Multi-source semi-supervised incremental working condition identification method and system for rod-pumped well

Abstract

The invention discloses a multi-source semi-supervised incremental working condition identification method and system for a rod-pumped well, and relates to the technical field of working condition identification of rod-pumped wells. The method is characterized by comprising the following steps: step 1, respectively storing a sample library of an actually measured ground indicator diagram and an actually measured electric indicator diagram; 2, respectively constructing a graph neural network teacher model for two data sources, namely an actually measured ground indicator diagram and an actually measured electric indicator diagram; 3, dynamically fusing the prediction probability of each teacher model; 4, multi-source data distillation learning is carried out; and 5, performing semi-supervised working condition identification by using a label propagation algorithm improved by a logistic regression classifier. According to the multi-source semi-supervised class increment working condition identification method and system for the rod-pumped well, a multi-source fusion technology based on an attention mechanism, class increment learning and semi-supervised learning are simultaneously applied to class increment working condition identification of the rod-pumped well, and a small number of multi-source marking working condition samples are fully utilized; and more efficient, robust and practical rod-pumped well working condition identification is realized by combining a large number of multi-source unknown working condition samples.

Description

Technical Field

[0001] The present invention relates to the technical field of working condition identification of pumping wells, and particularly to a multi-source semi-supervised incremental working condition identification method and system for pumping wells. Background Art

[0002] In the field of oilfield operations, the real-time monitoring and identification of the working conditions of pumping wells are crucial for preventing equipment failures, optimizing production scheduling, and improving resource utilization. The existing methods for identifying the working conditions of pumping wells mainly include the following several types:

[0003] 1. Identification based on the dynamometer card. The identification method based on the dynamometer card mainly relies on the pump dynamometer card (the dynamometer card obtained by model calculation) or the measured ground dynamometer card, and combines artificial intelligence methods for working condition identification. For example, the technical solutions described in the Chinese invention patent with the application number 202411505865.3, the application date of October 28, 2024, and the patent name of "An Intelligent Analysis Method and System for the Working Conditions of a Pumping Unit". The technical solutions described in the Chinese invention patent with the application number 202310347518.1, the application date of April 3, 2023, and the patent name of "A Training Method for a Working Condition Identification Model, a Working Condition Diagnosis Method, and a Device".

[0004] 2. Identification based on electrical parameters. The working condition identification based on the electric power card belongs to the category of working condition identification based on electrical parameters. The working condition identification based on the electric power card includes the following two types: (1) Model-converted electric power card. (2) Measured electric power card, such as the technical solutions described in the Chinese invention patent with the application number 202311441260.8, the application date of December 1, 2023, and the patent name of "A Method and System for Identifying the Working Conditions of a Pumping Well with a Measured Electric Power Card".

[0005] 3. Identification based on multi-source information. The working condition identification based on multi-source information mainly combines multiple information sources such as the dynamometer card and oil well production information (such as well conditions data, production volume, pumping parameters, etc.) for working condition identification. For example, the technical solutions described in the Chinese invention patent with the application number 202311284299.3, the application date of September 30, 2023, and the patent name of "Fault Diagnosis Method, Computing Device, and Readable Storage Medium".

[0006] The above-mentioned working condition identification methods generally have the following limitations: (1) Single information source identification limitation: In a complex nonlinear system of electromechanical and hydraulic coupling, relying on a single information source to judge the working condition of the pumping well is prone to false alarms; (2) Feature parameter accuracy problem: The calculation of the pump dynamometer diagram and the electric dynamometer diagram may be subject to the damping coefficient and the "division by zero" problem, which will affect the accuracy of the characteristic parameter value calculation; (3) Multi-source data fusion technology limitation: Due to the technical limitations of traditional multi-feature connection identification methods, the complexity and variability of well conditions, the uncertainty of manual statistical production data and other factors, the existing working condition identification methods based on multi-source data are in urgent need of improvement in terms of identification effect and model robustness; (4) Difficulty in obtaining labeled samples: Most working condition identification methods rely on a large number of labeled training samples. In actual engineering, it is difficult and expensive to obtain labeled samples. At the same time, the methods of unlabeled sample training often have poor recognition accuracy; (5) With the continuous progress of oilfield operations, the categories of pumping well working conditions continue to expand, gradually evolving into a typical quasi-incremental learning scenario. In this scenario, the emergence of new working condition categories requires the model to not only adapt to the introduction of new categories, but also maintain the ability to identify old working condition categories. This poses new challenges to traditional working condition identification methods.

