A pumping unit well working condition identification method based on measured multi-source incomplete data

Abstract

A pumping unit well working condition recognition method based on measured multi-source incomplete data belongs to the technical field of oil field intelligent oil production. Its characteristics are as follows: step 1, through a double-path structure, multi-source data features are extracted by means of a residual injection mechanism. Step 2, the reliable information of high confidence samples is adaptively propagated to the missing view samples, realizing effective diffusion of global reliable information to local uncertain areas. Step 3, by maximizing cross-view shared information mutual information and minimizing redundant noise information, the class discrimination of the completed feature is strengthened, and the fusion bias is relieved. Step 4, a total loss function composed of a basic classification task loss and a mutual information constraint loss is constructed to realize incomplete multi-source data working condition recognition. In the technical scheme of the application, through the collaborative optimization of the whole link of encoding, completion and fusion, the accuracy, noise resistance and generalization performance of incomplete multi-view classification are significantly improved.

Description

Technical Field

[0001] A method for identifying the operating conditions of pumping wells based on incomplete multi-source measured data belongs to the field of intelligent oil production technology. Background Technology

[0002] In actual oil production at oilfield sites, pumping units, as core production equipment, rely on the collaborative analysis of multi-source monitoring data, including load, displacement, current, and pressure, for operational monitoring. Complete operational data is the foundation for achieving well fault early warning, production parameter optimization, and digital oil production management. However, oilfield production sites are harsh environments, often located in remote areas, and are susceptible to interference from external factors such as high temperatures, wind, sand, vibration, and corrosion. In addition, uncontrollable factors such as sensor aging, transmission failures, and signal interruptions can easily lead to partial, fragmented, or even large-scale data loss in the measured multi-source operational data of pumping units. Such incomplete data directly results in distorted operational condition judgments and ineffective early warnings, seriously hindering intelligent monitoring of pumping units.

[0003] To address the aforementioned issues, researchers have proposed numerous incomplete multi-view learning methods, which can be broadly categorized into two types: (1) Methods based on available views. These methods directly discard missing view information and rely solely on existing view data for classification. While computationally efficient, they suffer from severe information loss, particularly in scenarios with high missing rates, where classification performance is poor. (2) Methods based on missing view generation. These methods utilize deep learning techniques such as autoencoders and generative adversarial networks to reconstruct missing views based on patterns from available views. These methods aim to provide comprehensive data representation by generating high-quality missing views, but their effectiveness heavily depends on the authenticity and reliability of the generated data—if the generated views deviate from the true data distribution, misleading information may be introduced, leading to a decline in model performance.

[0004] Trustworthy cross-view completion is a representative method in the field of incomplete multi-view classification. It demonstrates superior performance compared to traditional methods through the collaborative work of three modules: cross-view feature learning, uncertainty-aware completion, and trustworthy fusion. Specifically, the cross-view feature learning module mines view-specificity and cross-view... Figure 1 The system obtains reconstructed features by merging consistency, the completion module recovers the missing view by using uncertainty-aware weights and neighborhood supervision, and the trustworthy fusion module fuses the completed data to generate classification predictions.

[0005] However, the method still has the following limitations, which restrict further performance improvement: (1) The feature encoding is a static fixed structure. The cross-view feature learning module of the trustworthy cross-view completion extracts view-specific features through an independent encoder, which fails to effectively separate view-private information from cross-view shared semantics, resulting in insufficient alignment of cross-view semantic space. In heterogeneous multi-view data, feature representation inconsistency is likely to occur, affecting the reliability of completion. (2) The completion process lacks dynamic trustworthy information propagation. Its completion module relies on Euclidean distance to filter local neighborhood samples and only achieves completion through local constraints and uncertainty weights, ignoring the dynamic propagation of sample confidence. In high missing rate or noisy scenarios, limited local information is difficult to provide sufficient supervision, and the neighborhood selection does not evaluate the reliability of the samples themselves, which easily introduces noise, resulting in poor stability of the completion results. (3) The information filtering efficiency in the fusion stage is low. The credible fusion module estimates view-level uncertainty based on evidence-based deep learning and performs fusion using Dempster-Shafer theory. However, it only focuses on view-level uncertainty and fails to specifically suppress redundant and noisy information, nor does it fully explore the discriminative value of shared information. When view quality is uneven, the retention of irrelevant information will seriously affect classification performance. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a collaborative optimization of the entire link of encoding, completion and fusion, which significantly improves the accuracy and robustness of incomplete multi-source data condition identification.

