Plunger gas lift process zero sample working condition diagnosis method

Abstract

The invention belongs to the field of oil and gas well exploitation technology, and relates to a plunger gas lift technology zero sample working condition diagnosis method, which comprises the processes of data preprocessing, model construction and evaluation test, and comprises a data preprocessing stage of abandoning manual data labeling, and describing text data by using a natural language in a production daily report on a production site; the quality of input data is ensured through preprocessing and enhanced optimization; in the model construction stage, a cross-modal learning algorithm is adopted, a CLIP model is constructed, text data and sensor data are combined through the CLIP model, semantic information is learned from a text through combined learning of image and text features, and effective association between the image and the text is achieved; in the evaluation test stage, image-text multi-mode learning is introduced into the field of plunger gas lift working condition diagnosis, and high-precision and zero-sample diagnosis of unseen new well working conditions is achieved by constructing a data organization mode, namely image-text pairs, with a production daily report text as key input.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas well development technology, and in particular relates to a zero-sample working condition diagnosis method for plunger gas lift technology. It is mainly used for plunger gas lift in oil and gas wells to maintain the production capacity of oil and gas wells and ensure normal and stable production. Background Technology

[0002] With the continued growth of global natural gas demand, the problem of liquid accumulation during gas field development is becoming increasingly prominent. In the later stages of gas well production, decreased formation pressure reduces the gas's liquid-carrying capacity, leading to liquid accumulation at the well bottom and severely impacting normal gas well production. Studies have shown that liquid accumulation has become one of the main factors restricting normal gas well production (e.g., ...). Figure 1 The figure shown is a bar chart illustrating the influencing factors and their extent on the productivity of a water-gas field in a certain region of China. To address this common technical challenge, plunger gas lift, as an economical and efficient artificial lifting technology, has been widely applied globally.

[0003] Plunger lift technology is a widely used artificial lift method in oil and gas well production. This technology utilizes the energy of gas to carry fluid accumulated at the bottom of the well to the surface by periodically reciprocating a plunger within the tubing, achieving continuous and efficient oil and gas well production. In actual production, the operating status of the plunger lift system directly affects the production efficiency and equipment lifespan of the oil and gas well. A typical plunger lift cycle includes four main stages: shut-in gas storage, plunger descent, plunger ascent, and fluid discharge, such as... Figure 2 As shown. However, due to complex formation conditions and variable production environment, plunger lifting systems often experience various faults such as plunger stagnation, gas-liquid ratio imbalance, and plunger wear. These faults not only reduce production efficiency but may also lead to equipment damage and safety hazards. Early and existing mainstream fault diagnosis methods rely heavily on the experience of field operators, inferring downhole conditions by observing the fluctuation characteristics of surface parameters such as oil pressure, casing pressure, and temperature. This method has inherent defects such as strong subjectivity, delayed response, and difficulty in detecting minor potential faults. With the advancement of information technology and the development of equipment automation and intelligent technology, deep learning methods, especially neural networks, have been widely used in plunger lifting fault diagnosis due to their powerful feature extraction capabilities (LECUN Y, BOSER B, DENKER J, et al. Backpropagation Applied to Handwritten Zip Code Recognition[J]. Neural Computation, 1989, 1(4): 541-551.).

[0004] Neural networks are a part of the field of deep learning research. For example... Figure 3As shown, a neural network mainly consists of an input layer, hidden layers, and an output layer. Each layer contains multiple neurons, each with two parameters: weights and biases, as well as the corresponding activation function for that layer. The input undergoes layer-by-layer calculations to produce an output with a non-linear relationship, which can better solve complex non-linear problems in the real world. During training, the neural network propagates the calculation results of each layer forward, calculates the error based on the predicted results and the actual labels, and propagates the error back to the neurons layer by layer. The parameters on each neuron are updated according to the gradient algorithm to minimize the error, achieving model convergence.

