Classification Method for Working Conditions of Conventional Well Workover Operations in Oil and Gas Fields

Abstract

The present application relates to the field of petroleum engineering, and provides a method for classifying the working conditions of conventional workover operations in oil and gas fields. It uses data processing technology based on deep learning to decompose and semantically encode the text descriptions of conventional workover operations in oil and gas fields. Then, the process semantic feature corresponding to the i-th process is extracted from the set of encoded process semantic features as the process semantic feature to be recognized, and the entire set is divided into the sets of the process semantic features of the previous text and the subsequent text with this as the dividing line. Then, the process semantic feature to be recognized is contextually encoded with the sets of the process semantic features of the previous text and the subsequent text respectively, so as to intelligently judge the operation category of the process to be recognized according to the positive and negative attention responses between the context semantic features of the previous text and the subsequent text of the process to be recognized. In this way, the efficiency and accuracy of process classification can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of petroleum engineering, and more specifically, to a method for classifying the working conditions of conventional workover operations in oil and gas fields. Background Art

[0002] Workover operations in oil and gas fields are a key link in ensuring the daily production, management, and optimization of oil, gas, and water wells. These operations are not only crucial for maintaining and restoring production capacity but also for effectively addressing problems in oil and gas reservoirs and wellbore structures, removing faults, and thus ensuring the stable and increased production of oil wells. With the continuous deepening of exploration and development, the technical complexity and diversity of workover operations are also increasing, resulting in a series of process technologies for different well types and conditions.

[0003] However, despite the numerous types of workover operation procedures and diverse processes, in actual operation, it faces the problem of being difficult to achieve efficient and targeted management. Traditional workover operation management in oil and gas fields mainly relies on manual experience to judge working conditions, which is particularly insufficient when faced with the operation details of different procedures and complex working conditions. In addition, manual identification and classification are not only inefficient but also prone to misjudgment, leading to waste of resources and an increase in potential risks.

[0004] Therefore, an optimized classification scheme for the working conditions of conventional workover operations in oil and gas fields is needed. Summary of the Invention

[0005] This application aims at the deficiencies in the prior art and provides a method for classifying the working conditions of conventional workover operations in oil and gas fields.

[0006] According to one aspect of the present application, there is provided a method for classifying the working conditions of conventional workover operations in oil and gas fields, which includes: obtaining a text description of the conventional workover operations in oil and gas fields; decomposing and semantically encoding the text description of the conventional workover operations in oil and gas fields to obtain a set of semantic encoding vectors of the workover operation processes in oil and gas fields; extracting the semantic encoding vector of the workover operation process corresponding to the i-th process from the set of semantic encoding vectors of the workover operation processes in oil and gas fields as the semantic encoding vector of the process to be identified; using the semantic encoding vector of the process to be identified as a boundary to divide the set of semantic encoding vectors of the workover operation processes in oil and gas fields into a set of semantic encoding vectors of the previous workover operation processes in oil and gas fields and a set of semantic encoding vectors of the subsequent workover operation processes in oil and gas fields; respectively performing context encoding on the set of semantic encoding vectors of the previous workover operation processes in oil and gas fields and the set of semantic encoding vectors of the subsequent workover operation processes in oil and gas fields to obtain a context semantic encoding vector before the process to be identified and a context semantic encoding vector after the process to be identified; performing forward and reverse attention field interaction response on the context semantic encoding vector before the process to be identified and the context semantic encoding vector after the process to be identified to obtain a two-way context semantic encoding vector of the process to be identified, wherein the forward and reverse attention field interaction response is performed by constructing a forward and reverse context semantic two-way attention balance field of the process to be identified before and after, and performing semantic mapping modulation and response encoding based on the constructed forward and reverse context semantic two-way attention balance field of the process to be identified before and after; based on the two-way context semantic encoding vector of the process to be identified, obtaining an identification result for representing the operation category label of the i-th process.

[0007] Due to the adoption of the above technical solutions, the present application has remarkable technical effects: The method for classifying the working conditions of conventional workover operations in oil and gas fields provided by the present application uses deep learning-based data processing technology to decompose and semantically encode the text description of conventional workover operations in oil and gas fields, then extracts the process semantic feature corresponding to the i-th process from the set of encoded process semantic features as the process semantic feature to be identified, and uses it as a boundary to divide the entire set into a set of previous and subsequent process semantic features, and then context encodes the process semantic feature to be identified with the set of previous process semantic features and the set of subsequent process semantic features respectively, so as to intelligently judge the operation category of the process to be identified based on the forward and reverse attention responses between the context semantic features before and after the process to be identified. In this way, the efficiency and accuracy of process classification can be effectively improved. Brief Description of the Drawings

[0008] The above and other objects, features, and advantages of the present application will become more apparent by describing the embodiments of the present application in more detail with reference to the accompanying drawings. The drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 It is a flowchart of a method for classifying the working conditions of conventional workover operations in oil and gas fields according to an embodiment of the present application.

[0010] Figure 2 It is a flowchart of S120 in the method for classifying the working conditions of conventional workover operations in oil and gas fields according to an embodiment of the present application.

[0011] Figure 3 It is a flowchart of S150 in the method for classifying the working conditions of conventional workover operations in oil and gas fields according to an embodiment of the present application.

[0012] Figure 4 It is a flowchart of S160 in the method for classifying the working conditions of conventional workover operations in oil and gas fields according to an embodiment of the present application. Detailed implementation manners

[0013] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0014] Workover operations are a crucial part of oil and gas field production management, directly related to the daily operation and maintenance of oil, gas, and water wells and the improvement of production capacity. Such operations can not only effectively solve the problems existing in the wellbore structure and oil and gas reservoirs, but also eliminate faults, restore or improve production performance, thus providing guarantee for the stable and increased production of oil wells. With the in-depth exploration and development of oil and gas fields, the technical means of workover operations have become increasingly rich, and their complexity and diversity have also increased significantly, gradually developing a series of technological processes for different well conditions and well types.

[0015] However, despite the fact that workover operations cover a variety of processes and operation flows, there are still many challenges in their management efficiency and pertinence. In actual operation, the traditional workover operation management mode mainly relies on manual experience to judge the working conditions. This method often seems powerless in the face of complex operation environments and diverse process details. Manual identification and classification not only have low efficiency, but also are prone to misjudgment, resulting in waste of resources, while increasing operation risks and potential hazards.

