Distributed private cloud system for logging data processing and control method thereof

Abstract

The embodiment of the invention provides a distributed private cloud system for logging data processing and a control method thereof, and belongs to the technical field of oil-gas exploration and development. The system comprises a data access unit used for carrying out format conversion and preprocessing on collected data to generate a standardized data set; the distributed storage unit is used for executing distributed storage of the standardized data set; the learning modeling unit is used for generating a trained model structure; the model deployment and computing power scheduling unit is used for deploying the trained model structure to a plurality of GPU computing nodes in a containerization manner and dynamically distributing training and reasoning tasks according to node loads; the remote cooperation unit is used for receiving underground real-time feedback data and transmitting the data back to the data access module; and the result output unit is used for generating a logging report. According to the scheme, the core problems of data splitting, slow model response and scheduling rigidity in a traditional system are solved, and the intelligence and automation level of the whole well logging process is improved.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and specifically to a distributed private cloud system for well logging data processing and a control method for such a system. Background Technology

[0002] Well logging technology is a core tool in the exploration and development of oil, gas, and geothermal resources, and is widely used in key aspects such as formation identification, fluid discrimination, and reservoir evaluation. With the widespread adoption of intelligent well logging equipment and various types of sensors, the scale, dimensions, and complexity of well logging data are growing exponentially, posing unprecedented challenges to traditional well logging data processing and interpretation workflows.

[0003] Existing well logging data processing systems mostly adopt a stand-alone deployment or a small-scale server operation mode within a local area network. These systems have fixed resources and poor scalability, making it difficult to meet the parallel computing needs of large-scale heterogeneous well logging data. Furthermore, inconsistent data standards between different well logging methods (such as conventional electrical resistivity tomography, acoustic imaging, and component analysis) lead to low data fusion efficiency, a high dependence on manual data cleaning in the processing workflow, and a significant "data silo" problem. In addition, traditional systems often deploy well logging data processing, modeling analysis, and engineering control in a decentralized manner. Model training typically relies on offline processes, lacking a unified model framework and interpretability, making it difficult to provide rapid response and dynamic adjustment support for complex well conditions and unconventional reservoirs.

[0004] Especially in emerging energy fields such as geothermal and shale gas, remote logging operations, fracture development layer identification, and reinjection adaptability evaluation place higher demands on real-time performance and model generalization capabilities. However, existing solutions generally lack flexible computing resources, have long model update cycles, and weak adaptability, which seriously restricts the improvement of logging intelligence.

[0005] Therefore, there is an urgent need for an intelligent logging data processing architecture with high concurrency processing capabilities, unified model expression capabilities, and system-level linkage mechanisms, which can break down the barriers between data acquisition, modeling analysis, and operation control, and realize cloud-based intelligent collaborative processing of the entire logging process. Summary of the Invention

[0006] The purpose of this invention is to provide a distributed private cloud system and its control method for well logging data processing, so as to at least solve the problems of resource bottlenecks and slow model response in the process of large-scale well logging data processing, which cannot meet the needs of real-time interpretation and remote control under complex well conditions.

[0007] To achieve the above objectives, the first aspect of the present invention provides a distributed private cloud system for well logging data processing. The system includes: a data access unit for collecting structured and unstructured data generated by downhole logging instruments, and performing format conversion and preprocessing on the collected data to generate a standardized dataset; a distributed storage unit for receiving the standardized dataset and storing it in an object-oriented distributed storage structure using a time-depth indexing method; a learning and modeling unit for reading the standardized dataset from the distributed storage module to construct a well logging learning model, and executing training of the well logging learning model to generate a trained model structure; a model deployment and computing power scheduling unit for containerizing and deploying the trained model structure to multiple GPU computing nodes, and dynamically allocating training and inference tasks according to node load; a remote collaboration unit for using model inference results to control well site operations based on a central control communication link, receiving real-time downhole feedback data and transmitting it back to the data access module; and a result output unit for receiving model inference results, generating a well logging report, and pushing it to the user terminal through a front-end interface.

