Decision-level multi-source fusion gas leakage tracing method and device

Abstract

The invention belongs to the technical field of multi-source fusion, and particularly relates to a decision-making-level multi-source fusion gas leakage tracing method and a decision-making-level multi-source fusion gas leakage tracing device. The method comprises the following specific processes: data acquisition and space-time alignment: realizing decision-level multi-source acquisition on gas leakage data, and performing space-time alignment on the acquired data; preliminary leakage source positioning based on the evidence theory: processing the multi-source data by using the evidence theory, removing conflicting evidences, performing weak evidence weighting enhancement, and fusing the corrected evidences to obtain a region where leakage may occur; in the processing process, recording and obtaining a Jousseme distance; and accurate positioning of the leakage source: determining a dynamic radius based on combined adjustment of a detection error, time decay and a Jousseme distance, and performing weighted positioning by using a least square method to realize accurate source tracing of the leakage source.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of multi-source fusion, and particularly relates to a decision-level multi-source fusion gas leakage source tracing method and device. BACKGROUND

[0002] The rapid urbanization process has driven the growing demand for efficient and clean natural gas among residents. The construction of natural gas infrastructure in various regions is continuously advancing, and the coverage of the pipe network is gradually expanding. The natural gas industry bears a significant responsibility for ensuring the safe operation of this extensive pipe network, which is a prominent challenge. The key to achieving this goal is to accurately locate the gas leakage source, detect hazards in a timely manner, and effectively control explosion and combustion accidents.

[0003] It is difficult to monitor the real-time health status of the current buried natural gas pipe network. Gas companies mainly rely on manual inspection (such as walking about 10 kilometers per day) to detect hidden dangers, but this mode is inefficient, costly, and has poor safety, requires manual recording and is prone to errors, cannot cover 24 hours, and the reporting of abnormalities is delayed. At the same time, natural gas leakage in detection is easily disturbed by underground biogas (also methane), and innovative inspection methods and high-precision detection technologies are urgently needed. Although the existing vehicle-mounted inspection mode overcomes many limitations of manual inspection and has the ability to distinguish between natural gas (methane / ethane) and biogas (methane), its inspection path and detection calculation method are relatively programmed, and during vehicle driving, gas concentration detection is easily disturbed by wind speed, lacking the flexible response capability of manual inspection.

[0004] For the method of gas leakage source tracing, one method involves flow field modeling, i.e., constructing an environmental flow field map to accurately locate the leakage source, such as the Lagrangian model, the Computational Fluid Dynamics (CFD) model, and the statistical model. However, relying on tracking smoke to locate the indoor gas source is both time-consuming and unreliable. Currently, scholars are also interested in using intelligent optimization methods to achieve more efficient gas source location estimation. Kang Jichuan et al. used Computational Fluid Dynamics (CFD) to study typical FPSO leakage scenarios and explored the impact of various parameters on leakage. In addition, Yi Hao et al. proposed a comprehensive signal processing method that combines VMD, BSS, and relative entropy to accurately locate the multi-point pipeline leakage signal and its source, thereby improving the accuracy of multi-point pipeline leakage source location.

[0005] With the development of probability theory and artificial intelligence technology, scholars have begun to use multi-source fusion methods and AI-based methods to assess the risk state of pipe networks. Since there are few methods for gas, some similar methods are listed. Based on multi-source data fusion, the natural gas station equipment operation monitoring method of Yu et al. collects and formats the equipment operation data by determining the safety level of the equipment, and fuses the upper and lower state information of the equipment to obtain the equipment failure probability, and then constructs the equipment operation state evolution model. Based on machine learning, multi-source data fusion intelligent diagnosis is based on various data sources such as thermal parameters, through data preprocessing and feature engineering, key features are extracted, and machine learning algorithms are used for fusion training to realize intelligent diagnosis of engine room faults.

[0006] At present, the mainstream gas leakage inspection and leakage source positioning scheme has significant limitations. On the one hand, the method relying on a single data source (such as relying only on methane concentration) has insufficient accuracy. This is mainly because gas leakage (especially mainly composed of methane) is easily disturbed by environmental factors (such as wind speed, temperature gradient, background methane source) in an open environment, and it is difficult to effectively distinguish between real pipeline leakage and natural release sources such as biogas, resulting in a high risk of false positives and false negatives. On the other hand, although some schemes use deep learning-based feature recognition models, their actual application effect is still not ideal: the efficiency of model training and reasoning is relatively low, making it difficult to meet the timeliness requirements of large-scale pipe network real-time monitoring; more importantly, the decision-making process of such models usually exhibits "black box" characteristics, with a serious lack of explainability. This makes it difficult for field personnel to understand the basis for model judgment, reducing the credibility of the results and hindering targeted optimization of model performance and abnormal diagnosis.

[0007] In summary, the shortcomings of the prior art include:

[0008] 1. Inaccurate tracing of a single detection source: In traditional methods, the leakage area range reported by a single detection source is large and is easily affected by errors, resulting in a risk of misjudgment or missed judgment.

[0009] 2. Uncertainty and conflict between multi-source information: Multi-source detection results often have inconsistencies or even conflicts, which may contain false information.

[0010] 3. Insufficient leakage positioning accuracy: Traditional tracing methods only provide a rough location area and cannot achieve precise positioning. SUMMARY

[0011] Therefore, the present application provides a decision-level multi-source fusion gas leakage tracing method and device, which can accurately locate the leakage position based on multi-source information.