[0005] Currently, scholars both domestically and internationally have conducted extensive research on the design, construction, production optimization, and development of supporting tools for plunger wells. Zhang Ting et al. (Zhang Ting, Tang Hanbing, Zhu Peng, et al. Research and application of plunger gas lift drainage and gas production technology in low-pressure deep wells [J]. Drilling and Production Technology, 2021, 44 (06): 124-128.) proposed a complete design method for plunger gas lift in low-pressure gas wells with depths of 4000 meters or more, addressing the challenges of plunger gas lift applications. By adding influencing factors such as wellbore inflow dynamics and casing pressure changes, a new plunger start-up pressure guidance chart was established. Simultaneously, a self-sealing low-leakage plunger was developed, reducing leakage by 20-30%, and the surface pipeline system was optimized, simplifying the diaphragm valve connection method. Field application results show that this method not only expands the applicable scope of the process but also increases single-well gas production by more than 20%, with a maximum increase of 6 times. Lei Wei, considering the special characteristics of plunger gas lift in horizontal wells, studied a process optimization scheme under high condensate oil content conditions. A special plunger tool system suitable for horizontal wells was developed, and an innovative anti-jamming structure was designed to solve the problem of traditional tools easily jamming in the horizontal section. Meanwhile, a parameter optimization model for horizontal well plunger gas lift was established, achieving a significant improvement in lift efficiency; Nandola (NANDOLA NN, KAISARE NS, GUPTA A. Online optimization for a plunger liftprocess in shale gas wells[J]. Computers & Chemical Engineering, 2018, 108:89-97.) et al. proposed a novel step-by-step optimization method to optimize the plunger lift process in shale gas wells, addressing the modeling and optimization problem of the plunger lift system. The effectiveness of the method was verified through simulation, and compared with the time-based controller used in the traditional method, proving that the performance of the method is superior to that of the time-based controller, which can keep the plunger speed within the designed set value range, while maximizing net production; Liu Xiong Hui (Liu Xiong Hui, Wang Xiao Hui, Wang Su Kai, et al. Application effect of plunger drainage gas production in Sulige high water-cut tight sandstone gas reservoir[J]. China Petroleum and Chemical Standards and Quality, 2021, 41 (19): (151-152+154.) To address the low development efficiency of the Sulige high water-cut tight sandstone gas reservoir, a study on the application of plunger drainage gas production technology was conducted. The characteristics of the gas reservoir under high water-cut conditions were systematically analyzed, and the structural design and operating parameters of the plunger tool were optimized.Field application results show that the optimized process can increase gas well production by more than 40% and reduce water cut by 15%, providing an effective technical solution for the development of high water-cut gas reservoirs. PCS Ferguson (BURROWS A. Sensor improvements for plunger lift wells[J]. World Pumps, 2018,(12):8) improved the 3DSO plunger arrival sensor for plunger lift wells, enabling it to accurately detect the arrival of the plunger and activate the controller to perform corresponding operations. This sensor represents a major improvement in plunger arrival sensing technology.

[0006] As can be seen from the above, scholars both domestically and internationally have made significant progress in the development of plunger gas lift technology. System optimization model and supporting equipment improvementWhile there are numerous successful cases in this area, research on intelligent diagnosis using multi-source heterogeneous data (especially text data) is still lacking. Particularly in the field of plunger gas lift, there is no precedent for using multimodal learning methods that combine images with text (daily production reports) for operational condition diagnosis. However, CNN-based single-modal image classification methods expose the following core problems in practical applications (i.e., the technical defects this invention aims to solve): Heavy reliance on large amounts of labeled data (sample scarcity problem): Traditional CNN model training requires massive amounts of manually labeled fault samples. In oil and gas fields, obtaining sufficient samples covering all fault types (especially rare faults) is extremely difficult and costly, leading to a sharp decline in model performance when facing "long-tailed" data. Poor cross-well generalization ability (zero-sample recognition failure): CNN models tend to "memorize" the feature distribution in the training set. When facing a new gas well (new well), due to differences in geological structure, wellbore diameter, and production parameters, its data features often exhibit significant "domain drift" compared to the training set. This results in the model performing poorly on unseen new well data, making zero-sample diagnosis impossible, meaning it cannot be directly applied without historical data for the new well. Low utilization of multi-source information (lack of semantic information): Existing intelligent diagnostic methods are almost entirely limited to single-modal analysis of sensor values ​​(time series) or their converted images (such as pressure curves). In actual production, textual data such as daily reports and operating condition descriptions contain rich expert experience and semantic information (e.g., "oil pressure and casing pressure decrease slowly and synchronously" directly corresponds to a certain fluid accumulation characteristic), but traditional CNNs cannot understand and utilize this unstructured textual knowledge, resulting in a lack of semantic support for diagnostic decisions and limited accuracy. Since CNNs mainly rely on feature representations learned from training data, their generalization ability may be limited when faced with unseen operating conditions or abnormal data. This means that the model may not be able to accurately identify new types of operating conditions in practical applications (Hu Haibin, Liu Renxin, Liu Rilong, et al. A review of the application of convolutional neural networks in mechanical fault diagnosis [J]. Mechanical Engineering and Automation, 2024, (04): 221-223.).

[0007] With the rapid development of computer vision and natural language processing technologies, cross-modal learning has become a research hotspot. Among them, the CLIP (Contrastive Language–Image Pre-training) model, as an innovative cross-modal learning method, has received widespread attention in recent years. The CLIP model achieves effective association between images and text by jointly learning image and text features. Its core idea is to use a large number of image-text pairs for contrastive learning, enabling the model to learn the semantic correspondence between images and text (Liu Jie, Qiao Wensheng, Zhu Peipei, et al. Zero-shot reference image segmentation based on fine-tuning of the large image-text model CLIP [J / OL]. Computer Applications Research, 1-8 [2024-11-07].), the model structure is shown in Figure 4 This allows the CLIP model to fully utilize the correlation information between images and text, improving classification accuracy. Furthermore, the CLIP model has the following advantages: because it utilizes a large number of image-text pairs for comparative learning, it can achieve accurate matching and recognition on unseen images and texts. This gives the CLIP model stronger generalization ability when faced with new working conditions or anomalous data (Wenbo Zhang, Yifan Zhang, Yuyang Deng, Ta-Adapter: Enhancing few-shot CLIP with task-aware encoders, PatternRecognition, Volume 153, 2024, 110559, ISSN 0031-3203.).