[0008] Optionally, the data access unit includes: a structured data acquisition module for logging curve data as a first acquisition dataset; an unstructured data parsing module for extracting text information from logging images and / or documents as a second acquisition dataset; and a data preprocessing module for sequentially performing normalization, missing data completion, and format conversion on the first and second acquisition datasets to generate a standardized dataset with well numbers, well segments, and depth indices.

[0009] Optionally, the distributed storage unit includes: an object storage module for storing standardized datasets in a key-value format; a data indexing module for constructing a multi-level index structure according to hash and depth; and a data synchronization module for automatically synchronizing write operations among multiple storage nodes.

[0010] Optionally, the learning modeling unit includes: a feature mapping module for mapping a standardized dataset to a high-dimensional feature space; a parameter tuning module for adjusting the learnable parameters in the model to adapt to different logging data characteristics; and a residual learning module for introducing data simulated based on a physical model and constructing a loss function structure that includes physical priors.

[0011] Optionally, the model scheduling unit includes: a model container packaging module for packaging the trained model into a deployable image file; a GPU node scheduling module for allocating training or inference tasks based on the resource load of each node; wherein, allocating training or inference tasks based on the resource load of each node includes: obtaining the dominant resource utilization rate of each GPU node, wherein the dominant resource is the type of resource with the highest proportion in the current GPU node resources; calculating the dominant resource consumption ratio of each task to be scheduled on each GPU node and constructing a resource competition matrix; allocating tasks to the target GPU node that minimizes the maximum dominant resource ratio; and a dynamic migration module for automatically migrating tasks to idle nodes when GPU resources are insufficient.

[0012] Optionally, the collaborative control unit includes: a central control command generation module for generating remote operation control commands based on model inference results; an edge agent communication module for establishing an encrypted communication link with the well site logging main control equipment; and a data feedback module for receiving real-time feedback on downhole operation status and logging data, and updating and writing the feedback data to the data access unit.

[0013] Optionally, the model inference results include: lithology identification tags for distinguishing reservoir lithology types; reservoir fluid type discrimination results for identifying oil, water, gas, or dry reservoirs; and fracture strike and dip estimation results for fault structure interpretation and reinjection path analysis. The model inference results are transmitted to the result output unit in the form of structured data so that the result output unit can generate corresponding visualization charts.

[0014] A second aspect of the present invention provides a control method for a distributed private cloud system for well logging data processing. The method is applied to the distributed private cloud system for well logging data processing as described in any one of claims 1-7. The method includes: acquiring structured and unstructured data from downhole logging instruments, performing format conversion and preprocessing on the data to generate a standardized dataset; storing the standardized dataset in an object-oriented distributed storage structure using a time-depth index; constructing a well logging learning model based on the standardized dataset and performing model training to generate a trained model structure; deploying the trained model structure to multiple GPU computing nodes and performing dynamic training and inference task scheduling according to the node resource status; generating a well logging report based on the model inference results and well logging feedback data, and pushing it to the user terminal through a front-end interface.

[0015] Optionally, dynamic training and inference task scheduling is performed based on node resource status, including: obtaining the dominant resource utilization rate of each GPU node, wherein the dominant resource is the type of resource with the highest proportion in the current GPU node resources; calculating the dominant resource consumption ratio of each task to be scheduled on each GPU node and constructing a resource competition matrix; and allocating the task to the target GPU node that minimizes the maximum dominant resource ratio.

[0016] On the other hand, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the aforementioned control method for a distributed private cloud system for well logging data processing.

