Pumping unit fault diagnosis method based on multi-modal fusion under heavy oil working condition

Abstract

The invention discloses an oil pumping unit fault diagnosis method based on multi-modal fusion under a heavy oil working condition, and the method comprises the following steps: collecting operation parameter data of an oil pumping unit, including an indicator diagram, an electric indicator diagram, and structured data such as stroke, stroke frequency, water content, working fluid level and the like, and carrying out the data cleaning; a fusion strategy based on multi-modal interaction is designed, and key features are extracted and subjected to weighted fusion through a cross-modal attention distribution matrix; a dynamic collaborative weight distribution mechanism is introduced, feature integration and complementarity of multi-source data are optimized, and the model constructed by the invention can accurately diagnose the fault type of the oil pumping unit under a complex working condition, improve the diagnosis efficiency and reduce manual intervention. The method has high practical value and economic benefit, and is especially suitable for equipment fault diagnosis scenes under high challenging working conditions such as heavy oil recovery.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and particularly relates to a fault diagnosis method for pumping units based on multi-modal fusion under heavy oil working conditions, which can be used for guiding the oil production industry. Background Art

[0002] Heavy oil extraction is an important and challenging process in the petroleum industry. Due to the high viscosity and poor fluidity of heavy oil, special production processes such as steam injection and heating are usually required to effectively increase the oil well production. During the heavy oil extraction process, the operating load of the pumping unit increases significantly, and the equipment is easily affected by the oil product characteristics and process conditions, resulting in increased mechanical wear, decreased pump efficiency and other problems. In addition, under heavy oil working conditions, the pumping unit often operates in a high-temperature and high-pressure environment, which further increases the probability of failure and the complexity of diagnosis.

[0003] Although existing expert systems and deep learning methods have been able to identify various pumping unit faults, most of these methods only rely on a single information source, namely the dynamometer card, which increases the risk of misjudgment during the diagnosis process. Especially when facing heavy oil working conditions, the fluctuations of equipment operating parameters are more complex, and a single information source is difficult to accurately reflect the overall state of the equipment. Therefore, it is difficult to effectively diagnose major and rare faults of the pumping unit.

[0004] Therefore, there is an urgent need for a fault diagnosis method for pumping units based on multi-modal fusion under heavy oil working conditions to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and propose a fault diagnosis method for pumping units based on multi-modal fusion under heavy oil working conditions, so as to improve the accuracy and efficiency of fault diagnosis, reduce the dependence on a single data source, and realize the automatic and efficient identification of the working conditions of the pumping unit through deep learning technology, thereby reducing manual intervention and improving the safety and economic benefits of oilfield operation.

[0006] The present invention proposes a fault diagnosis method for pumping units based on multi-modal fusion under heavy oil working conditions, and the specific steps are as follows:

[0007] 1. A fault diagnosis method for pumping units based on multi-modal fusion under heavy oil working conditions, characterized by comprising the following steps:

[0008] S1. Obtain the operation parameter data of the pumping units in different work areas, and select engineering parameters related to the working conditions of the pumping units according to expert knowledge. The engineering parameters include: dynamometer card, electric power card, and four structured data of stroke, pumping speed, water cut and dynamic liquid level;

[0009] S2. Design a multimodal fusion method. This method generates a cross-modal attention distribution by calculating the interaction matrix of two modalities, namely the indicator diagram and the electric work diagram, and uses the generated attention distribution to perform weighted fusion on the original feature maps, highlighting the key features of the modalities and enhancing the discrimination ability and adaptability of the fused features.

[0010] S3. Design a multi-source data collaborative optimization method. Through a dynamic collaborative weight allocation mechanism, it integrates the feature representations of the indicator diagram, the electric work diagram, and four types of structured data, namely the stroke, the pumping frequency, the water content, and the fluid level. The model utilizes the feature complementarity to significantly improve the discrimination ability and prediction accuracy for complex working conditions.