[0007] The technical solution adopted by this invention to solve its technical problem is: a method for identifying the operating conditions of pumping wells based on measured multi-source incomplete data, characterized by including the following steps: Step 1: Construct a dual-path feature extraction structure, then input the incomplete multi-view data into the dual-path feature extraction structure to obtain the private features and shared semantic features of the incomplete multi-view data respectively. Use the residual injection mechanism to extract features from the multi-source data respectively, and finally generate the evidence vector. Step 2: Based on the component graph structure of feature similarity between samples, the credible information of high-confidence samples is adaptively propagated to the missing view samples, realizing the diffusion of global credible information to local uncertain regions; Step 3: In the fusion stage, by maximizing cross-view shared information and mutual information and minimizing redundant noise information, the class discriminativeness of the completed features is enhanced, while mitigating the fusion bias caused by uneven view quality. Step 4: Construct a total loss function consisting of the basic classification task loss and the mutual information constraint loss to achieve incomplete multi-source data condition identification.

[0008] Preferably, step 1 includes the following steps: Step 1-1, for the first View data Using a private feature extractor Extract the first Private features of each view The corresponding formula is: Simultaneously, a parameter-sharing feature extractor is used. Extract the first Each view shares semantic features This maps different views to a unified semantic space, and the corresponding formula is: ; Step 1-2: Introduce learnable residual weights Dynamically adjust the contribution of shared features to generate enhanced features. The corresponding formula is: in, For learnable residual weights, Use the Sigmoid activation function; Steps 1-3, through the evidence vector transformation layer Generate evidence vectors The calculation is as shown in the formula below: in, For evidence vectors, To enhance features.

[0009] Preferably, step 2 includes the following steps: Step 2-1, for the first Sample of a view Its evidence vector Parameterized Dirichlet distribution, with distribution parameters as follows: The uncertainty of this view The calculation is as follows: in, Indicates the number of categories; Corresponding confidence level The calculation formula is: ; Step 2-2: Construct a dynamic adjacency matrix based on feature space. ,in The batch size; for any two samples within the batch and The formula for calculating connection weights is: in, For cosine similarity, Temperature coefficient; They represent the first The i-th and j-th samples of the view.

[0010] Steps 2-3: Confidence propagation and reweighting; Design an iterative graph propagation mechanism to correct local uncertainty estimation; initialize the confidence matrix. In the t-th iteration ( The confidence update formula is: in, It is a balancing factor.

[0011] Steps 2-4: Calculate global confidence using multi-view information. The original evidence was then reweighted.

[0012] Preferably, steps 2-4 include the following steps: Step 2-4-1, using learnable parameters Calculate the fusion weights of each view The calculation formula is: in, They represent the first Learnable parameters for the j-th and j-th views.

[0013] Step 2-4-2, calculate the global confidence score. The calculation formula is: Step 2-4-3: Adaptively reweight the original evidence using global confidence, calculated as follows: in, Indicates the global confidence level. This represents the evidence vector.

[0014] Preferably, in step 3, mutual information The calculation formula is: in, To discriminate neural networks, and These represent the product of the joint distribution and the marginal distribution, respectively.

[0015] Preferably, in step 4: define the view pair collection Regularization loss The calculation formula is: Total loss function Loss from basic classification task Mutual information constraints The composition and calculation formula are as follows: in, The parameters used to control the strength of regularization.

[0016] Compared with the prior art, the beneficial effects of this invention are: In this application, a method for identifying the operating conditions of pumping unit wells based on measured multi-source incomplete data significantly improves the accuracy and robustness of identifying operating conditions from incomplete multi-source data through collaborative optimization of the entire chain of encoding, completion, and fusion. Specifically addressing the industry pain point of easily missing and difficult-to-accurately complete multi-source operating condition data of pumping unit wells in oilfields, this method achieves accurate and reliable completion of operating condition data under different missing rates and high-noise scenarios, laying a solid data foundation for subsequent downstream tasks such as oil well operating condition identification, fault early warning, and production optimization.