[0017] Through the above technical solution, this invention forms a complete closed loop from well logging data acquisition to model training and deployment, operation control, and result generation by constructing six functional units: data access, distributed storage, learning and modeling, model scheduling, remote collaboration, and result output. This system can achieve unified formatting and efficient storage of multiple types of well logging data, significantly improving data fusion efficiency. Simultaneously, relying on a distributed computing power scheduling mechanism, it supports parallel deployment and dynamic migration of models across multiple nodes, effectively alleviating resource bottlenecks. Furthermore, combined with a remote collaboration mechanism, inference results can directly drive well site control commands, achieving real-time closed-loop optimization of the well logging process. The overall solution solves the core problems of data fragmentation, slow model response, and rigid scheduling in traditional systems, improving the intelligence and automation level of the entire well logging process.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a system architecture diagram of a distributed private cloud system for well logging data processing provided in one embodiment of the present invention;

[0021] Figure 2 This is a flowchart of the steps of a control method for a distributed private cloud system for well logging data processing provided by one embodiment of the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0023] Figure 1 This is a system architecture diagram of a distributed private cloud system for well logging data processing provided in one embodiment of the present invention. Figure 1 As shown, this invention provides a distributed private cloud system for well logging data processing. The system includes: a data access unit for collecting structured and unstructured data generated by downhole logging instruments, converting and preprocessing the collected data to generate a standardized dataset; a distributed storage unit for receiving the standardized dataset and storing it in an object-oriented distributed storage structure using a time-depth indexing method; a learning modeling unit for reading the standardized dataset from the distributed storage module to construct a well logging learning model, and executing the training of the well logging learning model to generate a trained model structure; a model deployment and computing power scheduling unit for containerizing and deploying the trained model structure to multiple GPU computing nodes, and dynamically allocating training and inference tasks according to node load; a remote collaboration unit for using the model inference results to control the well site operation process based on the central control communication link, receiving real-time feedback data from the well and transmitting it back to the data access module; and a result output unit for receiving the model inference results, generating a well logging report, and pushing it to the user terminal through a front-end interface.

[0024] The present invention provides a distributed private cloud system for well logging data processing. The system includes a data access unit, a distributed storage unit, a learning and modeling unit, a model deployment and computing power scheduling unit, a remote collaboration unit, and a result output unit, thus constructing an intelligent architecture covering the entire process of well logging data acquisition, processing, modeling, control, and interpretation.

[0025] The system comprises several components: a data access unit that collects structured and unstructured data generated by downhole logging instruments and standardizes multi-source heterogeneous data through format conversion and preprocessing to provide a unified input for subsequent processing; a distributed storage unit that stores standardized data in an object-oriented structure and organizes the data based on time and depth indexes to improve access efficiency and retrieval flexibility; a learning and modeling unit that builds logging learning models based on stored data and achieves functions such as formation identification and fluid discrimination through model training; a model deployment and computing power scheduling unit that distributes models to multiple GPU nodes in a containerized manner and schedules tasks based on resource status to achieve parallelization and dynamic optimization of inference and training; a remote collaboration unit that uses model results to guide well site operations through communication links, supports logging equipment control and real-time data feedback, and constructs a closed-loop operation control path; and a results output unit that receives inference results, generates structured logging reports, and pushes them to users through a front-end interface.

[0026] This system can effectively solve problems such as data islands, model isolation, and rigid computing resource scheduling existing in traditional logging processes, and achieve unified management of logging data, automated operation of modeling analysis, and real-time linkage control of wellsite operations. Through a distributed architecture and an elastic computing power scheduling mechanism, the computing efficiency and adaptability of the system are improved, significantly enhancing the intelligence, collaboration, and efficiency levels of logging work.

[0027] Preferably, the data access unit includes: a structured data acquisition module for logging curve data as the first acquisition data set; an unstructured data parsing module for extracting text information from logging images and / or documents as the second acquisition data set; and a data preprocessing module for successively performing normalization, missing value completion, and format conversion on the first and second acquisition data sets to generate a standardized data set with well numbers, well sections, and depth indexes.