[0007] In the context of big data oil production, the production system of pumping wells can obtain massive working condition information including ground dynamometer diagrams, electrical parameters and other measured information sources, and also obtain a large amount of multi-source unknown working condition information. Faced with the quasi-incremental actual production operation scenarios caused by complex and changeable working conditions, it is necessary to explore how to effectively utilize the various measured information sources of pumping wells, and to achieve more efficient, robust and practical working condition identification by fully combining a large number of multi-source unmarked working condition samples with a small number of multi-source labeled working condition samples, so as to provide scientific, timely and accurate guidance for oil production decisions. This has become a key issue to be solved in the production, construction and development of intelligent oilfields. Therefore, the development of a multi-source semi-supervised quasi-incremental working condition identification method and system for pumping wells has important theoretical and application value for effectively solving the above key problems and promoting the production, construction and development of intelligent oilfields. Summary of the invention

[0008] The technical problem to be solved by the present invention is: to overcome the shortcomings of the prior art and provide a method and system for multi-source semi-supervised incremental condition identification of pumping wells, which simultaneously applies multi-source fusion technology based on attention mechanism, incremental learning and semi-supervised learning to the incremental condition identification of pumping wells, makes full use of a small number of multi-source labeled condition samples, and combines a large number of multi-source unknown condition samples to achieve more efficient, robust and practical pumping well condition identification.

[0009] The technical solution adopted by the present invention to solve the technical problem is: the multi-source semi-supervised incremental working condition identification method of the pumping well is characterized by comprising the following steps:

[0010] Step 1, respectively store the sample libraries containing two data sources of the measured surface dynamometer cards and the measured electric power cards corresponding to the marked and unmarked pumping well conditions;

[0011] Step 2, respectively construct graph neural network teacher models for the two data sources of the measured surface dynamometer cards and the measured electric power cards;

[0012] Step 3, use the Squeeze-and-Excitation attention mechanism to dynamically fuse the prediction probabilities of each teacher model;

[0013] Step 4, use the Kullback-Leibler divergence for multi-source data distillation learning;

[0014] Step 5, use the label propagation algorithm improved by a logistic regression classifier for semi-supervised condition identification.

[0015] Preferably, Step 3 specifically includes the following steps:

[0016] Step 3-1, refer to the Squeeze-and-Excitation attention mechanism to implement a multi-source feature representation fusion module, which is composed of a fully connected layer - ReLU activation layer - fully connected layer - Sigmoid activation layer. The prediction probability output by the teacher model of the v-th data source branch is expressed as Its corresponding attention score ω v Is expressed as:

[0017]

[0018] Among them, sigmoid(·) is the Sigmoid activation function, relu(·) is the ReLU activation function, and W1 and W2 are the weights of the two fully connected layers respectively;

[0019] Step 3-2, weight the obtained attention weights to the prediction results output by the multi-source teacher model to obtain the prediction results p of the fused multi-source teacher model t Is expressed as:

[0020]

[0021] Among them, p t Represents the prediction result of the fused multi-source teacher model, including the prediction result p t (L t ) and the prediction result p of unlabeled data t (U t ).

[0022] Preferably, Step 4 specifically includes the following steps:

[0023] Step 4-1, when learning task t, usually model G t is used as the student model, and model G of task t-1 t-1 is used as model G t 's teacher model. The distillation loss L KL between the output prediction probabilities of the teacher model and the student model is calculated by KL divergence:

[0024]

[0025] where T t-1 is the number of learned categories after task t-1, represents the predicted probability output by the multi-source teacher model of the (t-1)-th task, represents the predicted probability output by the student model of the t-th task;

[0026] Step 4-2, the total loss function L t of task t consists of the cross-entropy loss function L c and the multi-source data distillation loss L KL . The cross-entropy loss function L c is expressed as:

[0027]

[0028] where c t represents the number of sample categories of task t, p t (L t ) represents the model prediction result after the fusion of the feature representations of the labeled samples in task t, represents the true label of the i-th labeled sample in task t;

[0029] Step 4-3, the calculation formula of the total loss function L t is:

[0030] L t =L c +λL KL

[0031] where L c represents the cross-entropy loss function, L KL is the distillation loss between the output prediction probabilities of the teacher model and the student model, and λ is the weight of the multi-source data distillation loss function.