[0017] In the technical solution of this application, the end-to-end full-link trusted completion and enhancement framework is the core. By integrating the three-stage collaborative optimization strategy of "dual path coding - dynamic trusted propagation - mutual information constraint", the incomplete multi-source data condition identification is realized.

[0018] By constructing a dual-path structure of "private feature branch + shared feature branch" and leveraging the residual injection mechanism, the private and shared semantics of the view are effectively separated, providing unified and transferable features for the completion task and solving the problem of insufficient cross-view semantic alignment.

[0019] A graph structure is constructed based on the feature similarity between samples, and the credible information of high-confidence samples is adaptively propagated to missing view samples, thereby achieving the effective diffusion of global credible information to local uncertain regions.

[0020] During the fusion phase, the category discrimination of the completed features is enhanced by maximizing cross-view shared information and minimizing redundant noise information, while mitigating the fusion bias caused by uneven view quality.

[0021] A total loss function consisting of basic classification task loss and mutual information constraint loss was constructed. Through iterative training, the entire link of feature encoding, missing information completion and information fusion was optimized to achieve identification of incomplete multi-source data in scenarios with high missing rate and noise interference. Attached Figure Description

[0022] Figure 1This is a flowchart of a method for identifying the operating conditions of pumping wells based on incomplete, multi-source measured data.

[0023] Figure 2 This is a schematic diagram of the network structure for a method of identifying the operating conditions of pumping wells based on measured multi-source incomplete data. Detailed Implementation

[0024] Figures 1-2 This is the preferred embodiment of the present invention, which is described below in conjunction with the accompanying drawings. Figures 1-2 The present invention will be further described below.

[0025] like Figure 1 As shown, a method for identifying the operating conditions of pumping wells based on incomplete multi-source measured data includes the following steps: Step 1: Using a dual-path feature extraction structure, features are extracted from incomplete multi-view data to obtain enhanced features, and then evidence vectors are generated.

[0026] Combination Figure 2 A dual-path feature extraction structure of "private feature branch + shared feature branch" is constructed. Then, incomplete multi-view data is input into the dual-path feature extraction structure of "private feature branch + shared feature branch" to obtain the private features and shared semantic features of the incomplete multi-view data respectively. Then, through the residual injection mechanism, enhanced features are generated, and finally, evidence vectors are generated.

[0027] Step 1 includes the following steps: Step 1-1: Decoupling of features of the two paths.

[0028] For the View data Two encoding paths are constructed in parallel: Path 1, using a private feature extractor Extract the first Private features of each view The corresponding formula is: Path 2, Parameter-Shared Feature Extractor Extract the first Each view shares semantic features This maps different views to a unified semantic space, and the corresponding formula is: By employing a dual-path feature extraction structure of "private feature branch + shared feature branch", this decoupling design establishes reliable cross-view semantic associations while preserving view specificity.

[0029] Steps 1-2: Adaptive residual fusion.

[0030] In this application, learnable residual weights are introduced. Dynamically adjust the contribution of shared features to generate enhanced features. The corresponding formula is: in, For learnable residual weights, Using the Sigmoid activation function, for learnable residual weights Apply nonnegativity constraints. The initial state is set to be close to the traditional direct splicing or superposition of two features, which can easily lead to confusion in the feature space and even reduce model performance when the shared information is noisy. By using adaptive residual fusion, the private and shared semantics of the view are effectively separated, providing unified and transferable features for the completion task and solving the problem of insufficient cross-view semantic alignment.

[0031] Steps 1-3: Evidence vector generation.

[0032] To quantify feature uncertainty in subsequent modules, feature enhancement will be implemented. Mapped to the evidence space.

[0033] Specifically, through the evidence vector transformation layer (Composed of fully connected layers and activation functions) Generate evidence vectors The calculation is as shown in the formula below: Among them, evidence vector This will be directly used to parameterize the Dirichlet distribution, providing a basic confidence metric for subsequent dynamic uncertainty propagation.