[0011] As the pumping unit fault diagnosis method based on multimodal fusion under heavy oil working conditions of the present invention, wherein in the step S2, the cross-modal attention distribution is specifically designed as follows:

[0012] First, perform global average pooling operations on the indicator diagram and the electric work diagram respectively, specifically:

[0013]

[0014] Among them, and respectively represent the feature information of the indicator diagram and the electric work diagram after global average pooling; A(H, W, c) and B(H, W, c) respectively represent the feature maps of the indicator diagram and the electric work diagram obtained after being processed by the convolutional layer; H and W respectively represent the height and width of the feature map; c represents the number of channels of the feature map.

[0015] Finally, compress the global information representations of these two modalities from the spatial domain to the channel domain, calculate the interaction relationship between the two modalities of the indicator diagram and the electric work diagram, and then apply the Softmax operation to the interaction matrix to generate a cross-modal attention distribution matrix, specifically:

[0016]

[0017] Among them, Softmax(·) represents the Softmax operation; Glatten(·) represents the compression and flattening operation; represents the outer product operation; A represents the cross-modal attention distribution matrix.

[0018] As the pumping unit fault diagnosis method based on multimodal fusion under heavy oil working conditions of the present invention, wherein in the step S3, the multi-source data collaborative optimization method adopts a dynamic collaborative weight allocation mechanism, specifically:

[0019]

[0020] Among them, GAP(·) represents the global average pooling operation; Norm(·) represents the weight normalization function; F represents the feature after weighted fusion; F 1 and F 2 respectively represent the feature vectors extracted from the image data and the structured data; w 1 and w 2 respectively represent the learnable weights corresponding to the feature vectors of the image data and the structured data, which are used to ensure the numerical stability of the weights, and these weights are adjusted according to their importance in fault diagnosis.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1) Combining multiple data sources enhances the robustness and accuracy of the model, making fault diagnosis more precise and reliable;

[0023] 2) The multi-modal deep learning framework can process a large amount of complex data, optimize the processing flow, reduce the dependence on professional knowledge, and simplify the operation process;

[0024] 3) The design of the model reduces human errors, improves the diagnosis speed and efficiency, and helps to quickly respond to and handle possible operation problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the implementation flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0026] The following further describes the present invention in conjunction with the drawings and implementation cases, so as to make the purpose, technical solution and advantages of the present invention clearer. Taking Figure 1 as an example, the process of the present invention is to collect data, clean and plot the data, enhance and identify the features of the image data, train the model, and diagnose the working conditions, which specifically include the following steps:

[0027] , S1. Obtain the operation parameter data of the pumping units in different work areas, and select the engineering parameters related to the working conditions of the pumping units according to expert knowledge. The engineering parameters include: the indicator diagram, the electric power diagram, and four structured data of the stroke, the pumping frequency, the water cut, and the dynamic liquid level;

[0028] Obtain the operation parameter data of the pumping units in different work areas, and preferably select the engineering parameters related to the working conditions of the pumping units according to expert knowledge, including the indicator diagram, the electric power diagram, and four structured data of the stroke, the pumping frequency, the water cut, and the dynamic liquid level. Some parameter statistics are shown in Table 1:

[0029] Some parameter statistics are shown in Table 1:

[0030] Table 1, Some Parameter Statistics Table

[0031]

[0032] S2. Design a multimodal fusion method. This method calculates the interaction matrix of two modalities, namely the indicator diagram and the electric work diagram, generates a cross-modal attention distribution, and uses the generated attention distribution to perform weighted fusion on the original feature maps, highlighting the key features of the modalities and enhancing the discrimination ability and adaptability of the fused features;

[0033] First, perform global average pooling operations on the indicator diagram and the electric work diagram respectively, specifically:

[0034]

[0035] Among them, and respectively represent the feature information of the indicator diagram and the electric work diagram after global average pooling; A(H, W, c) and B(H, W, c) respectively represent the feature maps of the indicator diagram and the electric work diagram obtained after being processed by the convolutional layer; H and W respectively represent the height and width of the feature map; c represents the number of channels of the feature map;

[0036] Finally, compress the global information representations of these two modalities from the spatial domain to the channel domain, calculate the interaction relationship between the two modalities of the indicator diagram and the electric work diagram, and then apply the Softmax operation to the interaction matrix to generate a cross-modal attention distribution matrix, specifically:

[0037]

[0038] Among them, Softmax(·) represents the Softmax operation; Flatten(·) represents the compression and flattening operation; represents the outer product operation; A represents the cross-modal attention distribution matrix.