[0012] The technical scheme of the present application is as follows:

[0013] In a first aspect, the application discloses a decision-level multi-source fusion gas leakage source tracing method, and the specific process is as follows:

[0014] Data acquisition and space-time alignment: realize decision-level multi-source acquisition of gas leakage data, and perform space-time alignment on the acquired data;

[0015] Preliminary leakage source positioning based on evidence theory: process the multi-source data by using the evidence theory, eliminate conflicting evidence, perform weak evidence weighting enhancement, fuse the modified evidence, and obtain a possible leakage area; in the processing, Jousselme distance is recorded;

[0016] Precise positioning of the leakage source: determine the dynamic radius based on the detection error, time decay and Jousselme distance joint adjustment, and perform weighted positioning by using the least square method to realize precise source tracing.

[0017] Optionally, the multi-source data of the application includes inspection vehicle collected data, manual inspection collected data and underground buried pipe collected data.

[0018] Optionally, the application regards each source data as an evidence, and determines the possible leakage area based on the evidence theory, and the specific process is as follows:

[0019] Suspected area grid division: taking the leakage position collected by the multi-source as the center, determining the influence range in combination with the detection accuracy, and performing spatial grid division in the corresponding range;

[0020] BPA distribution of the space network: learning the data at the inspection vehicle end by using the BP neural network model, outputting the probability of each grid corresponding to the leakage area, and constructing the corresponding mass value in combination with the softmax result; for the manual and buried pipe data, the mass value is distributed based on the distance decay model;

[0021] Conflict evidence identification and replacement: by calculating the Jousselme distance between each evidence and the average mass value, the conflict evidence is identified, and the mass value of the conflict evidence is replaced by the average mass value;

[0022] Weak evidence weighting enhancement: judging the difference between the mass value of the largest focal element and the mass value of the second largest focal element in each evidence, if the difference is less than a set threshold, it is determined as weak evidence, and the mass value of the largest focal element is linearly weighted and enhanced;

[0023] Multi-source evidence fusion based on DS theory: by using the Dempster-Shafer synthesis rule, the evidence after the conflict processing and enhancement is fused, and the mass value of each grid point, i.e. the leakage risk probability distribution, is obtained.

[0024] Optionally, the specific process of the conflict evidence identification and replacement of the present application is as follows:

[0025] Firstly, the average evidence is calculated:

[0026]

[0027] The Jousselme distance of the average evidence and each evidence is calculated:

[0028]

[0029] wherein D is the Jousselme distance matrix constructed according to the Jaccard similarity;

[0030] Secondly, a conflict threshold δc is set, if d Jousselme > δc, it is considered as conflict evidence, and the mass value thereof is replaced by the average mass value

[0031] Optionally, the specific process of the weak evidence weighting enhancement of the present application is as follows:

[0032] Firstly, the difference Δ between the mass value of the largest focal element and the mass value of the second largest focal element in each evidence is calculated, when the difference is less than a preset threshold τ w , it is considered as weak evidence;

[0033] Secondly, for the weak evidence, the largest focal element mass value promotion parameter w max is calculated, then w max is multiplied by the mass value of the largest focal element:

[0034]

[0035] wherein α>0 is an enhancement coefficient; for the remaining focal elements, the normalization method is used for recalculation.

[0036] Optionally, for the inspection vehicle end, the BP neural network model is used to learn the inspection vehicle end data, to output the probability of each grid corresponding to the leakage area, and the corresponding mass value is constructed combining the softmax result, and the specific process is as follows:

[0037] A. Network input: the gas data detected by the inspection vehicle at a certain time, including the methane concentration and the ethane concentration, and the Euclidean distance di between the current grid center and the suspected leakage position presumed by the inspection vehicle, a total of 3 input features;

[0038] ​B. Network structure: a simple BP neural network classifier is adopted; the input layer is a full connection layer with an input dimension of 3; the network contains two hidden layers, each of which is provided with 16 nodes and 8 nodes, respectively, and the activation function is ReLU, which is used to extract nonlinear features; the output layer is a softmax layer with two neurons, which is used to output the probability distribution of the current grid being in the "leakage" and "non-leakage" states;

[0039] C. Network output: the network output is the probability of the current grid belonging to the "leakage" and "non-leakage" categories; during the training process, the actual leakage condition and the artificial expert experience judgment are used as the supervision label;

[0040] D. Probability distribution function construction: according to the probability result output by the softmax layer, the predicted probability of the grid belonging to the "leakage" category is wi, which is assigned to the grid, i.e., the focal element Xi, and after normalization, the complete mass value m(Xi) is obtained.

[0041] Optionally, the specific process of the leakage source precise positioning according to the present application is as follows:

[0042] Strong support evidence screening: strong evidence with high support for a certain leakage area is screened out from the fusion result, and the spatial position information reported by the strong evidence is extracted as a high-confidence positioning reference;

[0043] Strong evidence confidence radius generation: for the screened strong evidence, the confidence radius of each evidence is dynamically determined in combination with the device detection error, the time decay factor and the Jousselme distance;

[0044] Least squares fusion positioning: based on the above confidence radius, a least squares multi-source fusion optimization model is constructed, and the position of the leakage point is accurately solved through the spatial geometric relationship of the multi-source strong evidence.