[0032] Preferably, step 5 specifically includes the following steps:

[0033] Step 5-1, perform mean fusion on the features output by the graph convolutional networks of different data sources . The feature matrix X t of task t after fusion is calculated as:

[0034]

[0035] Among them, X t contains the global feature matrix X of labeled data L and the global feature matrix X of unlabeled data U , where: X t = [X L , X U ;

[0036] Step 5-2, the multi-source semi-supervised graph classification method based on enhanced label propagation is expressed as:

[0037]

[0038] Among them, is the adjacency matrix constructed using the fused features, contains the adjacency matrix constructed from labeled data and the adjacency matrix constructed from unlabeled data The calculation formula is:

[0039]

[0040] Among them, represents the probability that the k-th sample belongs to class c, The calculation formula is:

[0041]

[0042] Among them, y k represents the label of the k-th sample;

[0043] Step 5-3, use a logistic regression classifier to generate an initial prediction probability distribution for unlabeled data, provide a more accurate initial label vector for label propagation, after completing the label propagation process, use multiple logistic regression classifiers to make P LR classify the label embedding vectors and calculate the average of the results of these classifiers to obtain the final classification label. The final label y of task t t The calculation formula is:

[0044]

[0045] Among them, is the label vector of task t after label propagation.

[0046] Preferably, step 2 specifically includes the following steps:

[0047] Step 2-1: For each sample in a single data source, calculate the distances between it and all other samples in its dataset, and select the two samples with the smallest distances as its nearest neighbor samples to construct the initial adjacency matrix of the graph convolutional network. The calculation formula of the graph convolutional network for the t-th task is as follows:

[0048]

[0049] where is the adjacency matrix constructed for the v-th data source in task t, is the data of the v-th data source in task t, including the labeled data and unlabeled data of the v-th data source;

[0050] Step 2-2: Use the convolutional layer of the graph convolutional network to extract node features:

[0051]

[0052] where I K is the identity matrix, relu(·) is the ReLU activation function, W is the trainable parameter matrix, is a diagonal matrix, and the elements on the diagonal are: is the feature output by the l-th layer of the graph convolutional network;

[0053] Step 2-3: The predicted probability

[0054]

[0055] output by the teacher model of the v-th data source branch, where FC is a fully connected layer used to classify the features output by the graph convolutional network, is the v-th in task t

[0056]

[0057] Preferably, in Step 1, the measured surface dynamometer card is a binary image composed of polished rod displacement and polished rod load, and the measured electric dynamometer card is a binary image composed of polished rod displacement and motor active power.

[0058] A multi-source semi-supervised class incremental working condition recognition system for pumping wells, characterized by comprising a working condition sample library storage module, a working condition recognition model construction module, and a working condition recognition module, wherein the working condition sample library storage module further includes a measured surface dynamometer card sample library and a measured electric dynamometer card sample library;

[0059] The measured surface dynamogram sample library and the measured electric work diagram sample library store the measured surface dynamograms and the measured electric work diagrams respectively, aiming to collect and manage the pumping well condition samples used for training and to be recognized in the system; the measured surface dynamogram sample library and the measured electric work diagram sample library are respectively connected to the condition recognition model construction module;

[0060] The condition recognition model construction module includes a multi-source semi-supervised class incremental condition recognition model for pumping wells, and its output end is connected to the input end of the condition recognition module;

[0061] The condition recognition module is used to classify and recognize the pumping well conditions to be recognized through the condition recognition model.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] This multi-source semi-supervised class incremental condition recognition method and system for pumping wells apply the multi-source fusion technology based on the attention mechanism, class incremental learning and semi-supervised learning to the class incremental condition recognition of pumping wells at the same time, make full use of a small number of multi-source labeled condition samples, and combine a large number of multi-source unknown condition samples to achieve more efficient, robust and practical pumping well condition recognition.