[0028] In the embodiments of the present invention, the data of a geothermal field generally includes geological exploration data, geothermal resource reserve data, geothermal logging data, engineering data, environmental and safety data, and other relevant professional basic data. Geological exploration data includes data such as underground temperature, surface temperature, rock type, and groundwater hydrogeological conditions. Geothermal resource reserve data can be used to evaluate the exploitable nature and utilization potential of geothermal energy, including parameters such as reservoir permeability, fracture characteristics, and geothermal gradient. Logging data such as geothermal well drilling data, wellbore temperature, and formation pressure. In addition, engineering data for the development and utilization of geothermal energy includes technical parameters of the project, engineering design data, operation monitoring data, etc. Environmental and safety data includes groundwater monitoring data, surface environment monitoring data, safety production monitoring data, etc. These data are of important value for the geological engineering evaluation of geothermal fields. Therefore, the management and processing of these data are particularly important. In addition, other logging historical data, etc. are also included. This forms the basis of logging big data.

[0029] The fusion of logging data includes the integration, preprocessing, etc. of logging data. The integration of logging data is to distributively store logging data from various sources in the storage medium of a cloud server to provide data sharing and processing for users. For the preprocessing of logging data in a general format, that is, to perform denoising, error correction, and completion processing on logging data. The methods applied include discretization, dimensionlessization, missing value completion, and distribution transformation, etc. Structured data is stored in the cloud platform through an ETL tool, which will打通 the data flow channels of the data source, near-source storage layer, governance layer, sharing layer, and analysis layer.

[0030] Furthermore, for unstructured logging data, such as images, text, and imaging data, due to their irregular or incomplete data structures and lack of a unified data type, they cannot be represented using the binary logic of a database. Different parsing tools, such as Transformer, can be used to extract key information. The cloud platform uses object storage technology based on the S3 protocol (Simple Storage Service Protocol) to enable the uploading and resuming of interrupted downloads of unstructured logging data such as logging curves and document reports. The logging cloud platform provides file management functions to support the rapid delivery of logging software, allowing for the storage and retrieval of any amount of unstructured, multi-source data on the platform. The combination of synchronization services, data routing, and object storage technology enables the loading of logging processing data and the nearby storage and access of post-processing data.

[0031] In this embodiment of the invention, the data access unit includes a structured data acquisition module, an unstructured data parsing module, and a data preprocessing module. The structured data acquisition module collects various logging curve data from downhole logging instruments, including parameters such as natural gamma, sonic transit time, density, resistivity, and neutron porosity, forming a first acquisition dataset. The unstructured data parsing module identifies and processes images or text documents such as downhole logging imaging maps and comprehensive logging interpretation reports, extracting useful information through image slicing, text extraction, and language model encoding to construct a second acquisition dataset. The data preprocessing module sequentially performs normalization, missing value completion, outlier removal, and unified format conversion operations on the two datasets to ensure alignment consistency of the logging data in the time, depth, and channel dimensions, and constructs a standardized dataset structure containing well number, well segment, and depth indexes for subsequent use by the distributed storage and modeling analysis modules.

[0032] By setting up this data access unit, unified processing of multi-source heterogeneous logging data is realized, which solves the problems of scattered data structure, cumbersome manual cleaning, and difficulty in quantifying and utilizing unstructured data in traditional logging systems. It provides a standardized and highly consistent input data structure for subsequent model training, significantly improving data fusion efficiency and model training quality, and is a fundamental support link for system intelligence.

[0033] Preferably, the distributed storage unit includes: an object storage module for storing standardized datasets in a key-value format; a data indexing module for constructing a multi-level index structure according to hash and depth; and a data synchronization module for automatically synchronizing write operations among multiple storage nodes.

[0034] In this embodiment of the invention, the distributed storage unit includes an object storage module, a data indexing module, and a data synchronization module. The object storage module receives a standardized dataset generated by the data access unit and stores it in a distributed object storage system using key-value pairs. The keys include dimensions such as well number, well segment number, logging data type, and time-depth interval, supporting unified storage and management of structured and unstructured data. The data indexing module constructs a multi-level index structure based on the aforementioned keys, enabling rapid retrieval and location by well number, depth, time, or logging project, and supporting combined condition queries. The index structure supports inverted indexes and spatial hashing optimization strategies, improving access efficiency under large-scale datasets. The data synchronization module enables automatic data replication and version consistency maintenance among multiple storage nodes, supporting master-slave backup, partitioned replicas, and write arbitration mechanisms to ensure strong consistency and high availability of data in a distributed environment.