[0034] Step 2: Construct a graph structure based on the feature similarity between samples using the evidence vectors, and adaptively propagate the credible information of high-confidence samples to the missing view samples.

[0035] Step 2 includes the following steps: Step 2-1, View uncertainty quantification.

[0036] For the Sample of a view Its evidence vector Parameterized Dirichlet distribution, with distribution parameters as follows: The uncertainty of this view The calculation is as follows: in, Indicates the number of categories.

[0037] Corresponding confidence level The calculation formula is as follows: When the model is able to extract effective features from the view, the uncertainty of the view is reduced. Approaching step 2-2, dynamic graph construction.

[0038] To establish information transmission channels between samples, a dynamic adjacency matrix based on feature space is constructed. ,in This is the batch size. For any two samples within the batch... and The formula for calculating connection weights is: in, For cosine similarity, This is the temperature coefficient. They represent the first The i-th and j-th samples of the view.

[0039] To suppress noisy connections and reduce computational complexity, only top- For the nearest neighbor sample, the connection weights of the remaining nodes are reset to 0.

[0040] Steps 2-3: Confidence propagation and reweighting.

[0041] Design an iterative graph propagation mechanism to correct local uncertainty estimation. Let the initial confidence matrix be... In the t-th iteration ( The confidence update formula is: in, It is a balancing factor.

[0042] By dynamically correcting uncertainty and constructing a graph structure based on the feature similarity between samples, the credible information of high-confidence samples is adaptively propagated to missing view samples, thus achieving effective diffusion of global credible information to local uncertain regions.

[0043] The weighting of historical confidence and neighborhood propagation information is controlled. After T rounds of propagation, the smoothed uncertainty is obtained as follows: .

[0044] Steps 2-4: Global evidence fusion.

[0045] Calculating global confidence using multi-view information The original evidence is then reweighted, including the following steps: Step 2-4-1, using learnable parameters Calculate the fusion weights of each view The calculation formula is: in, They represent the first Learnable parameters for the j-th and j-th views.

[0046] Step 2-4-2, calculate the global confidence score. The calculation formula is: Step 2-4-3: Adaptively reweight the original evidence using global confidence, calculated as follows: in, Indicates the global confidence level. This represents the evidence vector.

[0047] Step 3: Estimate the lower bound of mutual information using the mutual information neural estimation method; A mutual information neural estimation method is used to estimate the lower bound of mutual information. During the fusion stage, the class discriminative power of the completed features is enhanced by maximizing cross-view shared mutual information and minimizing redundant noise information, while mitigating the fusion bias caused by uneven view quality. Specifically: According to the Donsker-Varadhan variational representation, mutual information The calculation formula is: in, To discriminate neural networks, They represent the joint distribution, Marginal distribution product. Discriminant network. Training is performed by distinguishing between jointly distributed samples and marginally distributed samples. During training, optimization is achieved using a maximization formula. To obtain a tight lower bound for mutual information.

[0048] Step 4: Construct a total loss function consisting of the basic classification task loss and the mutual information constraint loss to achieve classification of incomplete multi-view data.

[0049] Constructing the loss from the basic classification task and mutual information constraint loss The total loss function is composed of features that are optimized through iterative training, and the entire process of feature encoding, missing data completion, and information fusion is optimized to achieve accurate classification of incomplete multi-view data in scenarios with high missing rates and noise interference.

[0050] Step 4 specifically includes the following steps: To impose consistency constraints across all view pairs, define a set of view pairs. Regularization loss The calculation formula is: Total loss function Loss from basic classification task Mutual information constraints The composition and calculation formula are as follows: in, The parameters used to control the strength of regularization.

[0051] The technical solution of this application will be verified through an example below.

[0052] The datasets used include multiple incomplete multi-view benchmark datasets such as BRAC, Caltech101-20, Scene15, ROSMAP, UCI, and GK, which comprehensively cover multiple fields such as biomedicine, image recognition, and oil well condition monitoring, and can fully verify the performance of the method in different scenarios.