[0008] In summary, compared with traditional deep learning methods (such as CNN), the CLIP model has significant advantages in plunger data condition classification tasks. By fully utilizing the correlation information between images and text, the CLIP model can achieve more accurate and robust condition classification. Therefore, cross-modal learning is essential to solve the problems of plunger gas lift fault identification and zero-shot diagnostic identification. Currently, the field of intelligent diagnosis of plunger gas lift conditions faces two fundamental technical bottlenecks: existing deep learning methods heavily rely on large-scale, high-quality manually labeled datasets; and models exhibit significant generalization limitations when facing new wells with different geological conditions and production parameters. These two problems together constitute the core obstacle hindering the large-scale implementation of artificial intelligence technology in oil and gas fields. Therefore, it is necessary to conduct in-depth research and exploration of plunger gas lift process condition diagnosis methods. Summary of the Invention

[0009] This invention addresses the shortcomings of existing technologies by proposing a zero-shot diagnostic method for plunger gas lift processes. The fundamental objective of this invention is to pioneeringly introduce multimodal learning (ML) into the field of plunger gas lift process diagnostics. By constructing a novel data organization method—a "text-image pair"—with daily production report text as the key input, it achieves high-precision, zero-shot diagnostics of unseen well conditions.

[0010] In the process of industrial intelligence, the successful deployment of deep learning models highly depends on the quality and scale of their training data. For plunger air lift systems, establishing an accurate and reliable fault diagnosis model requires a large amount of sample data with precise labels. However, obtaining such high-quality labeled data in actual production environments faces severe challenges. First, accurate data labeling requires extensive field experience and professional knowledge. Only engineers familiar with the process flow and equipment principles can judge the actual operating conditions behind data fluctuations, creating a very high knowledge barrier. Second, manual labeling is a time-consuming and labor-intensive task. Analyzing and labeling millions of minutes of data one by one incurs enormous human and time costs. Finally, real fault events are low-probability events, accounting for a very small percentage of the massive amount of normal data, resulting in an extremely uneven distribution of labeled samples, making it difficult to meet the needs of model training. Therefore, traditional supervised learning models have fallen into a "data hunger" dilemma in the field of plunger diagnosis. The primary objective of this invention is to overcome this bottleneck. We abandon the reliance on expensive manually labeled data and instead utilize the "free resources" already existing in the production site, which contain rich semantic information—the natural language descriptions in daily production reports. These unstructured texts, written by frontline operators, are the most direct language describing the equipment's operating status. By combining these texts with sensor data, the model can learn semantic information from the text.

[0011] Meanwhile, the geological conditions of oil and gas fields are highly heterogeneous. Even within the same block, different gas wells exhibit significant differences in parameters such as reservoir pressure, permeability, water cut, well depth, and tubing structure. This means that the "feature representations" learned by a diagnostic model trained for well A may be completely inapplicable to well B, leading to a sharp drop in recognition accuracy. This is the so-called "transfer failure" problem. In actual production, gas fields often have hundreds or thousands of wells. If data needs to be re-collected, re-labeled, and the model re-trained for each new well, the efficiency will become extremely low. This presents a significant bottleneck for the large-scale application of existing AI diagnostic technologies. The second core objective of this invention is to solve this problem. Our proposed solution is to enable the model to learn to understand "semantics," rather than simply memorizing "patterns." Specifically, through a graph-text multimodal learning framework, we teach the model to understand the physical meaning of the phrase "valve abnormally open for a long time" and associate it with specific patterns in sensor data. When the model encounters a new well, even if its data distribution is completely different from the training set, as long as the semantic representation of "the valve remaining abnormally open for an extended period" is similar, the model can make the correct judgment based on its understanding of the "semantics." This method is essentially a form of "concept transfer," which doesn't care about the specific numerical values ​​of the data, but only about the "meaning" expressed by the data. This gives our model powerful zero-shot generalization capabilities, allowing it to be seamlessly applied to any new well without any additional annotation or retraining, thus greatly reducing the barriers and costs of technology promotion.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: A zero-sample operating condition diagnostic method for plunger gas lift process includes data preprocessing, model building, and evaluation testing. In the data preprocessing stage, manually labeled data is discarded, and natural language descriptive text data from daily production reports at the production site is used. Through preprocessing and enhancement optimization, the quality of input data is ensured. In the model building phase, a cross-modal learning algorithm is used to build the CLIP model (Contrastive Language–Image Pre-training). The CLIP model combines text data with sensor data, and learns semantic information from the text by jointly learning image and text features, thus realizing an effective association between images and text. During the evaluation and testing phase, Multimodal Learning was introduced into the field of plunger gas lift operation condition diagnosis. By constructing a data organization method with daily production report text as the key input—that is, “text-image pairs”—high-precision, zero-shot diagnosis of unseen new well operation conditions was achieved.

