indicators between raw data; at the same time, it is necessary to explore the characterization method of multi-source features under common reference coordinates. Only by establishing a unified processing template can the intrinsic value of the data be fully explored and deep fusion analysis be performed. Therefore, overcoming the fusion of multi-source heterogeneous data in gas pipelines is a problem that needs to be solved urgently. It is necessary to establish a suitable data fusion model and processing method to better utilize multi-source data and improve the efficiency and reliability of gas pipeline monitoring. Summary of the invention

[0005] In view of the above technical problems, the present invention proposes a gas pipeline leakage monitoring method based on deep flow pattern feature extraction, including: using the absolute difference median method to perform preliminary screening on the monitoring data such as gas pipeline pressure, flow and concentration from multiple sensors, specifically calculating the absolute difference between the monitoring data of each sensor, taking the median of these differences as the filtering threshold, and eliminating the data with errors exceeding the threshold; then, converting the data of different sensors into a standardized form with zero mean and unit standard deviation through a normalization method to eliminate the scale difference of the data; further using whitening processing to decorrelate the normalized data to reduce data redundancy and provide high-dimensional monitoring data with good consistency for subsequent feature extraction.

[0006] The present invention provides a method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data, comprising the following steps:

[0007] Step (1): Collect multi-source sensor data: pressure, flow, concentration and other data, use the absolute value difference median method to pre-process the data, and conduct preliminary screening of the data;

[0008] Step (2) The original high-dimensional vector of the gas pipeline leakage monitoring sequence after normalizing and whitening the collected multi-source data is y 1 ,y 2 ,…,y M ,in M is the number of multiple sources, and T is the length of the observed time series;

[0009] Step (3) constructing N an objective function in feature extraction of structured manifold learning for multiple gas pipeline monitoring sequences;

[0010] Step (4) Analyze and construct the completed objective function N, unify the original high-dimensional vectors of the normalized and whitened pressure, flow, concentration and other sequences into the same low-dimensional vector to complete feature extraction. The main process of the analysis is: calculate F k+1 , W k+1 , U k+1 , if it is not converged, the three matrices are calculated again until convergence, and the transformation matrix W is obtained and the original high-dimensional matrix is ​​converted into a low-order matrix F, completing the multivariate data feature extraction of the gas pipeline pressure, flow, and concentration data;

[0011] Step (5) utilizes the data features of the gas pipeline monitoring data extracted above, and fuses the feature data based on uncertainty theory and weighted average method to determine whether the gas pipeline is leaking.

[0012] Technical effects and advantages of the present invention:

[0013] The present invention provides a deep flow pattern feature extraction and state determination method for multi-source asynchronous heterogeneous data for gas pipeline monitoring, aiming to solve the dimensionality difference and asynchronism problems faced by multi-source heterogeneous data in the feature extraction and fusion process. The method includes the following core steps: first, collect monitoring data from various sensors such as pressure sensors, flow meters and concentration monitors, pre-screen the multi-source data by the absolute value difference median method, remove outliers and noise, and preliminarily improve the data quality; then, normalize and whiten the pre-processed data, eliminate the dimensional differences and linear correlations between different sensor data, and uniformly convert them into high-dimensional sequences of the same dimension, providing a basis for subsequent feature extraction. In the feature extraction stage, the present invention uses flow pattern learning technology to reduce the dimension of the normalized and whitened high-dimensional sequence, and map the data to a low-dimensional flow pattern space. Flow pattern learning is a nonlinear dimensionality reduction method that can extract potential low-dimensional features based on the manifold structure relationship between multi-source data, and realize unified dimensionality reduction processing of high-dimensional heterogeneous data by constructing and optimizing the objective function. This method is particularly suitable for unified feature processing of heterogeneous data. It not only effectively retains the global and local information of the original data, but also converts heterogeneous data from different sources into consistent low-dimensional feature representations, providing a unified feature input for subsequent data fusion. In general, due to the different characteristics and dimensions of data collected by different sensors, traditional methods are often difficult to directly fuse data during feature extraction, especially the processing of asynchronous heterogeneous data faces huge challenges. However, the present invention successfully solves this problem based on the intrinsic structural relationship of the data through flow learning technology. By converting multi-source heterogeneous data into feature sequences of the same low dimension, deep fusion between heterogeneous data is achieved, providing an accurate and efficient data foundation for gas pipeline leakage monitoring. In addition, this method not only shows significant advantages in gas pipeline monitoring, but can also be extended to other multi-source heterogeneous data scenarios, such as industrial equipment fault diagnosis, intelligent manufacturing and urban comprehensive management, and has wide applicability and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to better express the specific process of the present invention and the technology used therein, the following briefly introduces the drawings required for the specific implementation of the present invention.