[0016] In view of the above technical problems, the technical concept of this application is to obtain the text description of conventional workover operations in oil and gas fields, use natural language processing and encoding techniques based on deep learning to decompose and semantically encode the text description of conventional workover operations in oil and gas fields. Then, extract the workover operation process semantic feature corresponding to the i-th process from the set of encoded workover operation process semantic features in the oil and gas field as the work process semantic feature to be recognized, and use it as the dividing line to divide the entire set into the front and back workover operation process semantic features and sets in the oil and gas field. Then, perform context encoding on the work process semantic feature to be recognized respectively with the set of front workover operation process semantic features and the set of back workover operation process semantic features in the oil and gas field, so as to intelligently judge the operation category of the work process to be recognized according to the positive and negative attention response representation between the context semantic features before the work process to be recognized and the context semantic features after the work process to be recognized. This application can automatically parse the text description of conventional workover operations in oil and gas fields and perform semantic encoding on it. This not only improves the speed of process classification, but also ensures the consistency and accuracy of the classification results, avoiding the subjective deviation caused by manual judgment. Moreover, it enables each process to be analyzed independently and in detail, which helps to achieve more refined management and optimization, especially in the face of complex and changing working conditions.

[0017] Figure 1 FIG. is a flowchart of a method for classifying the working conditions of conventional workover operations in oil and gas fields according to an embodiment of the present application. As Figure 1As shown, the method for classifying the working conditions of conventional workover operations in oil and gas fields according to the embodiments of the present application includes: S110, obtaining the text description of the conventional workover operations in oil and gas fields; S120, decomposing and semantically encoding the text description of the conventional workover operations in oil and gas fields to obtain a set of semantic encoding vectors of the workover operation processes in oil and gas fields; S130, extracting the semantic encoding vector of the workover operation process corresponding to the i-th process from the set of semantic encoding vectors of the workover operation processes in oil and gas fields as the semantic encoding vector of the process to be recognized; S140, using the semantic encoding vector of the process to be recognized as a boundary, dividing the set of semantic encoding vectors of the workover operation processes in oil and gas fields into a set of semantic encoding vectors of the previous workover operation processes in oil and gas fields and a set of semantic encoding vectors of the subsequent workover operation processes in oil and gas fields; S150, respectively performing context encoding on the set of semantic encoding vectors of the previous workover operation processes in oil and gas fields and the set of semantic encoding vectors of the subsequent workover operation processes in oil and gas fields to obtain the context semantic encoding vector of the previous context of the process to be recognized and the context semantic encoding vector of the subsequent context of the process to be recognized; S160, performing positive and negative attention field interaction response on the context semantic encoding vector of the previous context of the process to be recognized and the context semantic encoding vector of the subsequent context of the process to be recognized to obtain the two-way context semantic encoding vector of the process to be recognized; S170, based on the two-way context semantic encoding vector of the process to be recognized, obtaining the recognition result of the operation category label representing the i-th process.

[0018] In step S110, the text description of the conventional workover operations in oil and gas fields is obtained. It should be understood that the text description of the conventional workover operations in oil and gas fields mainly includes operation process information, operation equipment and tool information, etc. Specifically, the operation process information includes the specific content of each process, such as well killing and well flushing in the flushing operation; milling and grinding in the rotary operation; the specifications and operation methods of pipe string running and pulling in the pipe string running and pulling operation and the special pipe string running and pulling operation, and also includes the operation details in different construction stages of each process. The operation equipment and tool information involves equipment such as mud pumps (used for flushing operations, circulating and squeezing fluids into the well), rotary tables, top drives, power drills, positive displacement motors (used for rotary operations, rotating the pipe string and tools in the well), well testing machines, workover rigs (used for pipe string running and pulling operations and special pipe string running and pulling operations), etc., and various construction tools. The usage conditions of these equipment and tools are one of the important bases for judging the operation type. Generally speaking, the text description of the conventional workover operations in oil and gas fields is the original data source for identifying the operation category labels of the processes. By analyzing this data, the specific operation content and working conditions of each process can be understood, so as to establish a corresponding relationship with the preset operation categories (such as flushing operations, rotary operations, etc.), thereby realizing the final identification of the operation category labels.

[0019] In step S120, the text description of the conventional workover operation in the oil and gas field is decomposed and semantically encoded to obtain a set of semantic encoding vectors of the conventional workover operation procedures in the oil and gas field. Specifically, Figure 2 FIG. Figure 2 is a flowchart of S120 in the method for classifying the working conditions of the conventional workover operation in the oil and gas field according to the embodiment of the present application. As Figure 2 shown, the step S120 includes: S121, decomposing the text description of the conventional workover operation in the oil and gas field based on the construction stage to obtain a set of text descriptions of the conventional workover operation procedures in the oil and gas field; S122, semantically encoding each text description of the conventional workover operation procedure in the set of text descriptions of the conventional workover operation procedures in the oil and gas field to obtain a set of semantic encoding vectors of the conventional workover operation procedures in the oil and gas field.

[0020] In step S121, the text description of the conventional workover operation in the oil and gas field is decomposed based on the construction stage to obtain a set of text descriptions of the conventional workover operation procedures in the oil and gas field. Accordingly, considering that the workover operation in the oil and gas field is a complex and highly specialized process involving multiple different types of operation links. For example, from pulling out the downhole string, inspecting and repairing the equipment, to reinserting the string and debugging, each major construction stage includes numerous specific procedures. Therefore, in order to better sort out the operation process to clarify each specific work link, the present application decomposes the text description of the conventional workover operation in the oil and gas field based on the construction stage to decompose the overall workover operation text description into text descriptions of specific procedures, obtaining a set of text descriptions of the conventional workover operation procedures in the oil and gas field. In this way, the detailed information of each operation link can be understood more carefully. For example, in the text description of the "oil pipe replacement" procedure, specific operation steps such as oil pipe disassembly, inspection and installation of the new oil pipe can be clarified, so as to accurately judge the function and category of this procedure in the entire workover operation.

[0021] In step S122, semantic encoding is performed on each of the oil and gas field conventional workover operation process text descriptions in the set of oil and gas field conventional workover operation process text descriptions to obtain the set of oil and gas field conventional workover operation process semantic encoding vectors. Specifically, in the embodiment of the present application, step S122 includes: using a semantic encoder including an embedding layer to perform semantic encoding on each of the oil and gas field conventional workover operation process text descriptions in the set of oil and gas field conventional workover operation process text descriptions to obtain the set of oil and gas field conventional workover operation process semantic encoding vectors. It should be understood that in order to further capture and excavate the specific semantics contained in each oil and gas field conventional workover operation process text description, in the technical solution of the present application, semantic encoding is performed on each of the oil and gas field conventional workover operation process text descriptions in the set of oil and gas field conventional workover operation process text descriptions to accurately capture the deep meaning in the text and obtain the set of oil and gas field conventional workover operation process semantic encoding vectors. In particular, in an embodiment of the present application, a semantic encoder including an embedding layer is used to perform semantic encoding on each of the oil and gas field conventional workover operation process text descriptions in the set of oil and gas field conventional workover operation process text descriptions to map each word to a low-dimensional continuous vector space, convert the high-dimensional sparse text data into a low-dimensional dense vector representation, and obtain the set of oil and gas field conventional workover operation process semantic encoding vectors. That is, through the combination of the embedding layer and the semantic encoder, the oil and gas field conventional workover operation process text description can be converted into a vector containing rich semantic information. These vectors not only contain the semantics of individual words, but also capture the semantic relationships such as the order and combination of words in the text through the processing of the entire text sequence by the semantic encoder, which can provide high-quality feature representations for subsequent process analysis and classification.