[0064] In the prior art, the Squeeze-and-Excitation attention mechanism was initially proposed for feature map fusion inside the model. It compresses the feature map through global pooling operations, and then learns the importance weights of each channel through a fully connected layer, so as to achieve dynamic weighting between the internal channels of the feature map, mainly used to enhance the feature representation ability inside a single deep learning model, thereby improving the classification accuracy. And in the field of multi-source data fusion, the common method is to fuse the features from different data sources by simple weighted summation, concatenation or fixed weight assignment. Although these methods can combine multi-source data, they cannot dynamically adjust the fusion weights according to the importance and relevance of the data sources, ignoring the importance differences of different modality data in different scenarios. In this multi-source semi-supervised class incremental condition recognition method and system for pumping wells, two data sources of measured surface dynamograms and measured electric work diagrams are adopted, the Squeeze-and-Excitation attention mechanism is applied to multi-source feature fusion, the SE mechanism is used to dynamically assign weights to the features of the measured surface dynamograms and the measured electric work diagrams, and the fusion ratio of the features of different data sources is dynamically adjusted according to these weights. This dynamic weight assignment method can better capture the complementary information between different data sources and enhance the adaptability of the model to multi-source data.

[0065] The multi-source semi-supervised class-incremental working condition recognition method and system for pumping wells uses a multi-source data distillation method based on Kullback-Leibler divergence to measure the difference between the output logics of the teacher and student models, helping the student model maintain memory of old tasks while learning new tasks, reducing the forgetting of old working condition categories during the class-incremental learning process, and improving the accuracy of class-incremental working condition recognition.

[0066] In the existing technologies, most patents on graph neural networks, distillation learning, and attention mechanisms apply distillation learning to ordinary classification tasks. Its main purpose is to improve the classification accuracy of the student model through distillation learning while reducing model parameters to achieve more efficient and lightweight model deployment. The working condition categories are fixed during the model training process. However, in this application, it is aimed at the special scenario of class-incremental classification tasks. During the class-incremental learning process, as new working condition categories continuously appear, the model needs to maintain the recognition ability of old categories while learning new categories. The present invention introduces a multi-source data distillation loss based on Kullback-Leibler divergence to measure the difference between the output logics of the teacher and student models, helping the student model maintain memory of old tasks while learning new tasks, and reducing the forgetting of old working condition categories during the class-incremental learning process. This application method of distillation learning is essentially different from the distillation learning for ordinary classification tasks in the existing technologies. It pays more attention to solving the problem of knowledge forgetting in class-incremental learning rather than simply improving classification accuracy or reducing model parameters.

[0067] The multi-source semi-supervised class-incremental working condition recognition method and system for pumping wells can break through the limitations of single information source prone to false alarms and the bottlenecks of traditional multi-source fusion technologies to further improve the working condition recognition effect, enhance the robustness and engineering practicability of the working condition recognition model in scenarios where the working condition types are complex and changeable and the training data is scarce, and has important theoretical, practical application value, and economic significance for improving the oil well fault prevention ability, optimizing oil production scheduling, increasing the oil well recovery rate, and promoting the construction and development of intelligent oilfields. Description of the Drawings

[0068] Figure 1 It is a flow chart of the multi-source semi-supervised class-incremental working condition recognition method for pumping wells.

[0069] Figure 2 It is a schematic diagram of the multi-source semi-supervised class-incremental working condition recognition method for pumping wells.

[0070] Figure 3 It is a block diagram of the principle of the multi-source semi-supervised class-incremental working condition recognition system for pumping wells. Detailed Embodiments

[0071] Figures 1 to 3 This is the best embodiment of the present invention. The following combines the attached Figures 1 to 3Further description of the present invention.

[0072] As Figures 1 to 2 shown, a multi-source semi-supervised class incremental working condition recognition method for pumping wells includes the following steps:

[0073] Step 1, respectively store sample libraries of two data sources, namely the measured surface dynamometer card and the measured electric power card corresponding to the marked and unmarked working conditions of the pumping well;

[0074] The measured surface dynamometer card and the measured electric power card data corresponding to the working conditions of the pumping well are respectively stored in the corresponding sample libraries. Each sample in the sample library is composed of data points actually collected by the production site under different working conditions, ensuring the authenticity and representativeness of the samples. These data points record in detail the operating state of the pumping well at a specific moment. The samples in each sample library are divided into marked and unmarked pumping well working condition samples. Among them, the marked samples are strictly selected according to the operation records of the oil wells, ensuring the accuracy and reliability of the samples.