[0045] Optionally, the process of generating the strong evidence confidence radius according to the present application is as follows:

[0046] Calculation of basic radius:

[0047]

[0048] wherein, σ i is the standard deviation of the device detection error, and k is the expansion multiple;

[0049] Time decay factor calculation:

[0050] λ i (t)=exp(-β·(tnow-t i )

[0051] wherein, β is the time decay coefficient, which can be fitted through historical data, and tnow is the fusion calculation time, ti The time when the data source i reports the result;

[0052] Jousselme consistency adjustment factor:

[0053] delta i = 1 + eta * d i

[0054] wherein, is an adjustment parameter;

[0055] Calculate the final fusion radius:

[0056]

[0057] Optionally, the least square fusion positioning of the application is:

[0058] R i is used as the evidence credible positioning range, and the established joint positioning model is shown in the formula,

[0059]

[0060] wherein, (x, y, z) is the to-be-determined accurate leakage point position, (x i , y i , z i ) is the position result reported by each strong evidence;

[0061] One of the suspicious leakage positions obtained by the vehicle-mounted inspection is taken as an initial iteration point (x0, y0, z0), in each iteration process, the residual error and the corresponding Jacobian matrix J are calculated, the weight is constructed as a diagonal weight matrix W, and then the position correction amount is solved and the position is updated.

[0062] In a second aspect, the application discloses a decision-level multi-source fusion gas leakage source tracing device, which comprises:

[0063] A data acquisition and space-time alignment module is used for realizing decision-level multi-source acquisition of gas leakage data and performing space-time alignment on the acquired data.

[0064] A preliminary leakage source positioning module based on evidence theory is used for processing the multi-source data by using the evidence theory, eliminating conflicting evidence, performing weak evidence weighting enhancement, fusing the corrected evidence, and obtaining a possible leakage area; in the processing process, Jousselme distance is recorded.

[0065] A leakage source accurate positioning module is used for determining a dynamic radius based on detection error, time attenuation and Jousselme distance joint adjustment, and performing weighted positioning by using a least square method, so as to realize accurate tracing of a leakage source.

[0066] Beneficial effects:

[0067] First, in the traditional method, the leakage area range reported by a single detection source is large and is easily affected by errors, which has the risk of misjudgment or omission. Therefore, the present scheme fuses multi-type detection source information (such as inspection vehicles, manual inspection, underground buried pipe sensors), significantly reduces the leakage point positioning range through multi-source joint reasoning, and improves the accuracy and reliability of leakage tracing.

[0068] Second, the multi-source detection results often have inconsistencies or even conflicts, which may be mixed with false information. To solve this problem, the present scheme introduces Dempster-Shafer evidence theory to construct a unified uncertain information fusion framework, uses Jousselme distance to identify and eliminate conflicting evidence, and effectively improves the stability and accuracy of the fusion results.

[0069] Third, the traditional tracing method only provides a rough location area and cannot achieve precise positioning. To solve this problem, the present scheme designs a two-stage optimization strategy of "evidence fusion + least squares positioning": in the early stage, the suspected leakage area is locked by using grid risk assessment, and in the later stage, a weighted least squares model is constructed based on the reliable evidence point information, and finally the high-precision spatial inversion of the leakage point is realized. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0071] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0072] The embodiments of the present application will be described in detail below with reference to the drawings.

[0073] It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict; and based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present disclosure.

[0074] It is important to note that the various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the teachings herein one skilled in the art should appreciate that an aspect described herein can be implemented independently of any other aspects and that an aspect described herein can be implemented both as any claim and as an aspect of a claim. For example, an apparatus can be implemented using any number of the aspects described herein. In addition, an apparatus can be implemented using other structure and / or functionality not expressly described herein. Similarly, a method can be implemented using any number of the aspects described herein. In addition, a method can be implemented using other structure and / or functionality not expressly described herein.

[0075] As shown in Figure 1 The embodiment of the present application provides a decision-level multi-source fusion gas leakage source tracing method, and the specific process is as follows:

[0076] Data acquisition and space-time alignment: realize decision-level multi-source acquisition of gas leakage data, and perform space-time alignment on the acquired data;

[0077] Preliminary leakage source positioning based on evidence theory: the multi-source data are processed by using the evidence theory, conflicting evidence is removed, weak evidence is weighted and enhanced, the modified evidence is fused, and a region where leakage may occur is obtained; in the processing process, Jousselme distance

[0078] Precise positioning of leakage source: the dynamic radius is determined based on the joint adjustment of detection error, time decay and Jousselme distance, and the weighted positioning is performed by using the least square method, so that the precise source tracing of the leakage source is realized.

[0079] The above process is step-byen subdivided and explained as follows, wherein data acquisition and space-time alignment correspond to step 1, preliminary leakage source positioning based on evidence theory corresponds to steps 2-6, and precise positioning of leakage source corresponds to steps 7-9. Specifically:

[0080] Step 1: Data acquisition and space-time alignment

[0081] Collect multi-source inspection data (mobile inspection vehicle, manual inspection, underground buried pipe sensor), and unify the space-time of each data to ensure that the coordinate systems are consistent and the time stamps are aligned.

[0082] Step 2: Suspected area grid division

[0083] Taking the leakage position reported by the device as the center, the influence range is determined in combination with the detection accuracy (such as error standard deviation), and the spatial grid division is performed in the corresponding area to provide a unified space framework for subsequent modeling.