[0015] Figure 1 It is a flow chart of the scheme of the present invention;

[0016] Figure 2 It is the construction and solution process of the objective function in steps (3) and (4);

[0017] Figure 3 This is a schematic diagram of the gas pipeline leakage identified by using the DS theory and the weighted average method in step (5). DETAILED DESCRIPTION

[0018] In order to more clearly illustrate the purpose, technical solutions and advantages of the embodiments of the present invention, the technical solutions of the embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. Please note that the described embodiments are only some embodiments of the present invention, not all embodiments. Those skilled in the art can obtain other embodiments including but not limited to these embodiments based on the embodiments disclosed in the present invention without making creative efforts, and these embodiments all belong to the protection scope of the present invention.

[0019] Figure 1 is an implementation flow chart of a method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to an embodiment of the present invention, Figure 2 It is a flowchart of objective function construction and solution. Figure 3 It is a flow chart for jointly judging gas pipeline leakage by using uncertainty theory and weighted average method.

[0020] like Figure 1 As shown, the present invention provides a deep flow pattern feature extraction and state determination method for multi-source heterogeneous data, which includes five steps.

[0021] Step (1): Collect data from multiple sensors, including pressure sensors, flow sensors, and concentration monitoring sensors, which are used to monitor the pressure changes inside the pipeline, the fluctuations in gas flow, and the diffusion of gas concentration in the surrounding environment in real time. The data source can be fixed sensors installed at key locations on the pipeline to continuously collect data, and mobile monitoring can be performed by inspection robots or handheld devices to cover the monitoring blind spots of fixed sensors. At the same time, the data from multiple sensors can be integrated with the help of a remote monitoring platform and uploaded to the cloud for centralized management, thereby ensuring the comprehensiveness and real-time nature of data collection.

[0022] Step (2) The raw sensor data usually has problems of noise, missing data or time asynchrony, so it needs to be strictly preprocessed to improve the reliability and accuracy of subsequent analysis. First, the absolute value difference median method is used to denoise the multiple collected data of the sensor. By calculating the absolute difference of the sensor data and taking the median as the screening threshold, outliers and noise data are eliminated; secondly, check whether the timestamps of different sensor data are consistent. For asynchronous data, linear interpolation or spline interpolation methods are used for time alignment, and the dimensions of the data are unified; thirdly, in the case of missing data, the data is supplemented by interpolation or mean filling methods to ensure data integrity; finally, the correlation coefficient between sensor data is calculated through correlation analysis, highly correlated samples are screened out, and outliers beyond the reasonable range are filtered. The original high-dimensional vector of the gas pipeline leakage monitoring sequence after processing the collected multi-source data is y 1 ,y 2,…,y M ,in M is the number of multiple sources, and T is the length of the observed time series;

[0023] Step (3) In the feature extraction stage of multiple gas pipeline monitoring sequences, it is first necessary to model the potential manifold structure of the data through structured manifold learning technology, and construct an optimization objective function to achieve efficient dimensionality reduction and feature extraction. Specifically, based on the high-dimensional monitoring data obtained after normalization and whitening, including: multi-source sequences such as pressure, flow, concentration, etc., each time series data is first high-dimensional Hankel matrixed, that is, the time series data is converted into an over-complete delay vector form to retain the historical characteristics of the time series; then the Hankel matrixed data is spliced ​​and reconstructed with other sequences to form a unified reconstructed data set. In order to extract manifold features from high-dimensional reconstructed data, an objective function based on Laplace eigenmapping is constructed to characterize the local geometric structural relationship of data points in the manifold space. Specifically, first construct an adjacency matrix, which uses the similarity between monitoring sequence data points as weights to reflect the local relationship between data points; then, calculate the normalized Laplace matrix, introduce it into the objective function, and set orthogonal constraints to avoid eigenvector solution degradation. The form of the objective function is: minTr(Y T LY),stY T Y=I

[0024] Among them, L is the normalized Laplace matrix, Y is the low-dimensional manifold feature representation, and I is the unit matrix. In order to further enhance the adaptability of the objective function to the gas pipeline monitoring data, a joint sparse penalty term is added to the objective function to constrain the feature consistency of the original data and the manifold space projection, thereby retaining both global and local information. The final optimization objective function is:

[0025]

[0026] Among them, X is the original data matrix, Z is the manifold space projection matrix, P is the transformation matrix, and λ and μ are adjustment parameters.