[0022] The following is a detailed elaboration of a specific implementation process of "using a semantic encoder including an embedding layer to perform semantic encoding on each of the oil and gas field conventional workover operation process text descriptions in the set of oil and gas field conventional workover operation process text descriptions to obtain the set of oil and gas field conventional workover operation process semantic encoding vectors": First, text preprocessing needs to be performed on the set of oil and gas field conventional workover operation process text descriptions, and this processing directly affects the performance of the final model. During this process, each oil and gas field conventional workover operation process text description needs to be first segmented into individual words or phrases, and this step is usually called word segmentation. By means of technical means such as removing stop words, punctuation marks, and performing stemming or lemmatization, the original text can be effectively purified to make it more easily understood and processed by a computer. Subsequently, a vocabulary is generated based on all text descriptions, and an index value is assigned to each unique word. This step not only simplifies the subsequent data processing flow, but also lays a foundation for constructing the embedding layer.

[0023] Next is the application stage of the embedding layer, which is one of the core parts of the entire implementation process. At this stage, word embedding techniques such as Word2Vec, GloVe, or FastText will be used to convert each word into a low-dimensional dense vector representation. This vector can not only capture the semantic information of the word itself but also reflect the relationships between words, thus providing strong support for subsequent semantic encoding. For the text description of each process, the corresponding word vectors are concatenated according to their original order to form a vector sequence representing the entire process. This process is essentially converting human-readable text information into a form that can be understood and processed by a computer.

[0024] Then, a suitable semantic encoder model architecture is built. In this scenario, deep learning models such as recurrent neural networks (RNNs) and their variants (such as LSTMs, GRUs) can be selected as the basic architecture of the semantic encoder. These models have their own characteristics. For example, RNNs can effectively process sequence data, especially when dealing with long-distance dependencies. LSTMs and GRUs perform well due to their special gating mechanisms. After selecting the model, it is trained using the labeled dataset. During the training process, the cross-entropy loss function or other suitable loss functions are used to optimize the model parameters, aiming to make the model learn how to convert the input word vector sequence into a vector representation that can better express the semantic features of the process. This stage may require a large amount of computing resources and time, which directly determines the quality of the final model.

[0025] Finally, a set of semantic encoding vectors for the conventional workover operations in oil and gas fields is output. The well-trained model can receive the set of processed process text descriptions, convert them into word vector sequences through the embedding layer, and then be processed by the core architecture of the semantic encoder, finally outputting a set of semantic encoding vectors for the conventional workover operations in oil and gas fields. Each vector in the set is not only a highly condensed representation of the process itself but also contains rich semantic information, which can be used for analysis in subsequent steps. Taking "tubing replacement" as an example, assume there is a process text description: "Pull out the old tubing, inspect the new tubing, and install the new tubing". After being processed through the above steps, this description will be converted into a series of word vectors and further generate a vector that can represent the semantic meaning of the entire process operation through the semantic encoder.

[0026] It should be noted that during the entire implementation process, issues related to data quality and model optimization need to be considered. A high-quality dataset is the key to success, so it is necessary to ensure the integrity and accuracy of the prepared data. At the same time, as the model training progresses, it is also very necessary to continuously adjust and optimize the model parameters. This includes, but is not limited to, adjusting hyperparameters such as the learning rate, batch size, number of layers, and number of units in each layer to find the configuration that best suits the current task. In addition, the use of pre-trained models for transfer learning can also be explored. This method can, to a certain extent, alleviate the problem of insufficient data and improve the generalization ability of the model.

[0027] In step S130, the i-th oil and gas field conventional workover operation process semantic encoding vector corresponding to the oil and gas field conventional workover operation process semantic encoding vector set is extracted as the workover operation process semantic encoding vector to be recognized. Correspondingly, considering that the oil and gas field conventional workover operation includes multiple processes, each process has its unique operation content and characteristics. Therefore, in order to extract a single process vector from the overall process semantic encoding vector set to focus on in-depth analysis of this specific process, in the technical solution of this application, the i-th oil and gas field conventional workover operation process semantic encoding vector corresponding to the oil and gas field conventional workover operation process semantic encoding vector set is extracted as the workover operation process semantic encoding vector to be recognized. In this way, it is possible to ensure a meticulous analysis of this process, which helps to accurately identify the specific category label of each process subsequently.

[0028] In step S140, using the semantic encoding vector of the operation to be recognized as a boundary, the set of semantic encoding vectors of the conventional workover operation procedures in the oil and gas field is divided into the set of semantic encoding vectors of the previous conventional workover operation procedures in the oil and gas field and the set of semantic encoding vectors of the subsequent conventional workover operation procedures in the oil and gas field. It should be understood that in natural language understanding and actual operation process understanding, relevant information before and after an event or operation will be naturally considered to comprehensively understand its meaning. Similarly, in the analysis of workover operation procedures in the oil and gas field, the meaning and characteristics of an operation are often affected by its previous and subsequent operations. For example, before the "tubing replacement" operation, there may be a "tubing removal" operation, and after that, there may be a "new tubing installation" operation. Therefore, in order to be able to understand the operation in combination with the context, in this application, using the semantic encoding vector of the operation to be recognized as a boundary, the set of semantic encoding vectors of the conventional workover operation procedures in the oil and gas field is divided into the set of semantic encoding vectors of the previous conventional workover operation procedures in the oil and gas field and the set of semantic encoding vectors of the subsequent conventional workover operation procedures in the oil and gas field. In this way, by separately constructing the sets of semantic encoding vectors of the previous and subsequent conventional workover operation procedures in the oil and gas field, the context semantic information of the operation to be recognized can be obtained from two directions. Specifically, the previous set can reflect the preconditions and preparatory operations that lead to the execution of this operation, and the subsequent set can reflect the subsequent operations and impacts caused after this operation is completed. For example, by analyzing the previous set, it can be known that the "tubing pulling out" operation was carried out before the "tubing cleaning" operation, providing conditions for tubing cleaning, and by analyzing the subsequent set, it can be known that the "tubing inspection" operation was carried out after cleaning to check the cleaning effect.