[0084] Step 3: BPA assignment based on neural network

[0085] The BP neural network model is used to learn the end data of the inspection vehicle, and the probability of each grid being a leakage area is output, and the mass function value (BPA) is constructed by combining the softmax result. For artificial and buried pipe data, BPA is allocated based on the distance attenuation model.

[0086] Step 4: Conflict evidence identification and replacement

[0087] By calculating the Jousselme distance between each evidence and the average mass function, the conflict evidence is identified. If the distance exceeds the set threshold, the evidence is replaced with the average mass, reducing the interference of abnormal information.

[0088] Step 5: Weak evidence weighting enhancement

[0089] Determine the difference between the maximum BPA focus element and the second largest BPA focus element in each evidence. If the difference is less than a set threshold (such as 0.1), it is determined as weak evidence, and the main focus element is linearly weighted and enhanced to improve the subsequent fusion ability.

[0090] Step 6: Multi-source evidence fusion based on DS theory

[0091] Using the Dempster-Shafer synthesis rule, the evidence after conflict processing and enhancement is fused to obtain the leakage risk probability distribution of each grid point.

[0092] Step 7: Strong support evidence screening

[0093] From the fusion results, strong evidence with high support for a certain leakage area is screened out, and its reported spatial position information is extracted as a high-confidence positioning reference, which is further used to accurately narrow down the location of the leakage point.

[0094] Step 8: Strong evidence credible radius generation

[0095] For the screened strong evidence, combined with the device detection error, time attenuation factor and Jousselme distance, the positioning radius of each evidence is dynamically determined.

[0096] Step 9: Least squares fusion positioning

[0097] Based on the above credible radius, a least squares multi-source fusion optimization model is constructed, and the leakage point position is accurately solved through the spatial geometric relationship of multi-source strong evidence.

[0098] In the first stage of the embodiment, the Dempster-Shafer evidence theory is used to build the basic probability assignment (BPA) of multi-source data such as inspection vehicles, manual inspection and underground buried pipes, and the neural network modeling, Jousselme distance conflict elimination and weak evidence enhancement mechanism are introduced to realize the fine description and probability fusion of the leakage risk of each area in the spatial grid, process the conflicting evidence and obtain the area with the largest leakage risk. In the second stage, the strong support evidence with high confidence is selected based on the fusion result, and the dynamic credible radius is constructed by combining the detection error, time decay factor and consistency weight, and finally the weighted least squares method is used to further accurately locate the source by using the uploaded position coordinates. Therefore, the embodiment comprehensively uses the evidence theory and the weighted least squares algorithm to improve the accuracy of leakage identification and positioning. At the same time, the method takes into account the expression of uncertain information and the accuracy of spatial positioning, and shows strong robustness and engineering adaptability in complex inspection scenarios.

[0099] Further, in the embodiment of the application, the specific process of step 1 data acquisition and spatio-temporal alignment is as follows:

[0100] In the current gas leakage detection and inspection process, there are various data sources that can be used. They include large-scale inspection vehicles, manual periodic inspection data, pipe network underground buried pipe detection equipment, etc.

[0101] 1. Inspection method of large-scale inspection vehicle:

[0102] The use of large-scale mobile inspection vehicles for gas leakage detection and source tracing has gradually become the main inspection technology used today. In terms of inspection principle, the inspection vehicle collects real-time methane and ethane concentration data in the air during driving through a high-sensitivity spectrum analyzer. When the methane / ethane gas concentration is detected to be high, it is determined that a leak has occurred. However, it is not possible to accurately determine the location of the leak at this time. Based on the Gaussian plume model that can simulate the physical process of gas diffusion and the PSO optimization algorithm, combined with the detected concentration data, the location of the leakage source is inverted. The final decision-level result data reported is: suspicious leakage area location (x, y), predicted methane concentration C_CH4, and predicted ethane concentration C_C2H6.

[0103] 2. Inspection method of manual periodic inspection:

[0104] Manual inspection mainly uses detection equipment to detect along the approximate position of the pipeline, for example, about 10 kilometers of walking per day. The inspection personnel record the inspection items by manually copying the data, and report the abnormal conditions after the inspection is completed. The final decision-level result data reported is: suspicious leakage area location (x, y), methane leakage concentration Q.

[0105] 3. Inspection method of pipe network underground buried pipe detection equipment:

[0106] In the process of underground pipe network construction, some pipe networks will be installed with detection devices at intervals, including pipe network pressure detection, methane gas detection, etc. If abnormal conditions are detected, suspicious result data will be uploaded: suspicious leakage area position (x, y), methane leakage concentration Q, pipe network pressure F.

[0107] 4. Multi-source data space-time unification:

[0108] Using unified space main reference + sliding time window alignment ensures that various types of sensor data, historical data and auxiliary data can be jointly modeled and analyzed under a unified space-time reference framework, improving the positioning accuracy and robustness of the overall system.

[0109] (1) Space reference calibration and space unification: Select CGCS2000 national geodetic coordinate system as the space main reference, as the reference system for space alignment of all data. At present, most large-scale inspection vehicles, manual periodic inspection data and pipe network underground pipe detection devices have been converted to CGCS2000 coordinate system.