[0035] By configuring this distributed storage unit, object-oriented management, on-demand access, and efficient distributed read / write operations of well logging data are achieved. This solves the problems of access conflicts, low retrieval efficiency, and data loss due to node failures caused by centralized data storage in traditional well logging systems. It provides high-concurrency, high-reliability data support for model training and remote task scheduling, significantly improving the system's stability and scalability.

[0036] In one possible implementation, the storage of well logging big data on a cloud platform typically needs to consider storage capacity requirements, data access speed, data security, scalability, and flexibility. The cloud platform's well logging big data storage can usually utilize distributed storage systems or cloud storage services, combining high-speed storage media and large-capacity disk arrays to achieve efficient and reliable storage of well logging big data. Simultaneously, data management and backup schemes must be compatible with the storage system to ensure the security and persistence of the well logging big data. In a cloud platform, Input / Output Per Second (IOPS) refers to the number of operations performed on the disk per second; it measures the speed at which a device reads or writes data to the storage medium. A higher IOPS value indicates faster processing of small file transfers, database queries, and similar tasks. When designing cloud platform infrastructure, factors such as CPU and GPU usage, memory requirements, network bandwidth, and IOPS rate need to be considered to ensure efficient execution of critical business processes. In a GPU cloud platform, throughput refers to the amount of data transferred between two endpoints in a computer system over time. It measures the speed at which information flows into or out of a specific component of a distributed system, typically expressed in bits per second (bps). Throughput takes into account various factors affecting data transmission, including latency, congestion control mechanisms, protocol implementation details, and overall performance. Throughput is calculated by dividing the total capacity by the elapsed time. Higher throughput indicates better performance because it allows the system to process large amounts of data quickly while maintaining low latency. By leveraging cloud platform analytics tools and monitoring technologies, administrators can monitor cloud platform technical bottlenecks such as throughput, packet loss, jitter, and latency, and take necessary technical measures to avoid significant business losses.

[0037] Preferably, the learning modeling unit includes: a feature mapping module for mapping a standardized dataset to a high-dimensional feature space; a parameter adjustment module for adjusting the learnable parameters in the model to adapt to different logging data characteristics; and a residual learning module for introducing data simulated based on a physical model and constructing a loss function structure that includes physical priors.

[0038] In this embodiment of the invention, the learning modeling unit includes a feature mapping module, a parameter adjustment module, and a residual learning module. The feature mapping module receives a standardized dataset from a distributed storage unit and converts it into a high-dimensional feature representation through kernel function mapping, encoder networks, etc., to improve the model's ability to express the nonlinear characteristics of complex formations. The mapping process not only extends conventional well logging curves but also fuses and models embedded vectors from images and text, forming a unified input vector structure. The parameter adjustment module dynamically adjusts learnable parameters in the neural network or ensemble learning model, such as weight initialization strategies, activation function types, and interlayer connection structures, according to the target well section, well logging curve type, and data characteristics, to adapt to the modeling accuracy requirements under different geological backgrounds. The residual learning module introduces prior physical model output results (such as wave velocity simulation, resistivity inversion, etc.), calculates the difference between the learned model output and the physical residual term, and embeds it into the loss function during model training, thereby improving the model's ability to fit actual geological laws and its physical interpretability.

[0039] This learning modeling unit enables high-precision formation identification and fluid type prediction without relying on large-scale manually labeled data. It breaks through the dependence of traditional modeling methods on a single model structure, achieving a fusion of data-driven and physically guided approaches. This effectively improves the model's generalization ability and robustness, providing reliable support for intelligent logging interpretation under complex well conditions.

