Pipeline leakage assessment method based on oversampling data enhancement and deep learning
The method addresses subjective expert-based assessments by using deep learning and data enhancement techniques to improve pipeline leakage detection accuracy and reliability, offering actionable insights into risk factors.
Patent Information
- Application Number
- CN202411360482.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-26
- Publication Date
- 2025-01-03
AI Technical Summary
Current pipeline leakage assessment methods rely heavily on expert judgment, leading to subjective assessments that are not universally applicable, and struggle with sample imbalance, reducing the accuracy and reliability of leakage detection in water supply networks.
A pipeline leakage assessment method using oversampling data enhancement and deep learning, incorporating mixed Gaussian distribution pattern learning, multi-head attention mechanisms, and residual connections to balance data sets and enhance feature interactions, coupled with SHAP analysis for interpretability.
Improves the accuracy and reliability of pipeline leakage detection by learning from actual patterns, balancing data samples, and providing actionable insights into risk factors, enhancing model performance and interpretability.
Smart Images
Figure CN119249277A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of municipal water supply pipe networks, and in particular to a pipeline leakage assessment method based on oversampling data enhancement and deep learning. Background Art
[0002] The pipeline system for delivering and distributing water to users in a water supply project is called a water supply / water supply network / pipeline, which includes water delivery channels, water distribution networks, booster pump stations, water towers, pools, and network ancillary facilities. The channels from the water source to the water plant only serve to deliver water, and are called water delivery networks; the pipelines from the water plant are called water distribution networks. The pipelines that mainly deliver water in the water distribution network are called trunk pipes, and the pipelines that branch out from the trunk pipes and serve to distribute water are called branch pipes, and the branches that connect to users are called user branches. Pipe burst refers to the situation where the pipeline leaks and rises to the ground due to structural damage to the pipeline, and must be repaired immediately. There are many reasons for pipeline bursts, such as pipeline-related factors, pipeline geography-related factors, pipeline process-related factors, etc. In reality, after a pipe burst accident occurs, it may sometimes take several days to be discovered, and sometimes it may not even be discovered. This is because the leakage caused by the pipe burst directly enters the nearby drainage channels and cannot reach the ground. Therefore, leakage assessment and monitoring of the water supply network are particularly important tasks.
[0003] At present, the health assessment scheme widely used in pipeline network assessment is the hierarchical analysis method based on fuzzy mathematics. It comprehensively considers pipeline risk factors, uses expert knowledge to assign weights to risk factors and quantifies pipeline risk assessment results. This method can evaluate the relative level of pipeline health, but the subjectivity brought about by the introduction of expert knowledge limits its practical application across regions and systems. Summary of the invention
[0004] In order to solve the problems existing in the background technology, the present invention proposes a pipeline leakage assessment method based on oversampling data enhancement and deep learning.
[0005] A pipeline leakage assessment method based on oversampling data enhancement and deep learning, comprising the steps of:
[0006] S100, obtaining pipeline characteristic data, and dividing the pipeline characteristic data into numerical characteristic data and attribute characteristic data;
[0007] S200, processing the numerical feature data into a pattern-type scalar, and processing the attribute-type feature data into an attribute-type feature mask;
[0008] S300, mapping the numerical feature data and the attribute feature data to a unified dimensional space;
[0009] S400, combines pipeline features and generates high-order interactions of risk factors;
[0010] S500, using residual connections to promote inter-layer information exchange and achieve risk probability mapping.
[0011] Based on the above, in step S200, when processing pipeline numerical feature data, firstly, mixed Gaussian distribution pattern learning is performed, the pattern type of generated data is used as part of the conditional vector, the generated numerical range is constrained, and then probability weights are assigned to the learned patterns, and the pattern classification to which the generated sample belongs is selected according to the weights, and finally the pattern class scalar is generated; this process is represented by the following formula, for all numerical features C i :
[0012]
[0013] Among them, c i,j is the jth numerical feature, μ k is the learned weight of the kth pattern, η k is the Gaussian distribution mean of the learned k-th mode, φ k is the standard deviation of the kth mode, α i,j The data value within the schema to which it belongs.