[0013] The zero-sample operating condition diagnostic method for plunger gas lift process includes the following steps in the data preprocessing stage: cleaning the original data and handling missing and duplicate values; using linear interpolation to handle missing data; and using the drop_duplicates function of the pandas library to remove duplicate data and retain the first occurrence of the data to maintain data integrity.

[0014] The zero-sample operating condition diagnostic method for the plunger gas lift process, in its data preprocessing stage, includes data labeling and classification, as well as data normalization and enhancement processes. Data labeling and classification: Establish a classification system containing six operating states: normal operation; extended valve closure anomaly; extended valve opening anomaly; improper parameter settings; other operating anomalies; data transmission anomalies; Data normalization and enhancement: a ratio-based normalization method; calculation of minute-by-minute ratios of oil pressure and casing pressure data; data segmentation using a sliding window method (window length 224, step size 48) to normalize and enhance the data.

[0015] The zero-sample operating condition diagnostic method for the plunger gas lift process, in the model building stage, includes the following model training steps: Model structure optimization: Integrating Vision Transformer as the visual backbone network; implementing cosine annealing learning rate decay; adding a novel data selector mechanism to ensure non-repetitive sampling of training batches; Split the dataset for training: Divide the dataset into training and test sets in a ratio of 0.8:0.2; set the small sample size to 6, the learning rate to 1e-4, the loss function to contrastive loss, and the optimizer to Adam; after training, test the model using data from 7 untrained wells, and after evaluation, save the model weights, and the model training is complete.

[0016] The zero-sample operating condition diagnostic method for the plunger gas lift process includes an evaluation and testing process, encompassing model performance evaluation, classification accuracy optimization, and zero-sample testing for unknown wells. Model performance evaluation involves plotting the loss-accuracy curve during training and generating a confusion matrix based on historical test data. Precision, recall, and F1 score are calculated using the confusion matrix as quantitative evaluation metrics. Simultaneously, t-SNE (t-Distributed Stochastic Neighbor Embedding) technology is employed to perform dimensionality reduction and visualization analysis on the high-dimensional features extracted by the model. The model's feature extraction capability is verified by observing the clustering distribution of features in the low-dimensional space. To optimize classification accuracy, for the visual encoder, the ViT-B / 16 architecture was selected as the optimal backbone network after comparative verification; for the text encoder, "detailed working condition description" containing rich semantic information was used as the input prompt word; in terms of training strategy, data augmentation techniques such as image rotation, cropping and flipping were introduced, and a cosine annealing strategy was used to dynamically adjust the learning rate to balance the convergence speed and final generalization performance of the model. In the zero-sample testing process for unknown wells, completely new gas well data that has not participated in model training is selected as the zero-sample test set. The test set data is preprocessed using the same ratio normalization, sliding window segmentation, and grayscale matrix conversion as in the training phase. The processed test images are then input into the trained CLIP model, which extracts image feature vectors and predefined text feature vectors for six different operating conditions. The cosine similarity between the image features and each text feature is calculated. Finally, the operating condition category with the highest similarity is selected as the diagnostic result for the test sample.

[0017] Beneficial effects of the invention: 1. This invention's zero-shot diagnostic method for plunger gas lift technology significantly improves the accuracy of identifying operating conditions in new wells. In practical applications at the Chongqing Gas Field, this invention demonstrated superior zero-shot diagnostic capabilities. Testing on seven "new wells" that had never participated in training, the model achieved an average accuracy of 0.6966 on production data from April to September 2022, representing a 73.4% performance improvement compared to traditional CNN methods relying solely on sensor data (accuracy 0.4017). This significant improvement fully validates the effectiveness of using daily production reports as the core input. Crucially, when identifying critical hazardous conditions such as "abnormal extended valve closure" and "improper parameter settings," the model's recall and precision remained consistently above 0.86. This means it can not only accurately capture the vast majority of real faults but also effectively avoid false alarms, providing a highly reliable decision-making basis for on-site emergency response.

[0018] 2. The zero-sample operational condition diagnostic method for plunger gas lift process of this invention provides zero-sample diagnostic capabilities, which is a core link in realizing intelligent monitoring of the entire oilfield and its entire life cycle, laying a solid foundation for building a safer, more efficient, and greener modern oil and gas production system. At the industry level, this invention has pioneering strategic significance. It is the first successful application of multimodal learning in the oil and gas field, providing a highly valuable paradigm for other industrial scenarios such as chemical, power, and metallurgy, proving that integrating unstructured text and structured sensor data is a feasible path to solve the bottlenecks in the implementation of industrial AI. It signifies that industrial intelligence is evolving from the traditional "data-driven" to a more advanced "semantic-driven" approach, that is, from simply mining data patterns to understanding the physical meaning and operational logic behind the data. As a key technological support for the construction of smart oilfields, the effects of this invention are comprehensive and multi-layered. It is not only a technological breakthrough but also an innovative achievement that can effectively solve industry pain points and create enormous economic and social value.