[0075] Among them, the measured surface dynamometer card is a binary image composed of polished rod displacement and polished rod load, and the measured electric power card is a binary image composed of polished rod displacement and motor active power. These sample libraries, as the basis for data storage and processing, provide necessary data support for subsequent working condition recognition.

[0076] Step 2, respectively construct graph neural network teacher models for the two data sources of the measured surface dynamometer card and the measured electric power card;

[0077] For the original data of each information source, use the histogram of oriented gradients to extract the local texture feature information of the data, and then use the nearest neighbor method to construct a graph structure for each data source respectively. Specifically, it includes the following steps:

[0078] Step 2-1, for each sample in a single data source, calculate the distance between it and all other samples in its dataset, and select the two samples with the smallest distance as its nearest neighbor samples to construct the initial adjacency matrix of the graph convolutional network. The calculation formula of the graph convolutional network for the t-th task is:

[0079]

[0080] Among them, The adjacency matrix constructed for the v-th data source in task t, The data of the v-th data source in task t, including the labeled data and unlabeled data of the v-th data source.

[0081] Step 2-2, use the convolutional layer of the graph convolutional network to extract node features, and the calculation process is as the formula:

[0082]

[0083] Among them, I K is the identity matrix, relu(·) is the ReLU activation function, W is the trainable parameter matrix, is a diagonal matrix, and the elements on the diagonal are: is the feature output by the l-th layer graph convolutional network.

[0084] Step 2-3, the predicted probability output by the teacher model of the v-th data source branch. The calculation process is as follows:

[0085]

[0086] Among them, FC is the fully connected layer, which is used to classify the features output by the graph convolutional network. is the data of the v-th data source in task t, is the feature output by the (l + 1)-th layer graph convolutional network.

[0087] Step 3, use the Squeeze-and-Excitation attention mechanism to dynamically fuse the predicted probabilities of each teacher model;

[0088] Use the attention mechanism to dynamically learn the weights of the predicted probabilities of the teacher models of different data sources. By assigning different weights to the predicted results of the teacher models of different data sources, the effective fusion of the multi-source teacher models is achieved. The specific steps are as follows:

[0089] Step 3-1, refer to the Squeeze-and-Excitation attention mechanism to implement the multi-source feature representation fusion module, which is composed of a fully connected layer - ReLU activation layer - fully connected layer - Sigmoid activation layer. The predicted probability output by the teacher model of the v-th data source branch is denoted as p t v , and its corresponding attention score ω v is expressed as:

[0090]

[0091] Among them, sigmoid(·) is the Sigmoid activation function, relu(·) is the ReLU activation function, and W1 and W2 are the weights of the two fully connected layers respectively.

[0092] Step 3-2, weight the obtained attention weights to the predicted results output by the multi-source teacher model to obtain the fused predicted results p t of the multi-source teacher model, which is expressed as:

[0093]

[0094] Among them, p t represents the prediction result of the multi-source teacher model after fusion, including the prediction result p t (L t ) of the labeled data and the prediction result p t (U t ).

[0095] Step 4: Use the Kullback-Leibler divergence for multi-source data distillation learning;

[0096] The specific process of the multi-source data distillation strategy is as follows:

[0097] In step 4-1, when learning task t, the general model G t serves as the student model, and the model G t-1 of task t-1 serves as the teacher model of model G t . The distillation loss L KL between the output prediction probabilities of the teacher model and the student model is calculated by the KL divergence:

[0098]

[0099] Among them, T t-1 is the number of learned categories after task t-1, represents the prediction probability output by the multi-source teacher model for the t-1th task, represents the prediction probability output by the student model for the tth task.

[0100] In step 4-2, the total loss function L t of task t is composed of the cross-entropy loss function L c and the multi-source data distillation loss L KL . The cross-entropy loss function L c is expressed as:

[0101]

[0102] Among them, c t represents the number of sample categories of task t, p t (L t ) represents the prediction result of the model after the feature representation fusion of the labeled samples in task t, represents the true label of the ith labeled sample in task t.