[0014] Based on the above, in step S200, when processing pipeline attribute feature data, for the attribute feature data D of the i-th row, i , generate a zero-filled mask vector m i , where m i The kth element of represents the mask of category k; for each category k, according to the category in column D i The number of times n appears in the calculation probability function P(D i =k) = n / N, select category k according to the probability function * , and the mask vector m i Corresponding to k * Update the element to 1; concatenate all generated attribute feature masks to form the final discrete quantity generation result The virtual leakage sample generated by the final generator is expressed as follows:
[0015]
[0016] in: is the jth pipeline sample feature finally generated, c is the dimension of the numerical feature, is the numerical feature of the generated sample, N is the total number of categories, is the mode quantity of the numerical feature of the generated sample, which is a one-hot encoding. If the calculated value belongs to the learned mode 3, then β i,j =[0,0,1].
[0017] Based on the above, in step S400, a representation learning layer for combining pipeline features to generate high-order interactions of risk factors is constructed, and the process is as follows:
[0018]
[0019] In the formula, e m ,e k are the embedded features m and k of a pipeline respectively, and in a separate attention head h, the inner product function ψ is used (h) (·,·) to calculate the similarity between two features, is the transformation matrix that maps the original embedding space to the new space, and calculates the attention score between m and k by similarity And update the pipeline embedding feature m as Concatenate the calculation results of all the head attention mechanisms to finally get the pipeline high-order interaction representation
[0020] Based on the above, in step S500, the results of the multi-layer representation learning layer are aggregated through residual connections, and the final binary classification prediction vector is output using the Sigmoid function. The process is as follows:
[0021]
[0022] In the formula, The residual representation generated for the mth attribute feature, W Res is the residual weight matrix to be learned, w T is the mapping layer weight matrix, b is the bias to be learned, Predict leakage risk results for the model.
[0023] Based on the above, in step S500, the leakage risk result is predicted. That is, the pipeline leakage probability, and pipelines with leakage probability within a certain range are divided into the same pipeline health.
[0024] Based on the above, the health threshold is:
[0025] 0 to 0.5: pipeline status is healthy;
[0026] 0.5 to 0.75: pipeline status is average;
[0027] 0.75 to 1: Pipeline status is dangerous.
[0028] Based on the above, including step S600, the influence of different risk factors is analyzed by SHAP, and the formula is:
[0029]
[0030] In the formula, S is the feature subset, |S| is the number (i.e., size) of the S feature subset, p is the number of the overall input features, x is the set of all features of the pipeline samples, and x S is the remaining feature set after removing the feature subset S, and f(x S ∪i) represents the model output including feature i, and f(x S ) represents the model output without feature i.
[0031] The present invention has prominent substantive features and remarkable progress compared with the prior art. Specifically, the present invention uses the prediction result of a data-driven model (i.e., a leakage assessment model) as the reference basis for pipeline health assessment, without relying on expert judgment or other artificial risk level labels. It learns knowledge from the typical patterns mined from pipeline data in practice, deeply analyzes the characteristics of pipeline attributes, and addresses the problem of sample imbalance by adopting the conditional vector optimization adversarial neural network technology to balance the sample ratio of leakage pipelines and normal pipelines in the database; by exploring the mutual relationship between pipeline risk factors, a multi-head attention mechanism is used to construct high-order combinations between risk features, overcoming the lack of accuracy of traditional models in identifying leakage pipelines and significantly improving the performance and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 is a schematic flow block diagram of the present invention.
[0033] Figure 2 is a schematic diagram of GAN data augmentation of the present invention.
[0034] Figure 3 is a schematic diagram of the pipeline risk prediction model architecture of the present invention.
[0035] Figure 4 is a comparison diagram of the sample balance results of the leakage sample data augmentation of the present invention.
[0036] Figure 5 is a schematic diagram of the SHAP analysis result of the present invention.
[0037] Figure 6 is a schematic diagram of the heuristic data analysis result of the present invention for the pipeline age. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0039] A pipeline feature database based on data governance was constructed, and the pipeline data of B2 and B3 metering areas in a city in central China were collected as data to be governed. The final governance results were 320,713 basic pipeline data, 3,055 leakage point data, and 1,964 maintenance ledger records. The data sources are mainly divided into the following three categories:
[0040] (1) Pipeline basic attribute data from GIS database;
[0041] (2) Maintenance record data of pipeline operation and maintenance units;
[0042] (3) The maintenance unit maintains record data of leakage points.