[0019] 3. This invention's zero-sample operating condition diagnostic method for plunger gas lift processes effectively solves the problems of class imbalance and gradient conflict in training with small-sample industrial data by introducing an innovative "Data Selector" mechanism. Existing technologies, when processing industrial data such as plunger gas lift, are often limited by the scarcity of anomalous samples (long-tail distribution), leading to a tendency for model training to "memorize" majority class samples. This invention integrates a data selector at the CLIP model training front end, forcing the uniqueness and class diversity of data within each training batch, fundamentally eliminating the calculation bias of contrast loss caused by sample duplication. Experimental data shows that compared to the original CLIP model, after introducing the data selector, the model's recall rate significantly increased from 0.56 to 0.77, and the F1 score increased from 0.55 to 0.76. This mechanism significantly improves the model's learning efficiency and recognition accuracy for rare fault samples, ensuring that the diagnostic system maintains high robustness when facing complex and ever-changing production data.

[0020] 4. The zero-sample operating condition diagnosis method for plunger gas lift process of this invention adopts data preprocessing techniques of "ratio normalization" combined with "grayscale matrixing," which significantly reduces the model's dependence on gas well physical parameters (such as well depth and pipe diameter), improving the algorithm's versatility and deployment efficiency. Traditional methods usually directly use the original pressure values ​​or simple maximum-minimum normalization, making the model susceptible to the influence of specific gas well base pressure values ​​(static pressure), and difficult to migrate between wells with different pressure systems. This invention innovatively calculates the ring ratio of oil pressure and casing pressure (i.e., the ratio of the current moment to the previous moment) and converts it into a grayscale image. This processing method effectively eliminates the interference of absolute pressure differences between wells and highlights the trend characteristics of operating condition changes. This technical feature allows the model to be deployed quickly and "plug and play" without retraining or fine-tuning for each new well, greatly saving data governance costs and model iteration cycles in field engineering applications, and significantly improving the intelligent efficiency of production management. Attached Figure Description

[0021] Figure 1 The figure shown is a bar chart illustrating the factors and their impact on the productivity of a water and gas field in a certain region of China. Figure 2 The diagram shows the working process of a plunger air lift. Figure 3 The diagram shown is a schematic of a neural network structure. Figure 4 The image shows the Clip model structure for the cross-modal learning method. Figure 5 The diagram shows the improved CLIP network structure of this invention; Figure 6 The diagram shown is a flowchart of the zero-sample operating condition diagnosis method for the plunger gas lift process of the present invention (overall technical route). Figure 7 The image shows the loss accuracy curve and confusion matrix; Figure 8 The results shown are the predictions from CNN and CLIP. Detailed Implementation

[0022] To make the technical concept and advantages of the invention clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are merely preferred embodiments for explaining and illustrating the present invention, and should not be considered as, nor constitute a limitation on, the scope of patent protection claimed by the present invention.

[0023] Example 1: See Figure 6 The present invention provides a zero-sample operating condition diagnostic method for plunger gas lift process, comprising data preprocessing, model building, and evaluation testing. In the data preprocessing stage, manually labeled data is abandoned. Natural language descriptive text data from daily production reports at the production site is used for preprocessing and enhancement optimization to ensure the quality of input data. Clean the original data, handling missing and duplicate values; use linear interpolation to handle missing data; use the drop_duplicates function of the pandas library to remove duplicate data, retaining the first occurrence of the data to maintain data integrity; Data labeling and classification: A classification system was established that includes six categories of operating states: normal operation; extended valve closure anomaly; extended valve opening anomaly; improper parameter settings; other operating anomalies; and data transmission anomalies. Data normalization and enhancement: Implement a ratio-based normalization method; calculate the minute-by-minute ratio of oil pressure and casing pressure data; use a sliding window method for data segmentation (window length 224, step size 48). In the model building phase, a cross-modal learning algorithm is used to build the CLIP model (Contrastive Language–Image Pre-training). The CLIP model combines text data with sensor data, and learns semantic information from the text by jointly learning image and text features, thus realizing an effective association between images and text. During the evaluation and testing phase, Multimodal Learning was introduced into the field of plunger gas lift operation condition diagnosis. By constructing a data organization method with daily production report text as the key input—that is, “text-image pairs”—high-precision, zero-shot diagnosis of unseen new well operation conditions was achieved.

[0024] Example 2: The zero-sample working condition diagnosis method for the plunger gas lift process in this example differs from that in Example 1 in that: further, in the model building stage, the model structure is optimized by: integrating Vision Transformer as the visual backbone network; implementing cosine annealing learning rate decay; and adding a novel data selector mechanism to ensure non-repetitive sampling of training batches. Split the dataset for training: Divide the dataset into training and test sets in a ratio of 0.8:0.2; set the small sample size to 6, the learning rate to 1e-4, the loss function to contrastive loss, and the optimizer to Adam; After training, the model was tested using data from seven untrained wells. After evaluation, the model weights were saved, and the model training was complete.