[0029] In step S150, context encoding is respectively performed on the set of semantic encoding vectors of the previous conventional workover operation procedures in the oil and gas field and the set of semantic encoding vectors of the subsequent conventional workover operation procedures in the oil and gas field to obtain the context semantic encoding vector of the previous context of the operation to be recognized and the context semantic encoding vector of the subsequent context of the operation to be recognized. Specifically, Figure 3 The flowchart of S150 in the method for classifying the working conditions of conventional workover operations in the oil and gas field according to an embodiment of the present application is as follows. As Figure 3 shown, the step S150 includes: S151, inputting the semantic encoding vector of the operation to be recognized and the set of semantic encoding vectors of the previous conventional workover operation procedures in the oil and gas field into a one-way context encoder based on a forward LSTM model to obtain the context semantic encoding vector of the previous context of the operation to be recognized; S152, inputting the semantic encoding vector of the operation to be recognized and the set of semantic encoding vectors of the subsequent conventional workover operation procedures in the oil and gas field into a one-way context encoder based on a backward LSTM model to obtain the context semantic encoding vector of the subsequent context of the operation to be recognized.

[0030] In step S151, the semantic encoding vector of the process to be recognized and the set of semantic encoding vectors of the previous oil and gas field conventional workover operation processes are input into a one-way context encoder based on a forward LSTM model to obtain the context semantic encoding vector of the previous context of the process to be recognized. Accordingly, considering that the semantic features before the semantic encoding vector of the process to be recognized express some previous work information of this process to be recognized. For example, before the "tubing installation" process, there may be a "check the pipe string and prepare a new pipe string" process immediately. Therefore, in order to comprehensively understand the context semantic information between each process in the previous context, in the technical solution of this application, the semantic encoding vector of the process to be recognized and the set of semantic encoding vectors of the previous oil and gas field conventional workover operation processes are input into a one-way context encoder based on a forward LSTM model to capture the dependency relationships between these information, and the context semantic encoding vector of the previous context of the process to be recognized is obtained. It should be understood that the forward LSTM model processes data in order from the starting position of the sequence, can make full use of the information in the set of semantic encoding vectors of the previous oil and gas field conventional workover operation processes, and gradually learn and remember the influence of the previous context on the current process to be recognized along the time or the process execution order. For the process to be recognized, the information such as the operation purpose and preparatory work contained in its previous process is crucial for accurately understanding the semantics of this process, and the forward LSTM model can effectively capture these information, so as to more comprehensively reflect the semantic features of the process to be recognized based on the previous context in the entire operation process.

[0031] The following is a detailed elaboration of a specific implementation process of "inputting the semantic encoding vector of the process to be recognized and the set of semantic encoding vectors of the previous oil and gas field conventional workover operation processes into a unidirectional context encoder based on a forward LSTM model to obtain the context semantic encoding vector of the previous text of the process to be recognized": First is the data preparation and format adjustment stage. In this stage, two key data parts are required. On the one hand, there is the semantic encoding vector of the process to be recognized, which is obtained by performing semantic encoding on the text description of the process to be recognized in the oil and gas field conventional workover operation. This vector contains the unique semantic information of this process. On the other hand, there is the set of semantic encoding vectors of the previous oil and gas field conventional workover operation processes. These vectors are also obtained by performing semantic encoding on the text descriptions of the previous processes. They represent the semantic features of a series of operation processes before the process to be recognized. To enable these data to enter the forward LSTM model for processing smoothly, some necessary operations need to be carried out. First, the order of the data needs to be determined. The vectors in the set of previous process vectors must be arranged in the order of the execution sequence of the processes. This is because the forward LSTM model processes sequence data in order. Only by ensuring the correct order can the model accurately capture the temporal dependence relationship and context information between the processes. At the same time, these vectors also need to be adjusted to a format suitable for input into the forward LSTM model. Generally, the input expected by the forward LSTM model is a three-dimensional tensor with a shape of (number of samples, number of time steps, feature dimension). In the application scenario of this application, the number of samples is generally 1 because it is for processing the context of a single process to be recognized. The number of time steps is the number of previous processes, which reflects how many processes before the process to be recognized need to be considered. The feature dimension is the dimension of each semantic encoding vector, and this dimension determines the richness of the information contained in each process vector.

[0032] Next is to initialize the forward LSTM model. The forward LSTM model is composed of a series of LSTM cells. Each LSTM cell has its unique structure and function, mainly including an input gate, a forget gate, an output gate, and a cell state. These gating mechanisms enable the forward LSTM model to effectively process sequence data and capture the long-term dependence relationship in the data. When defining the model structure, it is necessary to ensure that the input dimension of the model is consistent with the dimension of the semantic encoding vector. This is because only when the dimensions match can the model correctly receive the input data and effectively process the information of the input process vector. At the same time, the output dimension of the model also needs to be adjusted according to actual needs. The selection of the output dimension usually needs to consider various factors, such as the complexity of the subsequent task, the feature scale of the data, etc. Generally, a suitable dimension needs to be selected to represent the context semantic information so that this information can be better utilized for classification and other operations in the future.

[0033] When initializing the model parameters, a random initialization method can be adopted. Common initialization methods include Xavier initialization or He initialization. The purpose of these initialization methods is to ensure that the model can converge stably in the initial stage of training. By reasonably initializing the weights and biases, problems such as vanishing gradients or exploding gradients during the training process can be avoided, enabling the model to more effectively learn the context relationship between processes.

[0034] After completing the model initialization, it enters the sequence input and forward propagation stage. In this stage, the previous process vectors are input into the forward LSTM model in sequence. Starting from the first previous process vector, it is used as the input to the LSTM model at the first time step. The LSTM unit will perform a series of complex calculations based on the current input and the hidden state of the previous time step. Specifically, first is the calculation of the forget gate, which processes the input and the hidden state of the previous time step through a sigmoid function to obtain a vector between 0 and 1, which determines which information in the cell state needs to be forgotten. Then is the calculation of the input gate, also using the sigmoid function, which determines which new information needs to be added to the cell state. At the same time, the candidate value of the cell state is calculated, using the hyperbolic tangent function to transform the input and the hidden state of the previous time step to obtain a new vector. Then, according to the outputs of the forget gate and the input gate, the cell state is updated. The cell state is a key part of the LSTM model for transmitting long-term information, and it can retain important information between different time steps. Finally, through the processing of the output gate and the cell state, the hidden state of the current time step is obtained. The hidden state contains the context information up to the current time step, which will be an important basis for the calculation of the next time step.

[0035] After processing the first previous process vector, the subsequent previous process vectors are sequentially input into the LSTM model, repeating the above calculation process. Each new process vector input will update the hidden state and the cell state, enabling the model to gradually accumulate the context information of the previous processes. In this way, the forward LSTM model can effectively capture the temporal dependence relationship and context semantic information between processes, which can lay a foundation for accurately representing the previous context of the process to be recognized subsequently.

[0036] After completing the processing of the foregoing process vectors, fuse the process vector to be recognized with the context information. First, it is necessary to obtain the hidden state at the last time step. This hidden state is obtained after processing all the foregoing process vectors, and it contains all the context information of the foregoing processes. Then, fuse the semantic encoding vector of the process to be recognized with the hidden state at the last time step. There are various ways of fusion, and a common one is concatenation in the feature dimension. By concatenation, a new vector can be obtained, which combines the information of the process to be recognized itself and the context information of the foregoing processes. The advantage of this is that when the model considers the process to be recognized, it can not only focus on its own semantic features but also combine the background information of the foregoing processes, so as to more comprehensively understand the position and meaning of the process to be recognized in the entire operation process.