[0027] In the objective function analysis process, the matrix P and Y are first jointly optimized based on the auxiliary function, and the eigenvalue decomposition of the matrix is ​​calculated respectively, and the eigenvector corresponding to the smallest eigenvalue other than zero in the eigenvalue decomposition is selected as the result; then, through repeated iterative optimization, the gradient update of the Laplace matrix L is calculated until the objective function converges. Finally, the obtained transformation matrix P can uniformly map the original high-dimensional monitoring sequence into a low-dimensional eigenvector, completing the construction of the objective function N in the structured feature extraction of multivariate data such as gas pipeline pressure, flow, and concentration;

[0028] Step (4) Analyze and construct the completed objective function N, unify the original high-dimensional vectors of the normalized and whitened pressure, flow, concentration and other sequences into the same low-dimensional vector to complete feature extraction. The main process of the analysis is: calculate F k+1 , W k+1 , U k+1 , if it is not converged, the three matrices are calculated again until convergence, and the transformation matrix W is obtained and the original high-dimensional matrix is ​​converted into a low-order matrix F, completing the multivariate data feature extraction of the gas pipeline pressure, flow, and concentration data;

[0029] Step (5) After feature extraction is completed, uncertainty analysis and weight assignment are performed on the gas pipeline monitoring data based on the DS evidence theory. First, based on the characteristic data of the pressure, flow and concentration sensors, the respective uncertainty evidence values ​​are calculated, which respectively represent the indicative strength of different sensors for the gas leakage status; for example, the pressure sensor evaluates its evidence value through abnormal pressure fluctuations, the flow sensor calculates through the flow change amplitude, and the concentration sensor determines its evidence value based on the rising trend of the gas concentration in the environment. Subsequently, weights are assigned according to the accuracy and reliability of each sensor, and the specific weight is determined by the following theoretical basis: First, the detection principle of the pressure sensor usually directly reflects the operating status inside the pipeline, and it is most sensitive to gas leakage, so it is assigned a higher weight; second, the flow sensor is less affected by the environment because it mainly detects dynamic changes in gas transportation, but its data fluctuations may be interfered by non-leakage factors, so the weight is slightly lower; finally, the role of the concentration sensor is to monitor the diffusion of environmental gases that may be caused by gas leakage, but because the diffusion is delayed and may be affected by external environmental conditions, its accuracy and real-time performance are relatively low, so the weight is the lowest. Next, the uncertainty evidence of each sensor is fused using the Dempster synthesis rule to calculate the overall uncertainty evidence value to represent the confidence of the gas leak. If the confidence after fusion is higher than the set threshold, it is determined that there is a leak; otherwise, it is determined that there is no leak. In addition, the overall confidence is further evaluated by the weighted average method, and the final result is comprehensively judged in combination with the weight of each sensor to ensure the accuracy and robustness of the leakage status judgment. At the same time, the evidence and weight information are recorded to provide feedback for sensor configuration optimization and data fusion process.

[0030] Step (6) After completing the uncertainty evidence calculation and weight assignment of the gas pipeline monitoring data, the uncertainty evidence obtained by different sensors is fused into an overall uncertainty theory through the Dempster synthesis rule to represent the comprehensive confidence of the gas pipeline leakage. The Dempster synthesis rule is a fusion method based on evidence theory, which can effectively combine multi-source sensor information to reflect the support degree of the overall data for gas leakage. Specifically, the uncertainty evidence value extracted by each sensor (including: Belief p 、Belief of flow sensor f 、Belief of concentration sensor c ) is weighted according to the weights to ensure that sensors with high confidence and high weight have a greater impact on the synthesis results. The synthesis formula is:

[0031] Where m(B) and m(C) represent the uncertainty evidence of different sensors, and the denominator is used to eliminate the influence of conflicting evidence. Finally, the fusion result of the uncertainty evidence of each sensor is calculated. After completing the synthesis of uncertainty evidence, in order to further enhance the comprehensive assessment of confidence, the weighted average method is used to perform weighted synthesis of the overall uncertainty theory. The core of the weighted average method is to adjust and optimize the overall evidence after fusion based on the importance of the sensor weight and its indicative strength of gas leakage under different conditions. The formula is as follows: Confidence = w p ·Belief p +w f ·Belief f +w c ·Belief c .