[0025] Example 3: The zero-sample operating condition diagnostic method for the plunger gas lift process in this example differs from Examples 1 and 2 in that it further evaluates the testing process, including model performance evaluation, classification accuracy optimization, and zero-sample testing for unknown wells. Model performance evaluation involves plotting the loss-accuracy curves during training and generating a confusion matrix based on historical test data. Precision, recall, and F1 score are calculated using the confusion matrix as quantitative evaluation metrics. Simultaneously, t-SNE (t-Distributed Stochastic Neighbor Embedding) technology is employed to perform dimensionality reduction and visualization analysis on the high-dimensional features extracted by the model. The feature extraction capability of the model is verified by observing the clustering distribution of features in the low-dimensional space. To optimize classification accuracy, for the visual encoder, the ViT-B / 16 architecture was selected as the optimal backbone network after comparative verification; for the text encoder, "detailed working condition description" containing rich semantic information was used as the input prompt word; in terms of training strategy, data augmentation techniques such as image rotation, cropping and flipping were introduced, and a cosine annealing strategy was used to dynamically adjust the learning rate to balance the convergence speed and final generalization performance of the model. In the zero-sample testing process for unknown wells, completely new gas well data that has not participated in model training is selected as the zero-sample test set. The test set data is preprocessed using the same ratio normalization, sliding window segmentation, and grayscale matrix conversion as in the training phase. The processed test images are then input into the trained CLIP model, which extracts image feature vectors and predefined text feature vectors for six different operating conditions. The cosine similarity between the image features and each text feature is calculated. Finally, the operating condition category with the highest similarity is selected as the diagnostic result for the test sample.

[0026] Example 4: The zero-sample operating condition diagnostic method for the plunger gas lift process in this example differs from the previous examples in that it uses a gas well in the Chongqing Gas Field as an example. Regarding model construction, it illustrates the CLIP-improved framework, which includes a visual encoder and a text encoder. In the visual encoder part, after comparative experiments, the VisionTransformer (ViT-B / 16) was ultimately selected as the optimal backbone network because it maintains high performance while exhibiting better inference efficiency compared to larger models (such as ViT-L / 14). Figure 5 The diagram shows the core of this invention, namely the improved CLIP framework structure.

[0027] In the text encoder section, several approaches to representing text labels were explored, including "simple operation identifier," "this is the running status of <category>," and "detailed operation description." Experiments showed that the "detailed operation description" format provides richer semantic information, enabling the model to perform best when recognizing complex operating conditions. Therefore, it was selected as the default configuration, as shown in Table 1.

[0028]

[0029] In terms of training strategy, to address the problem of model convergence difficulties caused by data duplication within batches during training on small datasets, this invention introduces a data selector mechanism to ensure that data in each batch is unique. Particularly in the plunger dataset, this mechanism guarantees that each training iteration covers six different working conditions to support effective contrastive learning. Furthermore, to further improve model performance, a cosine annealing learning rate decay strategy is employed. This strategy maintains a high learning rate in the early stages of training to accelerate convergence, and gradually reduces the learning rate in the later stages to fine-tune model parameters, thus achieving the optimal balance between convergence speed and final performance. The optimizer chosen is Adam, whose adaptive learning rate adjustment feature helps the model train stably and efficiently. Model training results are shown in [link to documentation]. Figure 7 .

[0030] The new data undergoes the same preprocessing steps as in the pre-training phase before being fed into CLIP and the traditional CNN model. Potential predicted labels are provided to the CLIP model, and accuracy is calculated by comparing these predicted labels with the ground truth labels. Figure 8 A comparison of the prediction results of the CNN and CLIP models on three representative wells.

[0031] like Figure 8 As shown, by comparing the three color bars of "True Label (Label_ori)," "CNN Prediction Result (CNN_pre)," and "CLIP Prediction Result (CLIP_pre)," the performance difference between the two models on the unseen well can be intuitively observed: Traditional CNN model performance: The predicted color bands of CNN models exhibit significant fragmentation and discontinuity, with a large amount of noise and misclassified regions. Especially during the transition phase of changing operating conditions, CNNs often fail to react accurately, and may even deviate completely from the actual operating conditions for some long periods. This indicates that traditional CNN models suffer from a serious "overfitting" phenomenon, that is, they rely too much on the data distribution of the training set wells, resulting in severely insufficient generalization ability when faced with new wells with different physical characteristics.

[0032] Improved CLIP Model Performance: In comparison, the CLIP model proposed in this invention exhibits extremely high consistency between the predicted color bands and the actual label color bands. Whether in stable normal operation or under complex abnormal conditions (such as fluid accumulation or prolonged well shut-in), the CLIP model provides continuous and accurate judgments. The smoothness and accuracy of its prediction results far surpass those of CNN models, demonstrating its superior feature extraction and matching capabilities in zero-shot scenarios.

[0033] In summary, the experimental results strongly demonstrate the superiority of using "detailed operational description" text as semantic guidance. The CLIP model does more than just match image pixels; it understands the physical semantics behind the operating conditions. The rich semantic information provided by the text encoder acts as a "universal anchor point" for cross-well diagnostics, enabling the model to accurately identify the corresponding operating condition category based on the semantic description, even when faced with pressure curves from newly discovered wells that have never been seen before.