[0037] Finally, generate the semantic encoding vector of the context before the process to be recognized. Automatically input the fused vector into a fully connected layer or perform other forms of linear transformation. A fully connected layer is a common neural network layer that can map the input vector to the required dimension. Through the processing of the fully connected layer, a new vector is obtained, which is the semantic encoding vector of the context before the process to be recognized. It combines the context information of the foregoing processes and part of the information of the process to be recognized itself, and can better represent the semantic environment of the process to be recognized in the entire operation process. This vector will be an important input for subsequent data processing steps and provide key feature information for accurately classifying the working conditions of conventional workover operations in oil and gas fields.

[0038] In step S152, input the set of the semantic encoding vector of the process to be recognized and the semantic encoding vectors of the subsequent processes of the conventional workover operation in the oil and gas field into a one-way context encoder based on the backward LSTM model to obtain the semantic encoding vector of the context after the process to be recognized. It should be understood that in the workover operation of the oil and gas field, the execution of one process often has an impact on the subsequent processes, and this impact is crucial for comprehensively understanding the nature and role of this process. For example, after the "tubing installation" process, the "sealing detection" process may follow immediately. Through the backward LSTM model, information such as how the "tubing installation" process creates conditions for the "sealing detection" process and the verification requirements of the "sealing detection" for the quality of the "tubing installation" can be mined. Therefore, in order to effectively capture the information contained in the processes after the process to be recognized and their associations with the process to be recognized, the present application inputs the set of the semantic encoding vector of the process to be recognized and the semantic encoding vectors of the subsequent processes of the conventional workover operation in the oil and gas field into a one-way context encoder based on the backward LSTM model to capture and mine the semantic associations between the processes, and obtain the semantic encoding vector of the context after the process to be recognized.

[0039] In step S160, the semantic encoding vector of the context preceding the process to be identified and the semantic encoding vector of the context following the process to be identified are subjected to positive and negative attention field interactive responses to obtain a bidirectional semantic encoding vector of the context of the process to be identified. Specifically, Figure 4 Flow chart of S160 in the method for classifying the working conditions of conventional well repair operations in oil and gas fields according to an embodiment of the present application. Figure 4 As shown, the step S160 includes: S161, performing homography projection transformation on the semantic coding vector of the preceding context of the process to be identified and the semantic coding vector of the succeeding context of the process to be identified to obtain the homography projection coding vector of the preceding context of the process to be identified and the homography projection coding vector of the succeeding context of the process to be identified; S162, constructing a positive and negative preceding-postponing context bidirectional attention balance field between the homography projection coding vector of the preceding context of the process to be identified and the homography projection coding vector of the succeeding context of the process to be identified; S163, based on the positive and negative preceding-postponing context bidirectional attention balance field, performing semantic mapping modulation and response coding on the homography projection coding vector of the preceding context of the process to be identified and the homography projection coding vector of the succeeding context of the process to be identified to obtain the bidirectional context semantic coding vector of the process to be identified.

[0040] It should be understood that the preceding context semantic coding vector and the succeeding context semantic coding vector of the process to be identified provide information about the process to be identified from the front and back directions respectively, and can only provide partial context information. In order to ensure that the model obtains comprehensive and integrated information, so as to more accurately capture the true situation of the current process, in the technical solution of the present application, the preceding context semantic coding vector of the process to be identified and the succeeding context semantic coding vector of the process to be identified are subjected to positive and negative attention field interactive responses to obtain a bidirectional context semantic coding vector of the process to be identified. In this way, the information from the preceding and succeeding contexts can be deeply integrated, so that the model can fully and effectively utilize all relevant information. For example, the preceding text may emphasize the preparatory work of the process to be identified, and the succeeding text focuses on the impact of the process on subsequent processes. Through interactive responses, these two aspects of information can be integrated to form a more complete understanding of the process, so as to more accurately judge the job category label of the process to be identified.

[0041] In detail, first, it is necessary to perform homography projection transformation on the semantic coding vector of the preceding context of the process to be identified and the semantic coding vector of the following context of the process to be identified to obtain the homography projection coding vector of the semantic coding vector of the preceding context of the process to be identified and the homography projection coding vector of the semantic coding vector of the following context of the process to be identified. The above process can be expressed as: ;in, is the semantic encoding vector of the context preceding the process to be identified, is the semantic encoding vector of the context after the process to be recognized, is the homography matrix of the context mapping of the process to be recognized before the process, is the homography matrix of the context mapping of the process to be recognized after the process, is the homography projection encoding vector of the context before the process to be recognized, is the homography projection encoding vector of the context after the process to be recognized.

[0042] It should be understood that the homography projection transformation can map the semantic encoding vectors of the context before and after the process to be recognized to a new feature space. In the workover operation scenario of oil and gas fields, the information represented by the original semantic encoding vectors of the context before and after may be scattered and complex, and information such as different processes, equipment usage, and operation conditions are intertwined. By projecting to a new feature space, these information can be reorganized and integrated, enabling the model to process this information more systematically. And in the new feature space, the relative position relationship between features is maintained. This means that the internal association between the context before and after is retained. For example, in the workover operation, there is a specific association between the preparation process described in the previous context and the subsequent impact process described in the subsequent context in the original semantic encoding vector. After the homography projection transformation, this association still exists in the new feature space. This helps the model to still capture the logical connection between the context before and after in the new space and will not damage the coherence of the information due to the transformation of the space. In addition, the homography projection transformation can also enable the model to observe features from different perspectives and scales. That is, the homography projection encoding vectors of the context before and after obtained by the homography projection transformation contain the context information before and after observed from different perspectives and scales, which enables the model to comprehensively consider more factors when judging the operation category label of the process to be recognized and can reduce misjudgments caused by incomplete or single-perspective information.

[0043] Specifically, in the embodiment of the present application, the step S162 includes: calculating the forward context semantic attention score field of the context before the process to be recognized relative to the context after the process to be recognized, and this process can be expressed as: ; where is the homography projection encoding vector of the context before the process to be recognized, is the homography projection encoding vector of the context after the process to be recognized, is 's transposed vector, is matrix multiplication, is 's length, It is the forward context semantic attention score field of the process to be recognized before and after the text.

[0044] Calculate the reverse context semantic homography projection coding vector of the context after the process to be recognized relative to the context semantic homography projection coding vector before the process to be recognized, and the reverse context semantic attention score field of the forward and backward context of the process to be recognized can be expressed as: ; where is the context semantic homography projection coding vector before the process to be recognized, is the context semantic homography projection coding vector after the process to be recognized, is 's transposed vector, is matrix multiplication, is 's length, is the reverse context semantic attention score field of the forward and backward context of the process to be recognized.