[0110] (2) Time alignment and sliding time window processing: The task cycle of the vehicle-mounted detection device is used as the overall time framework. A sliding time window + time slicing mechanism is adopted to define a standard time window (1 minute) for time slicing archiving and alignment of all real-time collected data and auxiliary data. In each sensor data, each device realizes time synchronization through BDS PPS + NTP combined time service, and the collected data is uniformly annotated with UTC standard time stamp. Through the sliding window mechanism, the whole process time is aligned to ensure time sequence consistency.

[0111] Further, in the embodiments of the present application, the specific process of step 2-6 based on evidence theory to determine the possible leakage area is: the leakage results detected by each source device at different times are obtained, including leakage conditions and leakage points. In order to more accurately determine the leakage area, the present application proposes a multi-source fusion model for further determination of the leakage area. Considering that gas leakage is extremely harmful, one-by-one checking is extremely time-consuming, and the credibility of detection results of different devices is different, there may be conflicts and coupling relationships between devices, the evidence theory is introduced to process the uncertainty of information, and the credibility and accuracy of the leakage area detection are enhanced.

[0112] Introduction of evidence theory

[0113] Dempster-Shafer (DS) requires decision makers to generate a belief assignment function on the hypothesis space based on existing evidence, thereby determining the support trust degree of a solution. DS theory describes uncertain information by using basic probability assignment function, belief function, plausibility function, etc. The related functions in the DS theory algorithm are briefly introduced as follows.

[0114] (1) Recognition frame: The recognition frame θ is the set of all possible solutions to a problem to be solved. The solution at any moment can only take a subset of θ, and all elements in θ satisfy mutual exclusion. The form of the recognition frame is: θ = {θ1, θ2,..., θn}, where θi is an element of the recognition frame.

[0115] (2) Basic probability assignment function: The function m is a mapping from 2 θ to [0, 1], and A is any subset of 2 θ , also known as a proposition, denoted as m is called a basic probability assignment (BPA), also known as a mass function, and m(A) is the basic probability assignment value (BPA value) of the subset A, representing the confidence in A.

[0116] (3) Focal element: For any subset A of the recognition frame, when the subset A contains only one element, it is called a single-element focal element (single focal element). When the subset A contains more than one element, it is called a composite-element focal element (composite focal element).

[0117] (4) Belief function Bel(A): The sum of the basic confidence of all subsets in any subset A of the recognition frame is called the belief function (Belief Function), denoted as Bel(A), which represents the total belief assigned to A.

[0118]

[0119] (5) Plausibility function Pl(A): The plausibility function (Plausibility Function) is the sum of the confidence of all propositions compatible with A, i.e. the belief in A as a non-false proposition, denoted as Pl(A).

[0120]

[0121] (6) Belief interval: The interval [Bel(A), Pl(A)] composed of the belief function Bel(A) and the plausibility function Pl(A) is called the belief interval of the proposition A. Bel(A) and Pl(A) can be considered as the upper and lower probability boundaries of the belief in the proposition A.

[0122] (7) Dempster combination rule: The Dempster combination rule of two mass functions m1, m2 in the same recognition frame is:

[0123]

[0124] where The degree of conflict between two evidences, when k = 1, it means that m1, m2 are completely conflicting, and the two evidences cannot be combined using the Dempster combination rule.

[0125] Application of evidence theory in the method

[0126] The input is multi-time, multi-source data. In order to enhance the accuracy of the leakage area identification, and because of false positives and false negatives, there may be different early warning results in a certain area. In this embodiment, evidence theory is used for leakage risk discrimination to obtain the area where leakage is most likely to occur, and to narrow down the subsequent patrol range.

[0127] In the multi-source gas leakage detection process, multiple suspicious leakage positions are reported on different devices. Each suspicious leakage position is regarded as an evidence. For each evidence, the following steps (1)-(3) are processed to obtain the maas value m(Xi) corresponding to each evidence. The maas value corresponding to the jth evidence is denoted as m j (Xi).

[0128] (1) Suspicious area grid generation: Each device has different detection accuracy (usually represented by standard deviation σ). In order to construct the mass function and perform leakage risk fusion, it is necessary to divide the spatial grid around these suspicious positions for modeling.

[0129] The grid division follows the following principles,

[0130] A, the reported suspicious leakage position is taken as the center

[0131] Each suspicious data record contains a leakage position point, denoted as (xi, yi), which is taken as the center of grid division.

[0132] B, determine the coverage radius according to the detection device accuracy

[0133] Each data source (inspection vehicle, manual inspection, underground buried pipe) corresponds to a different detection error standard deviation σi. Based on this, a spatial coverage radius R = k x σi is defined, where k is a adjustable safety expansion coefficient (usually taken as 2-3), used to cover the influence range of the leakage point.

[0134] C, determine the size of grid division

[0135] In order to realize the discretization analysis of space, each circular coverage area is divided by a fixed grid side length l. The grid side length l is usually determined according to the actual data situation and experience.

[0136] D, two-dimensional grid division within the coverage range

[0137] In each coverage area range, the horizontal and vertical coordinates are equally divided by step length l to generate grid center point coordinates (xi, yi).

[0138] (2) Define the risk identification framework: all generated grid points are uniformly numbered to form the set θ, thereby defining the uniform framework of the risk identification framework:

[0139] θ = {X1, X2,..., XN}

[0140] Where X1~XN represents the grid number, and the center coordinates of the region are (xi, yi).