[0034] The improved CLIP framework proposed in this invention successfully overcomes the dependence of traditional deep learning models on large amounts of labeled data and identically distributed data. Even without fine-tuning using new well data, the CLIP model maintains high-precision diagnostic capabilities (as shown in the aforementioned experimental data, the average accuracy improved from 0.40 for CNN to over 0.69). This indicates that the method possesses extremely strong robustness and transferability. This "plug-and-play" zero-shot diagnostic capability solves the "cold start" problem of lacking labeled samples in the early stages of new well production in industrial fields. This technical solution can significantly reduce data labeling costs in oilfields, shorten model deployment cycles, and provide an efficient, reliable, and low-cost solution for intelligent management and full lifecycle monitoring of gas fields.

[0035] In CLIP model training, when switching to small-scale datasets, duplicate data may appear in each batch, posing a challenge to the model's learning process. Specifically, duplicate data can cause the model to treat multiple identical images under the same label as equivalent when calculating cosine similarity, forcing the contrast loss on the diagonal to be 1, thus failing to accurately distinguish the loss for other identical data locations, thereby affecting the model's convergence performance. To address this issue, we introduce a data selector mechanism before training the model to ensure that the data in each batch is unique. Especially in the plunger dataset, this mechanism guarantees that each training session covers six different working conditions to support effective contrastive learning.

[0036] The research data for this invention comes from the plunger lift system of the Chongqing Gas Field, including real-time production data from 21 wells, covering key sensor information such as oil pressure, casing pressure, gas transmission pressure, and production rate. Four wells (Wells 1, 5, 8, and 51) were selected as the main research objects, providing over 680,000 minute-level data points from January 2020 to April 2021. Data from the remaining 17 wells, covering only the period from April to September 2022, were designated as zero-sample test cases. The PyTorch deep learning framework in Python was used to build and train the clip model. Through comparative learning of the feature relationships between different modes within the clip, the plunger lift operating conditions of unknown wells were identified.

[0037] The specific training steps for the model are as follows: 1. Data preprocessing: Clean the raw data, handle missing and duplicate values; use linear interpolation to handle missing data; use the drop_duplicates function of the pandas library to remove duplicate data, keeping the first occurrence of the data to maintain data integrity; 2. Data Labeling and Classification: A classification system was established that includes six categories of operating states: normal operation; extended valve closure anomaly; extended valve opening anomaly; improper parameter settings; other operating anomalies; and data transmission anomalies. 3. Data Normalization and Enhancement: Implement a ratio-based normalization method; calculate the minute-by-minute ratio of oil pressure and casing pressure data; and use a sliding window method for data segmentation (window length 224, step size 48). 4. Model structure optimization: Integrate Vision Transformer as the visual backbone network; implement cosine annealing learning rate decay; add a novel data selector mechanism to ensure non-repetitive sampling of training batches; 5. Split the dataset for training: Divide the dataset into training and test sets in a ratio of 0.8:0.2; set the small sample size to 6, the learning rate to 1e-4, the loss function to contrastive loss, and the optimizer to Adam; 6. After training, data from 7 untrained wells were used for testing (data from April to September 2022). 7. After the evaluation is completed, save the model weights, and the model training is complete.

[0038] This invention significantly improves the accuracy of identifying new well operating conditions. In practical applications at the Chongqing Gas Field, this invention demonstrates superior zero-shot diagnostic capabilities. Through testing on seven "new wells" that had never participated in training, the model achieved an average accuracy of 0.6966 on production data from April to September 2022, representing a 73.4% performance improvement compared to traditional CNN methods relying solely on sensor data (accuracy 0.4017). This significant improvement fully validates the effectiveness of using daily production reports as the core input. Crucially, when identifying critical hazardous conditions such as "abnormal extended valve closure" and "improper parameter settings," the model's recall and precision remained consistently above 0.86. This means it can not only accurately capture the vast majority of real faults but also effectively avoid false alarms, providing a highly reliable decision-making basis for on-site emergency response.

[0039] At the industry level, this invention has pioneering strategic significance. It represents the first successful application of multimodal learning in oil and gas field production, providing a highly valuable paradigm for other industrial scenarios such as chemical, power, and metallurgy. It demonstrates that integrating unstructured text with structured sensor data is a feasible path to overcome the bottlenecks in industrial AI implementation. It signifies that industrial intelligence is evolving from traditional "data-driven" to a more advanced "semantic-driven" approach, moving from simply mining data patterns to understanding the physical meaning and operational logic behind the data. As a key technological support for the construction of smart oilfields, the zero-sample diagnostic capability provided by this invention is a core element in achieving intelligent monitoring of the entire oilfield and its entire lifecycle, laying a solid foundation for building a safer, more efficient, and greener modern oil and gas production system.

[0040] In summary, the effects of this invention are comprehensive and multi-layered. It is not only a technological breakthrough, but also an innovative achievement that can effectively solve industry pain points and create huge economic and social value.