[0043] In order to achieve high availability of the data base, the pipeline segment number is used as the main axis keyword in the three types of data sources, and the "three-in-one" data positioning criteria are referred to to preliminarily complete the arrangement of the pipeline data base with high completeness. After that, the pipeline risk factors are screened by referring to a large number of simulations, field experiments, historical data analysis and relevant standards that have been conducted in the academic community. These factors are shown in Table 1 and can be divided into three categories: (1) physical and pipeline related factors; (2) geographical related factors; and (3) process related factors.
[0044]
[0045]
[0046] Table 1
[0047] The data is further managed according to the risk factors that have been screened out, and the data of each pipeline unit is collected according to the pipeline section number to build a water supply pipeline feature database. In this step, we use data management to identify the key factors affecting pipeline health and determine the input of the pipeline health status assessment model.
[0048] The scarcity of leakage records in real scenarios means that existing model training can only be performed on unbalanced data sets, which reduces the reliability of model results and the ability to migrate across regions, especially in supervised model training. In this data set case, the data governance results also show that the number of common samples is more than 100 times that of leakage samples, which is extremely unbalanced. In order to avoid the decline of generalization performance caused by model training on the data set, a data enhancement solution specialized for pipeline features is needed to balance the data set.
[0049] Considering that in addition to the numerical attributes such as pipeline age and pipeline, pipeline features also include sparse high-dimensional attributes such as pipeline material and connection type, and the traditional interpolation data enhancement method cannot effectively generate similar samples when sampling sparse high-dimensional features. This embodiment models the numerical and attribute features separately, and introduces conditional vectors represented by masks during the training process of the GAN model to standardize the conditions to be followed when generating synthetic data.
[0050] like Figure 1 As shown in the figure, a pipeline leakage assessment method based on oversampling data enhancement and deep learning is used to perform pipeline network leakage assessment using GAN data enhancement specialized for pipeline features, risk factor interactive learning based on multi-head attention mechanism, and pipeline risk factor analysis method based on SHAP value. The data enhancement method used in this example and the excellent performance of the data-driven model were tested in two main experiments: (1) The distribution of generated data and original data in high-dimensional space was observed through visualization methods; (2) The evaluation accuracy of leaky pipelines and all pipelines was compared with the commonly used data-driven model. The higher the recognition accuracy, the more reliable the probability mapping result of the model. Among them, the comparative experiment is used to screen the optimal model by dividing the data in the pipeline feature database into two parts, 80% for training the model and 20% for testing experiments.
[0051] When processing pipeline numerical features, we first perform mixed Gaussian distribution pattern learning, take the pattern type of generated data as part of the conditional vector, constrain the generated numerical range, assign probability weights to the learned patterns, and select the pattern classification to which the generated sample belongs based on the weights, and finally generate the pattern class scalar. This process can be expressed as follows: For all numerical features C i :
[0052]
[0053] In this formula: c i,j is the jth numerical feature, μ k is the learned weight of the kth pattern, η k is the Gaussian distribution mean of the learned k-th mode, φ k is the standard deviation of the kth mode, α i,j The data value within the schema to which it belongs.
[0054] When processing pipeline attribute features, the goal of this embodiment is to constrain the generator results Satisfy as much as possible Specifically, for the attribute feature D of the i-th row i , generate a zero-filled mask vector m i , where m iThe kth element of represents the mask of category k; for each category k, according to the category in column D i The number of times n appears in the category (the total number of categories is N), calculate the probability function P(D i = k) = n / N, select category k according to the probability function * , and the mask vector m i Corresponding to k * Update the element to 1; concatenate all generated attribute feature masks to form the final discrete quantity generation result
[0055] The virtual leakage samples generated by the final generator can be expressed as:
[0056]
[0057] Where: is the jth pipeline sample feature finally generated, c is the dimension of the numerical feature, is the numerical feature of the generated sample, is the mode quantity of the numerical feature of the generated sample, which is a one-hot encoding. If the calculated value belongs to the learned mode 3 (mode η3. The variational Gaussian mixture mode VGM found three modes, named η1, η2 and η3), then β i,j =[0,0,1].