[0141] (3) Basic probability assignment value (BPA) assignment: introduce a neural network model, according to the detection data of the inspection vehicle, to identify the leakage risk of each grid X1, X2,..., XN, and based on the network output, build the corresponding basic probability assignment (BPA), that is, assign mass value to each grid. At the same time, for artificial inspection and underground buried pipes and other fixed detection ends, according to the reported leakage position and detection accuracy, a mass function model based on spatial distance attenuation is constructed, thereby realizing the unified modeling and evidence expression of different types of data sources.

[0142] For the inspection vehicle end: the inspection vehicle has high gas detection sensitivity, but its moving speed is fast, and the detection position is usually far away from the real leakage source, resulting in certain error in tracing and positioning, so more grids are divided around its suspicious area. For these grids, a BP neural network is used to model the detection data, and a corresponding mass function is constructed according to the softmax probability to express the leakage possibility.

[0143] A. Network input: the gas data detected by the inspection vehicle at a certain time (methane concentration C_CH4, ethane concentration C_C2H6), and the Euclidean distance di between the current grid center and the suspected leakage position estimated by the inspection vehicle, a total of three input features.

[0144] B. Network structure: a BP neural network classifier with simple structure is used. The input layer is a fully connected layer with an input dimension of 3. The network contains two hidden layers with 16 nodes and 8 nodes respectively, and the activation function is ReLU, which is used to extract nonlinear features. The output layer is a softmax layer with two neurons, which is used to output the probability distribution of the current grid being "leakage" and "non-leakage" state.

[0145] C. Network output: the network output is the probability of the current grid belonging to "leakage" and "non-leakage" categories. During the training process, the actual leakage situation and the experience judgment of artificial experts are used as supervision labels to build a binary classification task (0: non-leakage, 1: leakage)

[0146] D. Construction of probability assignment function: Based on the probability results output by the softmax layer, the predicted probability of the grid belonging to the "leakage" category is denoted as wi, and it is assigned to the grid (i.e. focal element) Xi. After normalization, the complete mass value m(Xi) is obtained.

[0147] For both underground pipe leak detection and manual inspection leak detection: Since manual inspection relies heavily on subjective judgment, and underground pipe detection uses direct sensors to monitor leak points, both can be considered relatively accurate data sources for location. Therefore, for these relatively certain leak locations, the following probability allocation model is designed:

[0148] Let the leak source reporting point be (x0, y0), and the grid center be (x0, y0). i y i )but:

[0149]

[0150] Where σ represents the control diffusion range, which can be determined based on the accuracy of the actual detection equipment.

[0151] Normalization and mass function generation: For each source of evidence, assign the probabilities obtained above to w. i Normalization

[0152]

[0153] Let the total confidence value be γ∈[0.7,0.95], which can be selected according to the actual situation, for example γ=0.85.

[0154] γ is adjusted according to the normalized weights. Assign to the corresponding grid:

[0155]

[0156] Where m(Xi) is the mass value.

[0157] The remaining 1-γ values ​​are allocated to the entire set (fuzzy support):

[0158] m(G) = 1 - γ

[0159] (4) Conflicting Evidence Removal Based on Jousselme Distance Model: In the process of multi-source information fusion, different data sources may produce contradictory judgment results due to factors such as detection principles, accuracy, or environmental interference. To identify and mitigate the interference of such conflicting information on the fusion results, this application introduces Jousselme distance to measure the similarity between various pieces of evidence. Specifically, the average mass value of all evidence is first calculated, and then the degree of difference between a single piece of evidence and the average value is measured by Jousselme distance. If the distance between a piece of evidence and the average mass function exceeds a set conflict threshold, it is considered conflicting evidence and is replaced by the average mass function, thereby improving the robustness and stability of the fusion process.

[0160] ① Calculate the average evidence:

[0161]

[0162] Calculate its with Distance to Jousselme:

[0163]

[0164] Where D is the Jousselme distance matrix constructed based on Jaccard similarity.

[0165] ② Conflicting Evidence Judgment and Replacement: Set a conflict threshold δc (e.g., 0.4). If d Jousselme If the value is greater than δc, it is considered conflicting evidence and is discarded, replaced by the average mass value.

[0166] (5) Weak Evidence Weighting Enhancement: In complex environments, some evidence may exhibit low discriminative power due to weak signals, environmental interference, or model uncertainty, meaning its tendency to support a specific leak area is not obvious. Directly incorporating such evidence into fusion may dilute key information, increasing the risk of missed alarms. Therefore, this application introduces a "weak evidence identification and enhancement" mechanism to improve the system's sensitivity at the evidence level. Specifically, The difference Δ between the focal element with the largest mass value (i.e., BPA value) and the focal element with the second largest mass value is defined as follows: when this difference is less than a preset threshold τ... w When the value is 0.1, the evidence is considered weak discriminant evidence.

[0167] For weak discriminative evidence, this application introduces a linear weighted enhancement strategy to increase the BPA value of the focal element with the largest weak discriminative evidence and to normalize other focal elements accordingly, so as to ensure the conservation of total probability quality.

[0168] Calculate the parameter w for increasing the maximum focal element BPA valuemax Then w max is multiplied by the BPA value of the largest focal element:

[0169]

[0170] where a > 0 is an enhancement coefficient (suggested to take 0.2-0.5).