[0103] In step 4-3, the calculation formula of the total loss function L t is:

[0104] L t = L c + λL KL

[0105] Among them, L c represents the cross-entropy loss function, and L KL is the distillation loss between the output prediction probabilities of the teacher model and the student model, and λ is the weight of the multi-source data distillation loss function.

[0106] Step 5: Use a logistic regression classifier to improve the label propagation algorithm for semi-supervised working condition identification;

[0107] By fusing multi-source data features and using a logistic regression classifier to improve the label propagation algorithm, make full use of the limited labeled data and a large amount of unlabeled data to improve the accuracy of pumping well working condition identification. The specific steps are as follows:

[0108] Step 5-1: Average fuse the features output by the graph convolutional networks of different data sources The feature matrix X of task t after fusion t The calculation formula is

[0109]

[0110] Among them, X t contains the global feature matrix X of labeled data L and the global feature matrix X of unlabeled data U , where: X t = [X L , X U .

[0111] Step 5-2: The multi-source semi-supervised graph classification method based on enhanced label propagation is expressed as:

[0112]

[0113] Among them, is the adjacency matrix constructed using the fused features, contains the adjacency matrix constructed from labeled data and the adjacency matrix constructed from unlabeled data The calculation formula is:

[0114]

[0115] Among them, represents the probability that the k-th sample belongs to class c. The calculation formula of

[0116]

[0117] Among them, y k represents the label of the k-th sample.

[0118] Step 5-3: Use a logistic regression classifier to generate an initial predicted probability distribution for unlabeled data, providing a more accurate initial label vector for label propagation.

[0119] After completing the label propagation process, use multiple logistic regression classifiers to make P LR Classify the label embedding vectors and calculate the average of the results of these classifiers to obtain the final classification label. The final label y of task t t The calculation formula is:

[0120]

[0121] where is the label vector of task t after label propagation.

[0122] To implement the above-mentioned multi-source semi-supervised class incremental working condition recognition method for pumping wells, in this application, an identification system is further included, as Figure 3 shown. The identification system includes: a working condition sample library storage module, a working condition recognition model construction module, and a working condition recognition module. The working condition sample library storage module further includes an actual measured surface dynamogram sample library and an actual measured electric power diagram sample library.

[0123] The actual measured surface dynamogram sample library and the actual measured electric power diagram sample library store the actual measured surface dynamograms and actual measured electric power diagrams respectively, aiming to collect and manage the pumping well working condition samples used for training and to be identified in the system. The actual measured surface dynamogram sample library and the actual measured electric power diagram sample library are respectively connected to the working condition recognition model construction module. The working condition recognition model construction module contains a multi-source semi-supervised class incremental working condition recognition model for pumping wells, and its output end is connected to the input end of the working condition recognition module. The working condition recognition module is used to classify and identify the pumping well working conditions to be identified through the working condition recognition model.

[0124] Next, take an example to verify the actual recognition effect of the above-mentioned multi-source semi-supervised class incremental working condition recognition method for pumping wells:

[0125] First, a dataset containing 1650 samples was constructed, covering 11 typical working conditions, with 150 samples for each working condition. In the class incremental learning setting, the initial task contains 2 categories, and then each new task adds 1 or 2 categories. This design simulates the scenario where new working condition categories gradually appear in oilfield operations. Specifically, the 11 typical working conditions in the dataset include: normal, insufficient liquid supply, continuous pumping with spraying, wax deposition, pump leakage, tubing leakage, fixed valve leakage, traveling valve leakage, sucker rod breakage, pump stuck, and pump traveling valve failure.

[0126] Then, following the actual operating conditions of oilfield pumping wells, the initial working conditions are set to two working condition categories, namely "normal" and "insufficient liquid supply", which are common in oilfield production. In the present invention, the semi-supervised learning ratios are set to 10%, 30%, 50%, 70%, and 100% respectively.