[0040] In one possible implementation, the geological engineering evaluation capability of a geothermal field is related to the lithology, reservoir type, physical properties, and formation pressure of the geothermal reservoir. For geothermal field evaluation, the application of machine learning algorithms mainly focuses on well logging data quality monitoring, reservoir lithology identification, and geothermal well monitoring. In complex geothermal wells, forward modeling iteration can be used to obtain the true electrical, acoustic, and nuclear characteristics of the formation or to correct well logging data. The principle of the unified well logging learning model is used to comprehensively characterize geothermal reservoir characteristics or predict performance. It can reveal the intrinsic correlation between the acoustic, electrical, and nuclear characteristics of the reservoir and its physical and fluid properties, thereby generating a more responsive, interpretable, and predictive rock physics model.

[0041] The unified logging learning model is suitable for rapid and automatic prediction of new data based on observations of known results. It is implemented through a high-level computer language on a logging big data cloud platform. The cloud platform's data model training process involves evaluating the algorithm's performance using performance metrics (RSME or MAE). Training and test data are also interchanged as much as possible to prevent overfitting. Various machine learning algorithms are tested, specifically by determining the machine learning category based on the geothermal field evaluation problem and then evaluating the preferred algorithm. The specific steps are as follows: First, the logging data is preprocessed, and the training and test data formats are unified. Based on the classification, regression, prediction, or clustering problem, no more than 10 standard algorithms are selected, and their performance is evaluated according to the test metrics. A typical machine learning workflow includes data preparation, model training, model scoring, and model fitting. To improve efficiency, smaller datasets are used to evaluate different machine learning algorithms, while the entire dataset is used for algorithm parameter tuning to obtain higher evaluation metrics. Classification and regression can be used for geothermal field lithology classification, reservoir fluid identification, helium inspection, cementing quality, and wellbore fracture evaluation. The advantages of the unified logging learning algorithm in the geological engineering evaluation of geothermal fields are shown in Table 1.

[0042] Table 1. Advantages of the Unified Logging Learning Model Algorithm in Geological Engineering Evaluation of Geothermal Fields

[0043]

[0044]

[0045] Preferably, the model scheduling unit includes: a model container packaging module for packaging the trained model into a deployable image file; a GPU node scheduling module for allocating training or inference tasks based on the resource load of each node; wherein, allocating training or inference tasks based on the resource load of each node includes: obtaining the dominant resource utilization rate of each GPU node, wherein the dominant resource is the type of resource with the highest proportion in the current GPU node resources; calculating the dominant resource consumption ratio of each task to be scheduled on each GPU node and constructing a resource competition matrix; allocating tasks to the target GPU node that minimizes the maximum dominant resource ratio; and a dynamic migration module for automatically migrating tasks to idle nodes when GPU resources are insufficient.

[0046] In this embodiment of the invention, the model scheduling unit includes a model container encapsulation module, a GPU node scheduling module, and a dynamic migration module. The model container encapsulation module encapsulates the trained model structure generated in the learning modeling unit, producing an image file that conforms to deployment standards. This module supports packaging model files, dependency libraries, configuration parameters, etc., based on container technologies (such as Docker or Kubernetes environments), ensuring that the model can be quickly and stably deployed on various heterogeneous computing nodes, improving the portability and automation of model deployment.

[0047] The GPU node scheduling module is used to perform task scheduling based on the resource status of each GPU node in a distributed computing environment. This module employs a scheduling algorithm based on Dominant Resource Fairness (DRF) strategy, specifically including the following steps: First, obtain the dominant resource utilization rate of each GPU node, where the dominant resource is the one with the highest proportion among GPU computing resources, memory, bandwidth, etc.; Second, calculate the dominant resource consumption ratio on each GPU node for all training or inference tasks to be scheduled, forming a task-node resource occupancy matrix; Finally, prioritize assigning each task to the target GPU node that minimizes its dominant resource proportion on that node, thereby achieving global resource balance and preventing any one type of resource from becoming a system bottleneck.