[0045] After feature splicing of the forward context semantic attention score field of the process to be recognized and the reverse context semantic attention score field of the forward and backward context of the process to be recognized, input it into a convolutional layer with a convolution kernel of 3×3 to obtain the bidirectional attention balance field of the forward and backward context semantic of the process to be recognized, and the process can be expressed as: ; where is the forward context semantic attention score field of the process to be recognized, is the reverse context semantic attention score field of the forward and backward context of the process to be recognized, is the feature splicing operation, is the convolutional coding with a convolution kernel of 3×3, is the bidirectional attention balance field of the forward and backward context semantic of the process to be recognized.

[0046] It should be understood that in complex workover operation scenarios, the text description covers a lot of information, from equipment usage, operation procedures to well condition details, etc. By calculating the forward context semantic attention score field of the operation to be recognized, the model can simulate the human visual system and quickly focus on the part of the context semantic homography projection coding vector in the previous context that is closely related to the vector in the subsequent context. For example, when judging whether an operation belongs to the operation of tripping special pipe strings, the information about the pipe string running-in depth and special connection methods in the subsequent vector is the key judgment basis. The forward attention mechanism enables the model to accurately locate relevant content such as special pipe string preparation and special tool allocation in the complex previous information, avoiding being interfered by other irrelevant information, thus providing strong support for accurately identifying the operation category. Moreover, the text data of workover operations may contain noisy information, which is not actually helpful for operation identification. The forward attention mechanism acts as an efficient filter, filtering out the noisy information that has nothing to do with the current operation judgment by quantifying the correlation degree of the previous and subsequent coding vectors, which can reduce the interference of noise on the recognition result and improve the accuracy and reliability of recognition.

[0047] Correspondingly, considering that the forward context semantic attention score field of the operation to be recognized mainly focuses on the importance distribution of the previous vector in the context of the subsequent vector, but this may not comprehensively capture all the relationships between the previous and subsequent contexts. In the text description of workover operation procedures, the subsequent vector also has important associations and influences on the previous vector. In this application, by calculating the reverse context semantic attention score field of the operation to be recognized, these information can be mined from the opposite direction to supplement the part not covered by the forward context semantic attention score field of the operation to be recognized, avoiding information omission. That is, through the supplementary information provided by the reverse context semantic attention score field of the operation to be recognized, the model can consider more comprehensive factors when identifying the operation category label of the operation, avoiding misjudgment caused by only considering the relationship in a single direction, and thus improving the accuracy of operation identification.

[0048] It should be understood that in workover operations in oil and gas fields, the judgment of processes requires comprehensive consideration of various complex information in the context. The forward and reverse context semantic attention score fields of the process to be recognized respectively reveal the relationship between the context semantic homography projection coding vectors of the previous and subsequent texts from different directions. First, through feature splicing, this information from different directions is completely integrated. For example, when judging whether a process belongs to the operation of tripping special pipe strings, the forward attention score field may highlight the selection and preparation of special pipe strings in the previous preparation work, while the reverse attention score field may emphasize the impact of this process on the use effect of special pipe strings in the subsequent process. After splicing the two, the model can obtain both aspects of information at the same time, and the understanding of the process is more comprehensive. Then, a convolutional layer with a convolution kernel of 3×3 is used to process the spliced features, which can extract more representative features and enhance the adaptability of the model to different data. That is, the generated forward and reverse context semantic bidirectional attention balance field of the process to be recognized enables the model to quickly adapt to the data characteristics when facing new workover operation data, accurately analyze the relationship between the previous and subsequent texts, and thus improve the generalization ability of the model. This means that whether it is a conventional workover operation or an operation in special circumstances, the model can stably and accurately identify the operation category label of the process.

[0049] Specifically, in the embodiment of the present application, the step S163 includes: mapping the context semantic homography projection coding vector of the previous text of the process to be recognized and the context semantic homography projection coding vector of the subsequent text of the process to be recognized to the forward and reverse context semantic bidirectional attention balance field of the process to be recognized to obtain the context semantic homography projection attention modulation coding vector of the previous text of the process to be recognized and the context semantic homography projection attention modulation coding vector of the subsequent text of the process to be recognized. This process can be expressed as: ; where is the context semantic homography projection coding vector of the previous text of the process to be recognized, is the context semantic homography projection coding vector of the subsequent text of the process to be recognized, is matrix multiplication, is the forward and reverse context semantic bidirectional attention balance field of the process to be recognized, is the context semantic homography projection attention modulation coding vector of the previous text of the process to be recognized, is the context semantic homography projection attention modulation coding vector of the subsequent text of the process to be recognized.

[0050] Calculate the element-wise division between the context semantic homography projection attention modulation coding vector of the previous text of the process to be recognized and the context semantic homography projection attention modulation coding vector of the subsequent text of the process to be recognized to obtain the bidirectional context semantic coding vector of the process to be recognized. This process can be expressed as: ; wherein, is the context semantic homography projection attention modulation coding vector of the previous context of the process to be recognized, is the context semantic homography projection attention modulation coding vector of the subsequent context of the process to be recognized, is the two-way context semantic coding vector of the process to be recognized.

[0051] Correspondingly, considering that workover operations in oil and gas fields involve many complex processes, the characteristics of each process not only include its own operation information, but also are closely related to the processes before and after. Although the original context semantic homography projection coding vectors of the previous and subsequent contexts of the process to be recognized each carry certain information, it is difficult to fully reflect the complex interaction relationships between the characteristics. In order to capture the complex activities between the characteristics, in this application, the context semantic homography projection coding vector of the previous context of the process to be recognized and the context semantic homography projection coding vector of the subsequent context of the process to be recognized are respectively mapped to the forward and reverse context semantic two-way attention balance field of the process to be recognized, so as to comprehensively consider the characteristics of the mutual influence of the previous and subsequent contexts by means of the balance field, and then deeply explore the complex connections between the characteristics. Through this process, the model can understand the feature relationship at a higher abstraction level. It is no longer limited to the surface operation description of the process, but can grasp the deeper logical connection between the processes. For example, when judging which operation category a complex process belongs to, the model can comprehensively consider the factors before and after from the perspective of the overall operation process and make a more accurate judgment, which helps to improve the accuracy of process classification.