Claims

1. A zero-sample operating condition diagnostic method for a plunger gas lift process, comprising data preprocessing, model building, and evaluation testing, characterized in that: In the data preprocessing stage, manually labeled data is abandoned. Natural language descriptive text data from daily production reports at the production site is used for preprocessing and enhancement optimization to ensure the quality of input data. In the model building phase, a cross-modal learning algorithm is used to build the CLIP model. The CLIP model combines text data with sensor data and learns semantic information from the text by jointly learning image and text features, thus realizing an effective association between images and text. During the evaluation and testing phase, multimodal learning of images and text was introduced into the field of plunger gas lift operation condition diagnosis. By constructing a data organization method with daily production report text as the key input—that is, "image-text pairs"—high-precision, zero-sample diagnosis of unseen new well operation conditions was achieved.

2. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 1, characterized in that: In the data preprocessing stage, the original data is cleaned to handle missing and duplicate values; a linear interpolation method is used to handle missing data; and the drop_duplicates function of the pandas library is used to remove duplicate data, retaining the first occurrence of the data to maintain data integrity.

3. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 2, characterized in that: The data preprocessing stage includes data labeling and classification, as well as data normalization and enhancement. Data labeling and classification: Establish a classification system containing six operating states: normal operation; extended valve closure anomaly; extended valve opening anomaly; improper parameter settings; other operating anomalies; Data transmission error; Data normalization and enhancement: Ratio-based normalization method; calculation of minute-by-minute ratios of oil pressure and casing pressure data; data segmentation using a sliding window method, followed by data normalization and enhancement.

4. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 1, 2 or 3, characterized in that: During the model building phase, the model training steps are as follows: Model structure optimization: Integrating Vision Transformer as the visual backbone network; implementing cosine annealing learning rate decay; adding a novel data selector mechanism to ensure non-repetitive sampling of training batches; Split the dataset for training: Divide the dataset into training and test sets in a ratio of 0.8:0.2; set the small sample size to 6, the learning rate to 1e-4, use contrastive loss as the loss function, and use Adam as the optimizer.

5. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 4, characterized in that: After training, the model was tested using data from seven untrained wells. After evaluation, the model weights were saved, and the model training was complete.

6. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 1, 2, 3 or 5, characterized in that: The evaluation and testing process includes model performance evaluation, classification accuracy optimization, and zero-shot testing for unknown wells. Model performance evaluation involves plotting the loss-accuracy curves after model training and generating a confusion matrix based on historical test data. Precision, recall, and F1 score are calculated from the confusion matrix as quantitative evaluation metrics. Simultaneously, t-SNE technology is used to perform dimensionality reduction and visualization analysis on the high-dimensional features extracted by the model. The clustering distribution of features in the low-dimensional space is observed to verify the model's feature extraction capability. To optimize classification accuracy, for the visual encoder, the ViT-B / 16 architecture was selected as the optimal backbone network after comparative verification; for the text encoder, "detailed working condition description" containing rich semantic information was used as the input prompt word; in terms of training strategy, data augmentation techniques such as image rotation, cropping and flipping were introduced, and a cosine annealing strategy was used to dynamically adjust the learning rate to balance the convergence speed and final generalization performance of the model. The zero-sample testing process for unknown wells involves selecting entirely new gas well data that has not been used in model training as the zero-sample test set. The test set data undergoes the same ratio normalization, sliding window segmentation, and grayscale matrix preprocessing as the training phase. The processed test images are then input into the trained CLIP model, which extracts image feature vectors and predefined text feature vectors for six different work conditions. The cosine similarity between the image features and each text feature is calculated. Finally, the work condition category with the highest similarity is selected as the diagnostic result for the test sample.

7. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 1, 2, 3 or 5, characterized in that: The CLIP model is constructed using an improved CLIP framework, which includes a visual encoder and a text encoder. In the visual encoder part, after comparative experiments, the Vision Transformer was selected as the optimal backbone network because it maintains high performance while having better inference efficiency compared to larger models. In the text encoder part, various schemes for representing text labels were explored, including "simple operation label", "this is the running status of <category>", and "detailed operation description". The "detailed operation description" format provides richer semantic information, making the model perform best when recognizing complex situations, and therefore it was selected as the default configuration, as shown in Table 1. 。 8. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 7, characterized in that: In CLIP model training, when training on a small dataset, duplicate data may appear in each batch. To address the problem of model convergence difficulties caused by duplicate data within batches when training on small datasets, a data selector mechanism is introduced before training the model to ensure that the data in each batch is not duplicated, thus supporting effective contrastive learning. To further improve model performance, a cosine annealing learning rate decay strategy is adopted. This strategy can maintain a high learning rate in the early stage of training to accelerate convergence, and gradually reduce the learning rate in the later stage to finely adjust the model parameters, thereby achieving the best balance between convergence speed and final performance.

9. The zero-sample operating condition diagnostic method for plunger gas lift process according to claim 8, characterized in that: The research data comes from real-time production data of the plunger gas lift system, covering key sensor information such as oil pressure, casing pressure, gas delivery pressure and production. The PyTorch deep learning framework in Python is used to build and train the clip model. By comparing and learning the feature relationships between different modes in the clip, the plunger lift conditions of unknown wells can be identified.

Images