[0058] In reality, the number of common samples is more than one hundred times that of leaky samples, which is extremely unbalanced. Therefore, in order to avoid the degradation of generalization performance caused by training the model on the data set, a specialized data enhancement scheme for pipeline characteristics is needed to balance the data set. To solve this problem, CTGAN is used to model the distribution of leaky data and sample from the distribution. This embodiment selects the general GAN paradigm as the generator and discriminator, such as Figure 2 As shown, the virtual pipeline features generated by the generator Compared with the actual pipeline characteristics r j The cross entropy of is used as the loss function. Using the cross entropy loss function can accurately reflect the distance between the model output and the true classification label. The cross entropy loss function formula is as follows:
[0059]
[0060] Where: is the virtual pipeline feature, r j is the real pipeline characteristic, is the cross entropy loss function, and the smaller the loss is when the prediction is closer to the true label.
[0061] According to the characteristics of the pipeline, we build a leakage assessment model with reference to the self-attention mechanism and residual connection commonly used in the field of deep learning, such as Figure 3 As shown, it includes the following three parts:
[0062] (1) Mapping numerical features and attribute features to a feature embedding layer with a unified dimensional space;
[0063] (2) Combining pipeline features to generate a representation learning layer for high-order interactions of risk factors;
[0064] (3) Residual connections are used to promote information exchange between layers and realize the result generation layer of risk probability mapping.
[0065] The representation learning layer is the core of the model, which uses a multi-head attention mechanism to combine features
[11] . Due to the complex interaction between pipeline features, this embodiment maps pipeline features to a high-dimensional space for high-order combined feature modeling. By stacking multiple layers of representation learning layers, the model can spontaneously learn the interaction results of pipeline features in higher-order dimensions and determine which features should be combined to form meaningful high-order features. The process is as follows:
[0066]
[0067] In the above formula: e m ,e k are the embedded features m and k of a pipeline respectively, and in a separate attention head h, the inner product function ψ is used (h) (·,·) to calculate the similarity between two features, is the transformation matrix that maps the original embedding space to the new space, and uses the similarity to calculate the attention score between m and k And update the pipeline embedding feature m as Concatenate the calculation results of all the head attention mechanisms to finally get the pipeline high-order interaction representation
[0068] The result generation layer aggregates the results of the multi-layer representation learning layer through residual connections, and uses the Sigmoid function to output the final binary classification prediction vector. The process is as follows:
[0069]
[0070] In the above formula: The residual representation generated for the mth attribute feature, W Res is the residual weight matrix to be learned, w T is the mapping layer weight matrix, b is the bias to be learned, Predict leakage risk results for the model.
[0071] The residual network enables the model to take into account the multi-order composite feature interaction combinations. Sigmoid maps the impacts of different risk factors on pipeline leakage to the probability space, and its output result is the pipeline leakage probability. Pipelines with leakage probabilities close to each other will be classified into the same pipeline health level. In this embodiment, the health level threshold is determined as follows:
[0072] (1) 0 to 0.5: The pipeline is in a healthy state;
[0073] (2) 0.5 to 0.75: The pipeline is in a general state;
[0074] (3) 0.75 to 1: The pipeline is in a dangerous state.
[0075] Taking the prediction result of the data-driven model as the reference basis for pipeline health assessment, it does not need to rely on expert judgment or other artificial risk level labels, and learns knowledge from the typical patterns mined from pipeline data in practice. By exploring the mutual relationships between pipeline risk factors and using the multi-head attention mechanism to construct high-order combinations between risk features, it overcomes the lack of accuracy of traditional models in identifying leaking pipelines and significantly improves the performance and reliability of the model.
[0076] For non-professional statisticians in pipeline operation and maintenance units, it is difficult to provide help for pipeline repair only by obtaining the quantitative leakage probability. Existing data-driven models can provide probability assessments of quantitative pipeline risks, but they usually ignore the analysis of individual pipeline risk factors, resulting in the model being unable to provide specific and operable suggestions for the maintenance and repair of specific pipelines. To improve this defect and enhance the interpretability of the model, here we introduce SHAP (SHapley Additive exPlanation, a machine learning model interpretation tool) to analyze the impacts of different risk factors. SHAP adopts the Shapley value in game theory, which quantifies the contribution of each feature to the final prediction result of the model, indicating the expected change in the model prediction result when assigning the attribute values of each feature to adjust the feature. As shown in the following formula:
[0077]
[0078] In the above formula: S is the feature subset, 丨S丨 is the number (i.e., size) of the S feature subset, p is the number of the overall input features, x is the set of all features of the pipeline sample, x S is the remaining feature set after removing the feature subset S, f(x S ∪i) represents the model output including feature i, and f(x S ) represents the model output without feature i.