[0127] Subsequently, the above-measured surface dynamometer cards and measured electric power diagrams are used for working condition identification. When the semi-supervised learning ratio is set to 10%, the number of sample categories in the training set is set to 2, 3, 5, 7, 9, 11 respectively, and the average working condition recognition rates are 99.33%, 99.12%, 98.98%, 97.89%, 98.09%, 96.69% respectively; when the semi-supervised learning ratio is set to 30%, the number of sample categories in the training set is set to 2, 3, 5, 7, 9, 11 respectively, and the average recognition rates are 99.47%, 99.30%, 99.15%, 98.35%, 98.98%, 98.29% respectively; when the semi-supervised learning ratio is set to 50%, the number of sample categories in the training set is set to 2, 3, 5, 7, 9, 11 respectively, and the average recognition rates are 99.51%, 99.34%, 99.28%, 98.85%, 99.17%, 98.72% respectively; when the semi-supervised learning ratio is set to 70%, the number of sample categories in the training set is set to 2, 3, 5, 7, 9, 11 respectively, and the average recognition rates are 99.61%, 99.45%, 99.39%, 98.98%, 99.19%, 98.84% respectively; when the semi-supervised learning ratio is set to 100%, the number of sample categories in the training set is set to 2, 3, 5, 7, 9, 11 respectively, and the average recognition rates are 99.78%, 99.57%, 99.41%, 99.03%, 99.33%, 98.99% respectively.

[0128] As can be seen from the above, the multi-source semi-supervised class incremental working condition recognition method for this pumping well shows good robustness under most labeled data ratios and different category numbers. In the case where the labeled data is relatively scarce (semi-supervised ratio of 10% and the number of categories is 11), the accuracy of the method proposed in the present invention reaches 96.69%. Under the condition of scarce data, the multi-source semi-supervised class incremental working condition recognition method for this pumping well can effectively utilize the limited labeled data, and at the same time make full use of the unlabeled data through semi-supervised learning to improve the recognition performance. As the ratio of labeled data gradually increases, the performance of the method of the present invention is further improved. When the semi-supervised ratio is 100% and the number of categories is 11, the accuracy of the multi-source semi-supervised class incremental working condition recognition method for this pumping well reaches 98.99%.

[0129] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention in other forms. Any person skilled in the art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for identifying multi-source semi-supervised incremental operating conditions of a pumping well, characterized by: The steps include: Step 1, respectively storing a sample library containing two data sources of measured ground dynamometer diagrams and measured electric dynamometer diagrams corresponding to marked and unmarked pumping well operating conditions; Step 2, constructing a graph neural network teacher model for two data sources, the measured ground dynamometer diagram and the measured electric dynamometer diagram; Step 3: Use the Squeeze-and-Excitation attention mechanism to dynamically fuse the prediction probabilities of each teacher model; Step 4: Use Kullback-Leibler divergence to perform multi-source data distillation learning; Step 5: Use the label propagation algorithm improved by the logistic regression classifier to perform semi-supervised working condition identification.

2. The multi-source semi-supervised incremental operating condition identification method for a pumping well according to claim 1 is characterized in that: Step 3 specifically includes the following steps: Step 3-1, refer to the Squeeze-and-Excitation attention mechanism to implement the multi-source feature representation fusion module, which is composed of a fully connected layer-ReLU activation layer-fully connected layer-Sigmoid activation layer. The predicted probability output of the teacher model for the vth data source branch is expressed as Its corresponding attention score ω v It is expressed as: ω v =sigmoid ( W2[relu(W1[p t v ])]) Among them, sigmoid(·) is the Sigmoid activation function, relu(·) is the ReLU activation function, W1 and W2 are the weights of the two fully connected layers respectively; Step 3-2, add the obtained attention weight to the prediction result output by the multi-source teacher model to obtain the fused multi-source teacher model prediction result p t It is expressed as: Among them, p t Represents the prediction results of the fused multi-source teacher model, including the prediction results of the labeled data p t (L t ) and unlabeled data prediction results p t (U t ).