[0052] It should be understood that workover operations involve a variety of complex processes. After the feature vectors of each process are modulated, they contain rich context correlation information. The element-wise division operation by position can deeply explore this information and reveal the new relationship between the two modulated feature vectors. For example, when judging whether a certain process belongs to the tubing running and pulling operation or the special tubing running and pulling operation, the previous vector may contain information related to tubing preparation work, and the subsequent vector contains operation details during the tubing running and pulling process. Through the element-wise division operation, the degree of association of information at each position between the previous and subsequent contexts can be clearly presented, helping the model to more accurately judge the characteristics of the process. Specifically, this element-wise division operation reveals the subtle proportional relationships between the features, and these proportional relationships provide key bases for the model's decision-making. In the workover operation scenario, different processes have differences in aspects such as equipment usage duration and complexity of operation steps, which are reflected in the feature vectors as different proportions of the element values in each dimension. For example, when distinguishing between flushing operations and rotary operations, the flushing operation may focus more on the proportional relationship between the working duration of the mud pump and the fluid flow rate, while the rotary operation pays more attention to the proportional relationship between the tubing rotation speed and the torque. The proportional relationships obtained by the element-wise division operation enable the model to accurately capture these differences, thereby more accurately identifying the process category.

[0053] In step S170, based on the bidirectional context semantic coding vector of the process to be identified, an identification result of the job category label for representing the i-th process is obtained. Specifically, in an embodiment of the present application, the step S170 includes: inputting the bidirectional context semantic coding vector of the process to be identified into a process identifier based on a classifier to obtain the identification result, and the identification result is the job category label of the i-th process. That is, the bidirectional context semantic coding vector of the process to be identified obtained by interactively responding with the preceding context semantic coding vector of the process to be identified and the following context semantic coding vector of the process to be identified is classified and processed to intelligently determine the job category to be identified. In particular, the job category labels include flushing operations, rotation operations, lifting and lowering tubular operations, lifting and lowering special tubular operations, empty wellbore, cable perforation, etc. In this way, managers can better plan resources, arrange manpower and technical equipment, and ensure that each process can be executed efficiently. In particular, in a specific embodiment of the present application, the bidirectional context semantic coding vector of the process to be identified is input into a process identifier based on a classifier to obtain the identification result, and the identification result is the job category label of the i-th process, including: using the fully connected layer of the classifier to fully connect the bidirectional context semantic coding vector of the process to be identified to obtain a bidirectional context semantic fully connected coding classification feature vector of the process to be identified; inputting the bidirectional context semantic fully connected coding classification feature vector of the process to be identified into the Softmax classification function of the classifier to obtain the identification result.

[0054] Preferably, the bidirectional context semantic encoding vector of the process to be identified is input into a process identifier based on a classifier to obtain an identification result, which includes: determining the label probability value corresponding to each job category label of the i-th process obtained by inputting the bidirectional context semantic encoding vector of the process to be identified into the process identifier based on a classifier , and calculate the probability value of each label The square root of the sum of the squares of is used to obtain the bidirectional context semantic probability value of the process to be identified: ;in, Represents the label probability value corresponding to each job category label, The number of labels representing the job category labels, Represents the bidirectional context semantic probability value of the process to be identified.

[0055] For each label probability value , the label probability value Subtract the bidirectional context semantic probability value of the process to be identified and divide it by the label probability value , and for all label probability values Sum to get the bidirectional context semantic encoding density gradient value of the process to be identified: ; wherein, represents the two-way context semantic coding density gradient value of the process to be recognized.

[0056] Multiply the feature mean and the feature variance of the two-way context semantic coding vector of the process to be recognized by the two-way context semantic coding density gradient value of the process to be recognized respectively to obtain the first two-way context semantic coding distribution field mapping value and the second two-way context semantic coding distribution field mapping value .

[0057] Calculate the dot product of the two-way context semantic coding vector of the process to be recognized and the first two-way context semantic coding distribution field mapping value as the exponent, and the exponential function with the natural constant as the base , and perform a dot product with the two-way context semantic coding vector of the process to be recognized to obtain the first two-way context semantic coding fine-grained response vector , wherein, represents the dot product.

[0058] Calculate the dot product of the two-way context semantic coding vector of the process to be recognized and the second two-way context semantic coding distribution field mapping value as the exponent, and the exponential function with the natural constant as the base , and perform a dot product with the two-way context semantic coding vector of the process to be recognized to obtain the second two-way context semantic coding fine-grained response vector .

[0059] Calculate the weighted sum of the first two-way context semantic coding fine-grained response vector and the second two-way context semantic coding fine-grained response vector to obtain the optimized two-way context semantic coding vector of the process to be recognized, wherein, represents the dot addition, and represent different weight hyperparameters.

[0060] Input the optimized two-way context semantic coding vector of the process to be recognized into a process recognizer based on a classifier to obtain a recognition result.

[0061] Since the semantic encoding vectors of the context before and after the operation to be recognized respectively represent the near - far bidirectional context - associated encoding semantic features of the i - th operation in the set of text descriptions of conventional workover operations in oil and gas fields with respect to the context before and after, during the interactive response based on the positive and negative attention fields of features, there will be a micro - macro interactive response representation deviation of the enhanced feature representation of the positive and negative attention fields of the encoding semantics between local semantic spatial domains compared to the global semantic spatial domain. Thus, when inputting the bidirectional context semantic encoding vector of the operation to be recognized into the operation recognizer based on a classifier, it causes a dynamic deviation in the mapping of features to class target probabilities, reducing the accuracy of the obtained recognition result.

[0062] Therefore, by calculating the probability density gradient features of the target sequence corresponding to the bidirectional context semantic encoding vector of the operation to be recognized and using them as the input to the distributed local feature aggregation unit, a collaborative mapping of the distribution field features of the bidirectional context semantic encoding vector of the operation to be recognized can be constructed based on the heterogeneous core representation framework. Thereby, enhancing the response ability of the fine - grained pattern representation of the global semantic dependence of the bidirectional context semantic encoding vector of the operation to be recognized to class probability modeling. That is, by coordinating the global semantic spatial correlation relationships of different - scale features of the bidirectional context semantic encoding vector of the operation to be recognized, to strengthen the non - linear transformation mechanism of the high - dimensional embedding into the probability space during the feature - class probability mapping optimization process, so as to ensure that on the feature space - class probability mapping path, the fine - grained pattern representation and the class probability distribution always maintain an iterative optimization dynamic balance, thereby improving the accuracy of the recognition result obtained by inputting the bidirectional context semantic encoding vector of the operation to be recognized into the operation recognizer based on a classifier.

[0063] In summary, the method for classifying the working conditions of conventional workover operations in oil and gas fields based on the embodiments of the present application is elucidated. It uses deep - learning - based data - processing techniques to decompose and semantically encode the text descriptions of conventional workover operations in oil and gas fields. Then, the operation semantic features corresponding to the i - th operation are extracted from the set of encoded operation semantic features as the semantic features of the operation to be recognized, and the entire set is divided into the sets of the operation semantic features of the context before and after using it as the dividing line. Then, the semantic features of the operation to be recognized are context - encoded with the sets of the operation semantic features of the context before and after respectively, and based on the positive and negative attention response representations between the context semantic features before and after the operation to be recognized, the operation category of the operation to be recognized is intelligently determined. In this way, the efficiency and accuracy of operation classification can be effectively improved.