[0032] Among them, w p 、w f 、w c are the weights of the pressure sensor, flow sensor, and concentration sensor, respectively, and are allocated according to the accuracy and reliability of the sensor. After weighted processing, the calculated Confidence represents the final confidence of the gas leak. Finally, the synthesis result is compared with the preset threshold. If the confidence is higher than the threshold, it is determined that the gas pipeline has leaked; if the confidence is lower than the threshold, it is determined that there is no leak. This step comprehensively considers the uncertainty characteristics and weight distribution of multi-source sensor data. Through the joint application of Dempster synthesis rule and weighted average method, it not only realizes the efficient judgment of gas pipeline leakage, but also ensures the robustness and accuracy of the final evaluation results. At the same time, the evidence contribution value and weight distribution of each sensor in the fusion process are recorded, which provides an effective reference for subsequent sensor configuration optimization and data fusion improvement.

[0033] like Figure 2 As shown in the figure, the pressure, flow, concentration, etc. of the gas pipeline monitoring data are normalized and whitened, and then multiple high-dimensional sequences with the same dimension are reduced to low-dimensional sequences through flow type learning to complete feature extraction. First, the objective function N is constructed: First, each gas pipeline monitoring sequence is high-dimensional Hankelized, and then the overcomplete delay vector and the high-dimensional Hankelized data are spliced ​​together to obtain the reconstructed data: The history length of each sequence is l i , and obtain the gas pipeline monitoring data set: X = [x1, x2, …, x s ], build the adjacency matrix: Get the objective function of the Laplacian eigenmap: Where L is the normalized Laplacian matrix, and the orthogonal constraint F is added T F = I prevents the solution from degenerating. Using the joint optimization method, the Laplacian feature map and the popular space projection of the original data are combined, and the weight joint sparse penalty is introduced in the optimization process to obtain the overall objective function of structured manifold learning:

[0034] And stF T F = I. Then analyze the objective function N: First construct the auxiliary function: After calculation, the optimization objective function for F is obtained And stF T F=I, and then calculate the matrix L-βX(X T X+αU) -1 X T The eigenvalue decomposition of , select the eigenvector corresponding to the first d smallest eigenvalues ​​other than zero in the eigenvalue decomposition. Then calculate W k+1 :

[0035] According to the formula W = (X T X+αU) -1 X T F,U k+1 :According to the formula Then judge its convergence until it converges and completes feature extraction.

[0036] Figure 3 In order to use uncertainty theory and weighted average method to jointly judge whether the gas pipeline is leaking. The weight distribution is based on the following mathematical basis: Assume that the weight w of each sensor i By its accuracy α i , data stability β i , and monitoring range γ i The weight calculation formula is: Among them, α i Indicates the accuracy of the sensor, reflecting its sensitivity to gas leak indication; β i Indicates data stability and measures the degree to which data is disturbed by external factors; γ i Indicates the monitoring range, which refers to the size of the range covered by the sensor for pipeline monitoring. This weighting method can ensure the relative rationality and scientificity of the weights. First, using the weighted average method, based on the quality and reliability of known data, it is assumed that the pressure sensor has the highest accuracy and is given a weight of 0.5. The flow meter is slightly less accurate and is given a weight of 0.3. The concentration environment monitoring instrument is relatively accurate and is given a weight of 0.2. Then, using uncertainty theory: convert the features extracted in step (4) into uncertainty evidence. Assume that the uncertainty evidence obtained for pressure is 0.8, the uncertainty evidence obtained for flow is 0.6, and the uncertainty evidence obtained for gas concentration is 0.4; then perform evidence synthesis: use the synthesis rules of uncertainty theory to synthesize all uncertainty evidence into a total uncertainty evidence, and use the Dempster synthesis rule for synthesis. The synthesis formula is: After synthesis, the overall uncertainty evidence is 0.55. A threshold is set to 0.6. If the overall uncertainty evidence is greater than or equal to 0.6, it is judged that leakage has occurred; if the overall uncertainty evidence is less than 0.6, it is judged that there is no leakage.