[0039] S3. Design a multi-source data collaborative optimization method. Through a dynamic collaborative weight allocation mechanism, integrate the feature representations of the indicator diagram, the electric work diagram, and four types of structured data, namely the stroke, the pumping speed, the water content, and the fluid level. The model utilizes the feature complementarity to significantly improve the discrimination ability and prediction accuracy for complex working conditions categories;

[0040] The multi-source data collaborative optimization method adopts a dynamic collaborative weight allocation mechanism, specifically:

[0041]

[0042] Among them, GAP(·) represents the global average pooling operation; Norm(·) represents the weight normalization function; F represents the weighted fused feature; F 1 and F 2 respectively represent the feature vectors extracted from the image data and the structured data; w 1and w 2 respectively represent the learnable weights corresponding to the feature vectors of the image data and the structured data, which are used to ensure the numerical stability of the weights, and these weights are adjusted according to their respective importance in fault diagnosis.

[0043] As mentioned above, it is not a restriction on the present invention in any form. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can, within the scope of the technical solution of the present invention, make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as the content does not depart from the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for fault diagnosis of an oil pumping unit based on multi-modal fusion under heavy oil conditions, characterized in that: The following steps are involved: S1. Obtaining the operating parameter data of the oil pumping units in different working areas, and selecting the engineering parameters related to the working conditions of the oil pumping units according to the expert knowledge, wherein the engineering parameters include: the dynamometer diagram, the electric dynamometer diagram, and four structured data of stroke, stroke frequency, water content and dynamic liquid level; S2. Design a multimodal fusion method. This method generates cross-modal attention distribution by calculating the interaction matrix of the two modalities of the dynamometer diagram and the electric dynamometer diagram, and uses the generated attention distribution to perform weighted fusion on the original feature map to highlight the key features of the modality and enhance the distinguishing ability and adaptability of the fused features. S3. Design a multi-source data collaborative optimization method, which integrates the feature representation of the dynamometer diagram, electric dynamometer diagram, and four types of structured data, namely stroke, stroke frequency, water content, and dynamic liquid level, through a dynamic collaborative weight allocation mechanism. The model utilizes feature complementarity to significantly improve the recognition ability and prediction accuracy of complex working conditions.

2. The method for diagnosing oil pumping unit faults based on multi-modal fusion under heavy oil conditions according to claim 1, characterized in that: In step S2, the cross-modal attention distribution is specifically designed as follows: First, the dynamometer diagram and the electric dynamometer diagram are subjected to global average pooling operations, specifically: in, and They represent the characteristic information of the dynamometer diagram and the electric dynamometer diagram after global average pooling, respectively; A(H, W, c) and B(W, W, c) represent the characteristic maps of the dynamometer diagram and the electric dynamometer diagram after processing by the convolution layer, respectively; H and W represent the height and width of the characteristic map, respectively; c represents the number of channels of the characteristic map; Finally, the global information representation of the two modalities is compressed from the spatial domain to the channel domain, and the interaction relationship between the two modalities of the dynamometer diagram and the electric dynamometer diagram is calculated. Then, the Softmax operation is applied to the interaction matrix to generate the cross-modal attention distribution matrix, which is: Among them, Softmax(·) represents the Softmax operation; Flatten(·) represents the compression and flattening operation; represents the outer product operation; A represents the cross-modal attention distribution matrix.

3. The method for diagnosing oil pumping unit faults based on multi-modal fusion under heavy oil conditions according to claim 1, characterized in that: In step S3, the multi-source data collaborative optimization method adopts a dynamic collaborative weight allocation mechanism, specifically: Among them, GAP(·) represents the global average pooling operation; Norm(·) represents the weight normalization function; F represents the weighted fused feature; F1 and F2 represent the feature vectors extracted from image data and structured data, respectively; w1 and w2 represent the learnable weights corresponding to the feature vectors of image data and structured data, respectively, which are used to ensure the numerical stability of the weights. These weights are adjusted according to their respective importance in fault diagnosis.

Images