Claims

1. A classification method for the working conditions of conventional workover operations in oil and gas fields, characterized in that, Including: Obtain the text description of the conventional workover operation in the oil and gas field; decompose and semantically encode the text description of the conventional workover operation in the oil and gas field to obtain a set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field; extract the semantic encoding vector of the conventional workover operation process corresponding to the i-th process from the set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field as the semantic encoding vector of the process to be recognized; using the semantic encoding vector of the process to be recognized as a boundary, divide the set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field into a set of semantic encoding vectors of the previous conventional workover operation processes in the oil and gas field and a set of semantic encoding vectors of the subsequent conventional workover operation processes in the oil and gas field; respectively perform context encoding on the set of semantic encoding vectors of the previous conventional workover operation processes in the oil and gas field and the set of semantic encoding vectors of the subsequent conventional workover operation processes in the oil and gas field to obtain a context semantic encoding vector of the previous context of the process to be recognized and a context semantic encoding vector of the subsequent context of the process to be recognized; perform positive and negative attention field interaction response on the context semantic encoding vector of the previous context of the process to be recognized and the context semantic encoding vector of the subsequent context of the process to be recognized to obtain a two-way context semantic encoding vector of the process to be recognized, wherein the positive and negative attention field interaction response is performed by constructing a positive and negative context semantic two-way attention balance field of the previous and subsequent contexts of the process to be recognized, and performing semantic mapping modulation and response encoding based on the constructed positive and negative context semantic two-way attention balance field of the previous and subsequent contexts of the process to be recognized; based on the two-way context semantic encoding vector of the process to be recognized, obtain an identification result for representing the operation category label of the i-th process.

2. The classification method for the working conditions of conventional workover operations in oil and gas fields according to claim 1, characterized in that, Decompose and semantically encode the text description of the conventional workover operation in the oil and gas field to obtain a set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field, including: performing process decomposition based on the construction stage on the text description of the conventional workover operation in the oil and gas field to obtain a set of text descriptions of the conventional workover operation processes in the oil and gas field; performing semantic encoding on each text description of the conventional workover operation processes in the set of text descriptions of the conventional workover operation processes in the oil and gas field to obtain the set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field.

3. The classification method for the working conditions of conventional workover operations in oil and gas fields according to claim 2, characterized in that, Performing semantic encoding on each text description of the conventional workover operation processes in the set of text descriptions of the conventional workover operation processes in the oil and gas field to obtain the set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field, including: using a semantic encoder including an embedding layer to perform semantic encoding on each text description of the conventional workover operation processes in the set of text descriptions of the conventional workover operation processes in the oil and gas field to obtain the set of semantic encoding vectors of the conventional workover operation processes in the oil and gas field.

4. The method for classifying the working conditions of conventional workover operations in oil and gas fields according to claim 3, wherein, Perform context encoding on the set of semantic encoding vectors of the aforementioned oil and gas field conventional workover operation processes and the set of semantic encoding vectors of the subsequent oil and gas field conventional workover operation processes respectively to obtain the context semantic encoding vector of the process to be recognized before and the context semantic encoding vector of the process to be recognized after, including: inputting the semantic encoding vector of the process to be recognized and the set of semantic encoding vectors of the aforementioned oil and gas field conventional workover operation processes into a unidirectional context encoder based on a forward LSTM model to obtain the context semantic encoding vector of the process to be recognized before; inputting the semantic encoding vector of the process to be recognized and the set of semantic encoding vectors of the subsequent oil and gas field conventional workover operation processes into a unidirectional context encoder based on a backward LSTM model to obtain the context semantic encoding vector of the process to be recognized after.

5. The classification method for the working conditions of conventional workover operations in oil and gas fields according to claim 4, characterized in that, Perform forward and backward attention field interaction response on the context semantic encoding vector of the process to be recognized before and the context semantic encoding vector of the process to be recognized after to obtain the bidirectional context semantic encoding vector of the process to be recognized, including: performing homography projection transformation on the context semantic encoding vector of the process to be recognized before and the context semantic encoding vector of the process to be recognized after to obtain the homography projection encoding vector of the context semantic of the process to be recognized before and the homography projection encoding vector of the context semantic of the process to be recognized after; constructing a forward and backward context semantic bidirectional attention balance field between the homography projection encoding vector of the context semantic of the process to be recognized before and the homography projection encoding vector of the context semantic of the process to be recognized after; based on the forward and backward context semantic bidirectional attention balance field, performing semantic mapping modulation and response encoding on the homography projection encoding vector of the context semantic of the process to be recognized before and the homography projection encoding vector of the context semantic of the process to be recognized after to obtain the bidirectional context semantic encoding vector of the process to be recognized.

6. The classification method for the working conditions of conventional workover operations in oil and gas fields according to claim 5, characterized in that, Construct a forward and backward context semantic bidirectional attention balance field between the homography projection encoding vector of the context semantic of the process to be recognized before and the homography projection encoding vector of the context semantic of the process to be recognized after, including: calculating the forward context semantic attention score field of the homography projection encoding vector of the context semantic of the process to be recognized before relative to the homography projection encoding vector of the context semantic of the process to be recognized after; calculating the backward context semantic attention score field of the homography projection encoding vector of the context semantic of the process to be recognized after relative to the homography projection encoding vector of the context semantic of the process to be recognized before; performing feature splicing on the forward context semantic attention score field and the backward context semantic attention score field and then inputting them into a convolutional layer with a convolutional kernel of 3×3 to obtain the forward and backward context semantic bidirectional attention balance field.

7. The classification method for the working conditions of conventional workover operations in oil and gas fields according to claim 6, characterized in that, Based on the front-back context semantic bidirectional attention balance field of the to-be-recognized process before and after, perform semantic mapping modulation and response coding on the to-be-recognized process front context semantic homography projection coding vector and the to-be-recognized process back context semantic homography projection coding vector to obtain the to-be-recognized process bidirectional context semantic coding vector, including: mapping the to-be-recognized process front context semantic homography projection coding vector and the to-be-recognized process back context semantic homography projection coding vector to the front-back context semantic bidirectional attention balance field of the to-be-recognized process before and after respectively to obtain the to-be-recognized process front context semantic homography projection attention modulation coding vector and the to-be-recognized process back context semantic homography projection attention modulation coding vector; calculating the division of the to-be-recognized process front context semantic homography projection attention modulation coding vector and the to-be-recognized process back context semantic homography projection attention modulation coding vector at each position point to obtain the to-be-recognized process bidirectional context semantic coding vector.

8. The method for classifying the working conditions of conventional workover operations in oil and gas fields according to claim 7, wherein, Based on the to-be-recognized process bidirectional context semantic coding vector, obtain the recognition result for representing the operation category label of the i-th process, including: inputting the to-be-recognized process bidirectional context semantic coding vector into a process recognizer based on a classifier to obtain the recognition result, and the recognition result is the operation category label of the i-th process.

Images