[0048] The dynamic migration module is used to monitor the running status of nodes in real time and trigger the task migration mechanism when GPU resource usage approaches saturation or nodes malfunction. The task migration process is based on node health assessment indicators. It can migrate running tasks to idle nodes after interruption and snapshot, and automatically resume execution, ensuring that the overall scheduling process is uninterrupted, without loss, and has self-recovery capabilities.

[0049] By configuring this model scheduling unit, the entire process from model training to distributed deployment and execution is automated. Compared to traditional fixed allocation or static configuration methods, this invention adopts a dynamic scheduling strategy driven by multiple resource indicators, which improves the overall utilization of the GPU cluster and significantly reduces resource waste and node waiting time. At the same time, the combination of containerization and migration mechanisms enhances the elasticity and robustness of the scheduling system, making it suitable for real-time concurrent execution scenarios of different types of logging tasks, and particularly suitable for oil and gas and geothermal exploration operation platforms with complex resource scheduling and large data scale.

[0050] In another possible implementation, the well logging cloud platform includes high-bandwidth memory, a tensor core architecture, CUDA, and a deep learning architecture. The high-bandwidth memory of the GPU is specifically designed for the cloud platform environment, making it ideal for large-scale training and inference of neural networks. The foundational layer of the well logging cloud platform provides abundant computing resources, enabling parallel execution across multiple nodes within the cloud platform, thereby accelerating tasks such as well logging big data mining and learning algorithms. The cloud platform features strong scalability, low application costs, low local resource requirements, and high efficiency. Its strong scalability means that when users need to expand the scale and workload of tasks, the GPU cloud platform can add more GPUs to improve efficiency. For acceleration algorithms in well logging big data processing, such as kernel machine learning (deep learning), parallelization of data and algorithms can significantly accelerate computation and improve algorithm execution efficiency. GPUs can not only perform computational processing but also support real-time data visualization and interactive analysis, allowing the big data processing process and results to be presented intuitively. This provides higher efficiency and performance for the geological engineering analysis of geothermal big data.

[0051] The development of hardware computing power has laid the foundation for the establishment of well logging big data cloud platforms. Multiple processor clusters, composed of single GPUs, can stream and process data, enabling thousands of threads to work simultaneously. This is ideal for the parallel execution of machine learning algorithms for well logging big data. In machine learning tasks, including deep learning, the main program runs on the CPU, while the parallel computing program (CUDA) is processed on the GPU. The CPU can allocate complex tasks such as 3D graphics rendering and large-scale calculations to the GPU. With the support of programming frameworks like TensorFlow / Keras or PyTorch, GPUs and neural networks can achieve optimal matching to complete well logging tasks within this framework. The well logging big data cloud platform provides a hardware platform and service framework for distributed storage and processing of well logging big data, data training, verification, and testing of artificial intelligence learning models. The network configuration and architecture of the well logging big data GPU cloud platform, including the physical addresses allocated within the platform, are described. Its internal modules include: a distributed file system, resource management and scheduling module, a distributed computing framework, various data operation and learning modules, a service connection bus layer, and a user application layer.

[0052] Traditional evaluation methods assess geothermal fields using the evaluation approach of the oil and gas industry. The well logging big data cloud platform, based on the fourth paradigm of well logging knowledge discovery, follows a research methodology of "well logging big data cloud platform - unified well logging learning model - large function space - knowledge learning and discovery - application." It achieves technological innovations in "data distribution, computational in-memory, algorithm scaling, program datafication, and parallel learning" on the well logging big data GPU cloud platform. First, the applicability of ML algorithms to geothermal field evaluation is analyzed. Then, the GPU cloud platform is used to perform lithological analysis, reservoir and structural description, and helium detection of geothermal reservoirs, and to evaluate geothermal cementing engineering through remote operations. This demonstrates the feasibility of implementing the GPU cloud platform. Analysis of algorithms, programming model libraries, cost, processing units, task complexity, data volume, multi-source data processing, learning data exchange, data result sharing, and professional software sharing shows that the ML algorithm driven by the GPU well logging big data cloud platform has advantages that traditional well logging methods cannot match. The characteristics and advantages of the GPU cloud platform are shown in Table 2.