[0171] For the remaining focal elements, the normalization method is used to recalculate.

[0172] (6) DS fusion: According to the Dempster combination rule, the modified evidence is fused to obtain the final fusion result m(Xi).

[0173] Finally, the model outputs the results based on multi-source evidence fusion, quantifying the leakage risk degree of each grid point, providing support for accurate leakage positioning and emergency response.

[0174] Further, the strong evidence leakage point accurate fusion traceability positioning model

[0175] Based on the results of evidence theory fusion, the probability distribution of each suspicious leakage area is obtained. In order to further improve the positioning accuracy, the Jousselme distance between each data source and the final result is calculated to filter out highly correlated evidence sources, and then a multi-source positioning mathematical model is designed to further optimize and converge the leakage position, thereby narrowing the range of the final traceability area.

[0176] 1. Evidence error determination

[0177] Assuming that there are K pieces of evidence pointing to the same area leakage, for the K pieces of leakage evidence pointing to the same area, the detection error of different devices, the timeliness of information and the consistency degree with the average evidence are considered, and the "trustworthy space" is modeled differently in leakage traceability. For this purpose, the application proposes a dynamic radius determination and weighted positioning method based on joint adjustment of detection error, time decay and Jousselme distance.

[0178] (1) Calculate the basic radius: for each evidence i, the initial radius is set to the standard deviation σ i of its device detection error multiplied by an expansion factor (k generally takes 2-4):

[0179]

[0180] (2) Time decay factor calculation: to weaken the influence of outdated information, a time decay function λ i (t) is introduced:

[0181] λ i (t) = exp(-β·(tnow-t i)

[0182] wherein, β is a time attenuation coefficient, which can be fitted by historical data, t now is the time of fusion calculation, t i is the time of the result reported by the data source i;

[0183] (3) Jousselme consistency adjustment factor: introduce a consistency weight (similar to the inverse expression of confidence), give the conflict evidence a radius expansion:

[0184] δ i = 1 + η · d i

[0185] wherein, is an adjustment parameter.

[0186] (4) Calculate the final fusion radius: after comprehensively considering the three factors, the final spatial weight radius of each piece of evidence is:

[0187]

[0188] 2, Multi-source fusion model:

[0189] R i is used as the evidence credible positioning range, and the joint positioning model is established as shown in the formula,

[0190]

[0191] wherein, (x, y, z) is a to-be-determined accurate leakage point position, (x i , y i , z i ) is a reported position result of each strong evidence, R i is the error of the fixed point device, the least square method is used for solving, one suspicious leakage position obtained by a vehicle-mounted inspection is taken as an initial iteration point (x0, y0, z0), in each iteration process, the residual error and the corresponding Jacobian matrix J are calculated, and the weight is constructed as a diagonal weight matrix W, then the position correction amount is solved and the position is updated.

[0192] The application uses multi-source joint analysis, depends on not only a single data source, and improves the accuracy of gas leakage detection traceability. Since the gas detection environment is disturbed, the evidence theory is used to process the credibility of each detection result, a credibility factor is generated for each source, and the leakage detection traceability accuracy is further improved. Finally, a more accurate position coordinate after fusion is obtained, the suspicious leakage range is reduced, and the subsequent problem troubleshooting is more convenient.

[0193] The embodiment of the application relates to a decision-level multi-source fusion gas leakage traceability device, which comprises:

[0194] A data acquisition and space-time alignment module is configured to implement decision-level multi-source acquisition of gas leakage data and to perform space-time alignment on the acquired data;

[0195] A preliminary leakage source positioning module based on evidence theory is configured to process the multi-source data by using evidence theory, to eliminate conflicting evidence, to enhance weak evidence by weighting, to fuse the modified evidence, and to obtain a region where a leakage is likely to occur. In the processing, Jousselme distance is recorded.

[0196] A leakage source accurate positioning module is configured to determine a dynamic radius by jointly adjusting a detection error, a time attenuation, and the Jousselme distance, and to perform weighted positioning by using a least square method, thereby realizing accurate tracing of a leakage source.

[0197] In summary, the above is only a preferred embodiment of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A decision level multi-source fusion gas leakage traceability method, characterized in that, The specific process is: Data acquisition and space-time alignment: realizing decision-level multi-source acquisition of gas leakage data, and performing space-time alignment on the collected data; Preliminary leakage source positioning based on evidence theory: processing the multi-source data by using the evidence theory, eliminating conflicting evidence, and performing weak evidence weighting enhancement, fusing the modified evidence, and obtaining a possible leakage area; During the processing, Jousselme distance is recorded; Precise positioning of the leakage source: determining a dynamic radius based on detection error, time attenuation and Jousselme distance joint adjustment, and performing weighted positioning by using the least square method to realize precise tracing of the leakage source.

2. The method according to claim 1, wherein, The multi-source data includes: inspection vehicle collected data, manual inspection collected data and underground buried pipe collected data.