3. The multi-source semi-supervised incremental operating condition identification method for a pumping well according to claim 1 is characterized in that: Step 4 specifically includes the following steps: Step 4-1, when learning task t, the model G is usually t As the student model, the model G of task t-1 t-1 As Model G t The distillation loss L between the teacher model, the output prediction probability of the teacher model and the student model KL Calculated by KL divergence: Among them, T t-1 is the number of categories learned after task t-1, represents the predicted probability output by the multi-source teacher model for the t-1th task, represents the predicted probability of the output of the student model for the tth task; Step 4-2, total loss function L for task t t By the cross entropy loss function L c and multi-source data distillation loss L KL Composition, cross entropy loss function L c It is expressed as: Among them, c t represents the number of sample categories of task t, p t (L t ) indicates that there are labeled sample features in task t, which represent the prediction results of the fused model. Represents the true label of the i-th label sample in task t; Step 4-3, total loss function L t The calculation formula is: THE t =L c +λL KL Among them, L c represents the cross entropy loss function, L KL is the distillation loss between the output prediction probabilities of the teacher model and the student model, and λ is the weight of the multi-source data distillation loss function.

4. The multi-source semi-supervised incremental operating condition identification method for a pumping well according to claim 1 is characterized in that: Step 5 specifically includes the following steps: Step 5-1: Output features of graph convolutional networks from different data sources Perform mean fusion, and the feature matrix X of task t after fusion t The calculation formula is: in, X t Contains the global feature matrix X of labeled data L And the unlabeled data global feature matrix X U , where: X t =[X L ,X U ]; Step 5-2, the multi-source semi-supervised graph classification method based on enhanced label propagation is expressed as: in, is the adjacency matrix constructed using the fused features, Contains the adjacency matrix constructed with labeled data And the adjacency matrix constructed from unlabeled data The calculation formula is: in, represents the probability that the kth sample belongs to class c, The calculation formula is: Among them, y k represents the label of the kth sample; Step 5-3, use the logistic regression classifier to generate the initial predicted probability distribution for the unlabeled data, and provide a more accurate initial label vector for label propagation. After completing the label propagation process, use multiple logistic regression classifiers for P LR Classify the label embedding vectors and calculate the average of these classifier results to get the final classification label; the final label y of task t t The calculation formula is: in, is the label vector of task t after label propagation.

5. The multi-source semi-supervised incremental operating condition identification method for a pumping well according to claim 1 is characterized in that: Step 2 specifically includes the following steps: Step 2-1, for each sample in a single data source, calculate the distance between it and all other samples in its data set, and select the two samples with the smallest distance as its nearest neighbor samples to construct the initial adjacency matrix of the graph convolutional network. The calculation formula of the graph convolutional network for the tth task is: in, The adjacency matrix constructed for the vth data source in task t, is the data of the vth data source in task t, including the labeled data and unlabeled data of the vth data source; Step 2-2, use the convolutional layer of the graph convolutional network to extract node features: in, I K is the identity matrix, relu(·) is the ReLU activation function, W is the trainable parameter matrix, is a diagonal matrix with the following diagonal elements: It is the feature output by the l-th layer graph convolutional network; Step 2-3, the predicted probability output by the teacher model of the vth data source branch Among them, FC is a fully connected layer, which is used to classify the features output by the graph convolutional network. is the vth data source data in task t, It is the feature output by the l+1th layer graph convolutional network.

6. The multi-source semi-supervised incremental operating condition identification method for a pumping well according to claim 1 is characterized in that: In step 1, the measured ground dynamometer diagram is a binary image composed of the displacement of the bare rod and the load of the bare rod, and the measured electric dynamometer diagram is a binary image composed of the displacement of the bare rod and the active power of the motor.

7. An identification system for implementing the multi-source semi-supervised incremental working condition identification method for a pumping well as claimed in any one of claims 1 to 6, characterized in that: It includes a working condition sample library storage module, a working condition identification model construction module and a working condition identification module, wherein the working condition sample library storage module further includes a measured ground dynamometer diagram sample library and a measured electric dynamometer diagram sample library; The measured ground dynamometer diagram sample library and the measured electric dynamometer diagram sample library respectively store the measured ground dynamometer diagram and the measured electric dynamometer diagram, and are intended to collect and manage the working condition samples of the pumping wells used for training and to be identified in the system; the measured ground dynamometer diagram sample library and the measured electric dynamometer diagram sample library are respectively connected to the working condition identification model construction module; The working condition identification model construction module includes a multi-source semi-supervised incremental working condition identification model for pumping wells, and its output end is connected to the input end of the working condition identification module; The working condition identification module is used to classify and identify the working condition of the pumping well to be identified through the working condition identification model.

Images