Claims

1. A method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data, characterized by: The following steps are involved: Step (1): Collect multi-source sensor data: pressure, flow, concentration and other data, use the absolute value difference median method to pre-process the data, and conduct preliminary screening of the data; Step (2) The original high-dimensional vector of the gas pipeline leakage monitoring sequence after normalizing and whitening the collected multi-source data is y 1 ,y 2 ,…,y M ,in M is the number of multiple sources, and T is the length of the observed time series; Step (3) constructing N an objective function in feature extraction of structured manifold learning for multiple gas pipeline monitoring sequences; Step (4) Analyze and construct the completed objective function N, unify the original high-dimensional vectors of the normalized and whitened pressure, flow, concentration and other sequences into the same low-dimensional vector to complete feature extraction; the main analysis process is: calculate F in turn k+1 , W k+1 , U k+1 , if it is not converged, the three matrices are calculated again until convergence, and the transformation matrix W is obtained and the original high-dimensional matrix is ​​converted into a low-order matrix F, completing the multivariate data feature extraction of the gas pipeline pressure, flow, and concentration data; Step (5) utilizes the data features of the gas pipeline monitoring data extracted above, and fuses the feature data based on uncertainty theory and weighted average method to determine whether the gas pipeline is leaking.

2. According to claim 1, a method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data is characterized by: For the preprocessed high-dimensional monitoring sequence data, the deep manifold learning method is used to construct an objective function containing orthogonal constraints and weighted joint sparse penalties. By optimizing the objective function, the intrinsic structural relationship of high-dimensional data in the manifold space is analyzed; the heterogeneous monitoring data collected by different sensors are reduced in dimension to generate a unified low-dimensional feature sequence, retaining the key feature information in the high-dimensional data, and completing the feature extraction and preliminary fusion of multi-source monitoring data.

3. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 1 is characterized by: Based on the feature extraction results of multiple sensors, the weighted average method is used to assign weights to the data features of different sensors, and the weight distribution is dynamically adjusted according to the accuracy and reliability of the sensors; the weighted feature data is converted into uncertain evidence, and multiple uncertain evidences are fused using the Dempster synthesis rule to generate an overall uncertain evidence value; according to the set judgment threshold, whether the gas pipeline is leaking is judged based on the overall uncertain evidence value, thereby achieving high-precision status judgment.

4. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 1 is characterized in that: include: The absolute difference median method compares multiple collected data pairwise, removes outliers that deviate greatly from other data, and saves the median value as the standard value for monitoring data consistency assessment; The normalized data has zero mean and unit standard deviation in the same dimension, which facilitates the comparability of data from different sensors.

5. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 1 is characterized in that: include: Whitening processing eliminates the linear correlation between monitoring data from different sensors by calculating the covariance matrix, further compressing data redundancy; The high-dimensional sequence data formed after whitening is more evenly distributed in the feature space, which facilitates the implementation of subsequent feature extraction steps.

6. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 2 is characterized in that: include: In the process of objective function optimization, weighted joint sparse penalty is introduced to solve the heterogeneity problem of data collected by different sensors. During optimization, the eigenvector matrix of manifold learning is generated through iterative calculation. The smaller the eigenvalue corresponding to the eigenvector, the more important the information it retains in the low-dimensional feature space.

7. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 2 is characterized in that: include: The low-dimensional feature sequence generated in the feature extraction stage contains the same indicators and scales, which can achieve unified representation of monitoring data from different sensors; in the dimensionality reduction process, the feature extraction algorithm is highly robust to the noise of high-dimensional sequence data, and can improve the feature expression ability of multi-source monitoring data.

8. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 3 is characterized in that: include: The weighted average method assigns weights to the quality and reliability of different sensors, among which the pressure sensor has the highest priority, followed by the flow meter and concentration monitoring instrument; The dynamic adjustment of weights is optimized based on the historical data quality and operating status of the sensor.

9. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 3 is characterized in that: include: The Dempster synthesis rule combines the uncertain evidence of each sensor data and synthesizes multiple uncertain evidence values ​​into an overall uncertain evidence value; when the overall uncertain evidence value exceeds the set judgment threshold, it can be judged that a gas pipeline leak has occurred.

10. The method for deep flow pattern feature extraction and state determination of multi-source heterogeneous data according to claim 3 is characterized in that: include: During the gas pipeline leakage judgment process, the set judgment threshold is adjusted according to the historical monitoring data of the gas pipeline to adapt to environmental changes or sensor performance fluctuations; the status judgment result can be further improved in accuracy by dynamically updating the model.

Images