3. The method according to claim 1 or 2, characterized in that, Each source data is regarded as an evidence, and the possible leakage area is determined based on the evidence theory, and the specific process is: Suspected area grid division: taking the leakage position collected by the multi-source as the center, determining the influence range in combination with the detection accuracy, and performing spatial grid division in the corresponding range; BPA distribution of the space network: learning the inspection vehicle end data by using a BP neural network model, outputting the probability of each grid corresponding to the leakage area, and constructing the corresponding mass value in combination with the softmax result; for the manual and buried pipe data, the mass value is distributed based on a distance attenuation model; Conflict evidence identification and replacement: the Jousselme distance between each evidence and the average mass value is calculated to identify the conflict evidence, and the mass value of the conflict evidence is replaced by the average mass value; Weak evidence weighting enhancement: judging the difference between the mass value of the largest focal element and the mass value of the second largest focal element in each evidence, if the difference is less than a set threshold, it is determined as weak evidence, and the mass value of the largest focal element is linearly weighted and enhanced; Multi-source evidence fusion based on DS theory: the Dempster-Shafer synthesis rule is adopted to fuse the evidence after conflict processing and enhancement, and the mass value of each grid point, i.e. the leakage risk probability distribution, is obtained.

4. The method according to claim 3, wherein, The specific process of the conflict evidence identification and replacement is: Firstly, the average evidence is calculated: Compute its Jousselme distance with the Jaccard distance: Wherein, D is the Jousselme distance matrix constructed according to the Jaccard similarity; Second, a conflict threshold δc is set. If d Jousselme >δc, it is considered as conflict evidence, and its mass value is replaced by the average mass value 5. The method according to claim 4, wherein, The specific process of the weak evidence weighting enhancement is: First, the difference Δ between the mass value of the largest focus in each evidence and the mass value of the second largest focus is calculated, and when the difference is less than a preset threshold τ w , the evidence is considered as weak evidence. Second, for weak evidence, compute the maximum focus mass value boost parameter w max Then multiply w max by the maximum focus mass value: Wherein, α>0 is the enhancement coefficient; for the remaining focal elements, the normalization method is used to recalculate.

6. The method according to claim 3, wherein, For the inspection vehicle end: the BP neural network model is used to learn the inspection vehicle end data, and the probability of each grid corresponding to the leakage area is outputted, and the corresponding mass value is constructed in combination with the softmax result, and the specific process is: A. Network input: the gas data detected by the inspection vehicle at a certain time, including methane concentration and ethane concentration, and the Euclidean distance di between the current grid center and the suspected leakage position estimated by the inspection vehicle, a total of 3 input features; B. Network structure: a simple BP neural network classifier is adopted; the input layer is a fully connected layer with an input dimension of 3; the network contains two hidden layers, each having 16 nodes and 8 nodes, respectively, and the activation function is ReLU, which is used to extract nonlinear features; the output layer is a softmax layer with two neurons, which is used to output the probability distribution of the current grid being "leak" and "non-leak"; C. Network output: the network output is the probability of the current grid belonging to the "leak" and "non-leak" categories; during the training process, the actual leakage and the expert experience are used as the supervision label; D. Probability distribution function construction: according to the probability output of the softmax layer, the predicted probability of the grid belonging to the "leak" category is wi, which is assigned to the grid, i.e., the focal element Xi, and the complete mass value m(Xi) is obtained after normalization.

7. The method according to claim 3, wherein, The specific process of the precise positioning of the leakage source is as follows: Strong support evidence screening: strong evidence with high support for a certain leakage area is screened from the fusion result, and its reported spatial position information is extracted as a high-confidence positioning reference; Strong evidence confidence radius generation: for the screened strong evidence, the confidence radius of each evidence is dynamically determined by combining the device detection error, the time decay factor and the Jousselme distance; Least squares fusion positioning: based on the above confidence radius, a least squares multi-source fusion optimization model is constructed, and the position of the leakage point is accurately solved through the spatial geometric relationship of the multi-source strong evidence.

8. The method according to claim 7, wherein, The process of generating the confidence radius of the strong evidence is as follows: Calculation of the basic radius: wherein σ i is the device detection error standard deviation, k is the expansion factor; Calculation of the time decay factor: λ i (t) = exp(-β · (tnow - t i ) wherein β is a time decay coefficient, which can be fitted by historical data, tnow is the time of fusion calculation, t i is the time when the data source i reports the result; Jousselme consistency adjustment factor: δ i = 1 + η · d i wherein η > 0 is a tuning parameter; Calculation of the final fusion radius:

9. The method according to claim 8, wherein, The least squares fusion positioning is as follows: R i The established joint positioning model is shown in the formula as evidence for a reliable positioning range. Wherein, (x, y, z) is the to-be-determined accurate leakage point position, (x i ,y i ,z i ) is the position result reported by each strong evidence Take the suspicious leakage position obtained by one of the vehicle inspections as the initial iteration point (x0, y0, z0), in each iteration process, calculate the residual and the corresponding Jacobian matrix J, and construct the weight as a diagonal weight matrix W, then solve the position correction amount and update the position.

10. A decision level multi-source fusion gas leakage tracing device, characterized in that, Including: A data acquisition and space-time alignment module for collecting decision-level multi-source data of gas leakage and aligning the collected data in space and time; A preliminary leakage source positioning module based on evidence theory, which processes the multi-source data using evidence theory, eliminates conflicting evidence, weights and enhances weak evidence, and fuses the modified evidence to obtain the area where leakage may occur; During the processing, the Jousselme distance is recorded; A precise leakage source positioning module that determines a dynamic radius based on detection error, time decay and Jousselme distance, and performs weighted positioning using the least squares method to achieve precise tracing of the leakage source.

Images