[0079] The results of SHAP analysis are generally visualized through a columnar overall characteristic diagram to show the contribution of all risk factors to the final prediction results. In the overall characteristic diagram, the order and width of the risk factors reflect the level of influence of different characteristics. The SHAP method analyzes the impact of risk factors on individual pipelines in a visual and intuitive way, which can guide the maintenance and repair plans for different pipelines in practice.
[0080] In this embodiment, the TSNE-KDE dimensionality reduction graph is used to jointly evaluate the performance of the GAN data enhancement method, such as Figure 4 As shown in the figure, both the traditional SMOTE method and the optimized GAN method can effectively generate clear cluster boundaries and balance different sample ratios. Compared with the existing SMOTE method for data interpolation, the method used in this case generates more evenly distributed samples, which reflects the richness of the generated data. In addition, the closer the distance between the generated samples of the optimized GAN method and the original samples in the TSNE dimensionality reduction graph, the higher their similarity. These samples with smaller differences carry more information, which can better refine the model decision boundary during training, and improve the training quality of the model while balancing the leakage and normal pipeline samples.
[0081] The optimal model screening results are shown in Table 2 below. In this embodiment, the evaluation accuracy of different data-driven models on leaking pipe sections and overall pipe sections is compared when different proportions of synthetic data participate in model training. The optimal model comparison results show that the leakage assessment model of this embodiment achieved the best performance in the test after adopting conditional GAN and introducing the attention mechanism. Compared with the commonly used data-driven model, the recognition accuracy of the leaking pipeline is improved by 2.30%, and the recognition accuracy of the overall pipeline is improved by 1.84%.
[0082]
[0083]
[0084] Table 2
[0085] The SHAP method not only provides the impact of risk factors, but also improves the interpretability of the model. The SHAP analysis results of the entire pipeline and two specific single pipelines are shown in the figure below. Figure 5 As shown, the age and material of the pipeline (especially gray cast iron) provide the greatest contribution to the prediction results of the model, which is the greatest risk factor for the pipeline. In this example, the reliability of the SHAP analysis results of this case is comprehensively verified by comparing the trend of the SHAP analysis results with the risk factor heuristic data analysis and the prediction results of the pipeline model in the study area.
[0086] Specifically, the leakage frequency of pipeline materials is first analyzed (socket-type asbestos cement joints are generally matched with gray cast iron). The analysis results are shown in Table 3. The leakage frequency per kilometer of PE pipes and gray cast iron pipes with larger contribution values of the model analysis is significantly higher than that of pipelines of other materials; among the number of pipelines predicted to be dangerous, the number of PE pipes and gray cast iron pipes is higher than that of other pipes. It is worth noting that in the area studied in this case, the pipelines with leakage are concentrated in pipelines with small diameters and short lengths, especially DN100 pipelines, accounting for 34.51% of the total number of leakage pipelines, and their main materials are gray cast iron and PE pipes. Gray cast iron is sensitive to corrosion inside and outside the pipe, and its probability of leakage is higher. Compared with gray cast iron pipes, PE pipes are shorter in length and the connection method is mainly hot melt. The installation environment is complex, the overall quality is directly related to the construction quality, and the structural reliability is relatively insufficient. The analysis results show that the model and its SHAP analysis results can significantly distinguish the impact of pipeline materials (including connection methods) on leakage risks.
[0087]
[0088]
[0089] Table 3
[0090] The results of the heuristic data analysis on pipeline age factors are as follows: Figure 6 As shown in the figure, the number of leakages per kilometer of pipelines with an age of more than 15 years is higher than that of pipelines with a lower age. The violin distribution diagram of leakage and normal pipeline age distribution shows that the leakage ratio of pipelines increases significantly with the increase of pipeline age. The model prediction results show that among the pipelines assessed as dangerous, pipelines with an age of more than 30 years account for 35.2%, pipelines with an age of 20-30 years account for 20.4%, and pipelines with an age of 15-20 years account for 39.1%. The age of pipelines is significantly correlated with the risk of pipeline leakage, which is basically consistent with the trend in the SHAP analysis that the older the pipeline, the higher the risk of leakage.