[0053] To verify the classification ability, the technical solution of this application was tested on the aforementioned six datasets. The test scenarios included missing rates ranging from 10% to 70%, with adjacent missing rate intervals of 10%. The results show that, in the six datasets, compared with existing mainstream methods, the classification performance of this invention is superior under different missing rates. Experimental results demonstrate that the proposed three-stage collaborative design of residual-shared semantic atomic encoding, dynamic uncertainty correction, and mutual information regularization can effectively compensate for the performance shortcomings of existing methods in cross-view alignment, completion reliability, and fusion efficiency. By optimizing feature representation quality, reliable information propagation efficiency, and fusion feature discriminability, the proposed method achieves superior performance in incomplete multi-view classification tasks with high missing rates and noise interference.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A pumping unit well operating condition identification method based on measured multi-source incomplete data, characterized by: Includes the following steps: Step 1: Construct a dual-path feature extraction structure, then input the incomplete multi-view data into the dual-path feature extraction structure to obtain the private features and shared semantic features of the incomplete multi-view data respectively. Use the residual injection mechanism to extract features from the multi-source data respectively, and finally generate the evidence vector. Step 2: Based on the component graph structure of feature similarity between samples, the credible information of high-confidence samples is adaptively propagated to the missing view samples, realizing the diffusion of global credible information to local uncertain regions; Step 3: In the fusion stage, by maximizing cross-view shared information and mutual information and minimizing redundant noise information, the class discriminativeness of the completed features is enhanced, while mitigating the fusion bias caused by uneven view quality. Step 4: Construct a total loss function consisting of the basic classification task loss and the mutual information constraint loss to achieve incomplete multi-source data condition identification; Step 2 includes the following steps: Step 2-1, for the first The evidence vector of sample i in view i Parameterized Dirichlet distribution, with distribution parameters as follows: The uncertainty of this view The calculation is as follows: Where C represents the number of categories; Corresponding confidence level The calculation formula is: ； Step 2-2: Construct a dynamic adjacency matrix based on feature space. Where B is the batch size; for any two samples within the batch and The formula for calculating connection weights is: in, The cosine similarity is T, where T is the temperature coefficient. , They represent the first The i-th and j-th samples of the view; Steps 2-3: Confidence propagation and reweighting; Design an iterative graph propagation mechanism to correct local uncertainty estimation; initialize the confidence matrix. In the t-th iteration The confidence level update formula is: in, It is a balancing factor; Steps 2-4: Calculate global confidence using multi-view information. And the original evidence is reweighted; In step 3, mutual information The calculation formula is: in, To discriminate neural networks, and These represent the product of the joint distribution and the marginal distribution, respectively.

2. The method for identifying pumping unit well conditions based on measured multi-source incomplete data according to claim 1, characterized in that: Step 1 includes the following steps: Step 1-1, for the first View data Using a private feature extractor Extract the first Private features of each view The corresponding formula is: Simultaneously, a parameter-sharing feature extractor is used. Extract the first Each view shares semantic features This maps different views to a unified semantic space, and the corresponding formula is: ； Step 1-2: Introduce learnable residual weights Dynamically adjust the contribution of shared features to generate enhanced features. The corresponding formula is: in, For learnable residual weights, Use the Sigmoid activation function; Steps 1-3, through the evidence vector transformation layer Generate evidence vectors The calculation is as shown in the formula below: in, For evidence vectors, To enhance features.

3. The method for identifying pumping unit well conditions based on measured multi-source incomplete data according to claim 1, characterized in that: Steps 2-4 include the following steps: Step 2-4-1, using learnable parameters Calculate the fusion weights of each view The calculation formula is: in, , They represent the first Learnable parameters for the j-th and j-th views; Step 2-4-2, calculate the global confidence score. The calculation formula is: Step 2-4-3: Adaptively reweight the original evidence using global confidence, calculated as follows: in, Indicates the global confidence level. This represents the evidence vector.

4. The method for identifying the operating conditions of pumping wells based on measured multi-source incomplete data according to claim 1, characterized in that: In step 4: Define the view collection Regularization loss The calculation formula is: Total loss function Loss from basic classification task Mutual information constraints The composition and calculation formula are as follows: in, The parameters used to control the strength of regularization.

Images