[0091] This example introduces the SHAP interpretation model to achieve visual analysis of risk factors at both the overall pipeline and individual pipeline levels, improving the model's interpretation capabilities while enabling the leakage assessment model to have the ability to specifically guide practice. At the same time, the rationality of the model is demonstrated. Among the pipeline risk factors, pipeline age and pipeline material are the two factors with the greatest impact.
[0092] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.
Claims
1. A pipeline leakage assessment method based on oversampling data enhancement and deep learning, characterized in that: Includes steps: S100, obtaining pipeline characteristic data, and dividing the pipeline characteristic data into numerical characteristic data and attribute characteristic data; S200, processing the numerical feature data into a pattern-type scalar, and processing the attribute-type feature data into an attribute-type feature mask; S300, mapping the numerical feature data and the attribute feature data to a unified dimensional space; S400, combines pipeline features and generates high-order interactions of risk factors; S500, using residual connections to promote inter-layer information exchange and achieve risk probability mapping.
2. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: In step S200, when processing pipeline numerical feature data, firstly, mixed Gaussian distribution pattern learning is performed, the pattern type of generated data is used as part of the conditional vector, the generated numerical range is constrained, and then probability weights are assigned to the learned patterns, and the pattern classification to which the generated sample belongs is selected according to the weights, and finally the pattern class scalar is generated; this process is represented by the following formula: for all numerical features C i : Among them, c i,j is the jth numerical feature, μ k is the learned weight of the kth pattern, η k is the Gaussian distribution mean of the learned k-th mode, φ k is the standard deviation of the kth mode, α i,j The data value within the schema to which it belongs.
3. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: In step S200, when processing pipeline attribute feature data, for the attribute feature data D of the i-th row i , generate a zero-filled mask vector m i , where m i The k-th element of represents the mask of category k; For each category k, the i The number of times n appears in the calculation probability function P(D i =k) = n / N, select category k according to the probability function * , and the mask vector m i Corresponding to k * Update the element to 1; concatenate all generated attribute feature masks to form the final discrete quantity generation result The virtual leakage sample generated by the final generator is expressed as follows: in: is the jth pipeline sample feature finally generated, c is the dimension of the numerical feature, is the numerical feature of the generated sample, N is the total number of categories, is the mode quantity of the numerical feature of the generated sample, which is a one-hot encoding. If the calculated value belongs to the learned mode 3, then β i,j =[0,0,1].
4. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: In step S400, a representation learning layer for combining pipeline features to generate high-order interactions of risk factors is constructed, and the process is as follows: In the formula, e m ,e k are the embedded features m and k of a pipeline respectively, and in a separate attention head h, the inner product function ψ is used (h) (·,·) to calculate the similarity between two features, is the transformation matrix that maps the original embedding space to the new space, and calculates the attention score between m and k by similarity And update the pipeline embedding feature m as Concatenate the calculation results of all the head attention mechanisms to finally get the pipeline high-order interaction representation 5. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: In step S500, the results of the multi-layer representation learning layer are aggregated through residual connections, and the final binary classification prediction vector is output using the Sigmoid function. The process is as follows: In the formula, The residual representation generated for the mth attribute feature, W Res is the residual weight matrix to be learned, w T is the mapping layer weight matrix, b is the bias to be learned, Predict leakage risk results for the model.
6. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: In step S500, the leakage risk result is predicted. That is, the pipeline leakage probability, and pipelines with leakage probability within a certain range are divided into the same pipeline health.
7. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 6 is characterized in that: The health thresholds are: 0 to 0.5: pipeline status is healthy; 0.5 to 0.75: pipeline status is average; 0.75 to 1: Pipeline status is dangerous.
8. The pipeline leakage assessment method based on oversampling data enhancement and deep learning according to claim 1 is characterized in that: The method comprises step S600, wherein the influence of different risk factors is analyzed by SHAP, and the formula is: Wherein, S is a feature subset, 丨S丨 is the number of the S feature subsets, p is the number of the overall input features, x is the set of all features of the pipeline samples, x S is the remaining feature set after removing the feature subset S, f(x S ∪i) represents the model output including feature i, f(x S ) represents the model output excluding feature i.