[0053] Table 2 Comparison of Well Logging Big Data GPU Cloud Platform and CPU / Machine Learning Platform

[0054]

[0055]

[0056] Preferably, the collaborative control unit includes: a central control command generation module for generating remote operation control commands based on model inference results; an edge agent communication module for establishing an encrypted communication link with the well site logging main control equipment; and a data feedback module for receiving real-time feedback on downhole operation status and logging data, and updating and writing the feedback data to the data access unit.

[0057] In this embodiment of the invention, the collaborative control unit includes a central control command generation module, an edge agent communication module, and a data feedback module, which are mainly used to realize dynamic linkage and closed-loop control between the logging model inference results and the well site operation process.

[0058] The central control command generation module automatically constructs remote operation control commands based on the model inference results output by the learning modeling unit. These inference results typically include information such as lithological identification tags, fracture development characteristics, and fluid type determination. This module maps these results into executable well site control parameters through a pre-set control strategy rule base. For example, after identifying an oil-bearing or rechargeable layer, it automatically issues operation commands such as triggering tests, adjusting pump speed, and adjusting winch depth. This module supports both manual intervention and automated generation modes, which can be flexibly switched according to different operating conditions.

[0059] The edge proxy communication module establishes a communication link between the cloud platform control system and the well site logging main control equipment. This communication channel is encrypted using the TLS protocol to ensure secure and reliable data transmission in public or private network environments. This module deploys a lightweight proxy program at the well site, providing functions such as command parsing, authentication, and status reporting. It can accurately translate platform commands into instrument-level operating commands and is compatible with various logging control system interface standards.

[0060] The data feedback module is used to collect real-time status information and raw logging data during downhole operations and upload them to the central platform for secondary processing and model calibration. The collected data includes the current operating depth, instrument operating status, real-time acquisition curves, and operating condition change information. The module features breakpoint resume and data redundancy caching mechanisms to ensure data integrity and continuity in weak network environments. After being formatted, the feedback data is automatically updated and written to the data access unit and used for the next round of model training or control strategy correction, enabling the system's adaptive learning and dynamic optimization.

[0061] Through the design and implementation of the aforementioned collaborative control unit, the system has for the first time achieved fully automated control capabilities from "model output to operational response, and then to data closed-loop feedback," breaking through the technical bottlenecks of data processing and operational control separation and feedback lag in traditional logging processes. It significantly improves the response speed and precision control level of logging operations, making it particularly suitable for complex operational scenarios such as geothermal and shale gas exploration where real-time performance and control accuracy are crucial, providing key support for the construction of intelligent well sites and automated exploration.

[0062] Preferably, the model inference results include: lithology identification tags for distinguishing reservoir lithology types; reservoir fluid type discrimination results for identifying oil, water, gas, or dry reservoirs; and fracture strike and dip estimation results for fault structure interpretation and reinjection path analysis. The model inference results are transmitted to the result output unit in the form of structured data so that the result output unit can generate corresponding visualization charts.

[0063] In this invention, the model inference results are structured interpretations obtained through inference calculations after the learning modeling unit completes model training and deployment. These include lithology identification tags, reservoir fluid type discrimination results, and fracture strike and dip estimation results. Lithology identification tags are used to automatically classify the lithology of downhole formations based on well logging data. Combined with natural gamma ray, density, and sonic curve features, they identify major reservoirs or caprocks such as sandstone, mudstone, limestone, and tuff, providing fundamental support for geological modeling and development planning. Reservoir fluid type discrimination results are based on multi-parameter joint analysis, outputting classification tags and confidence levels for oil, water, gas, or dry layers, used to determine reservoir productivity characteristics and development potential. Fracture strike and dip estimation results are obtained based on imaging logging and sonic response pattern recognition algorithms, extracting fracture direction, extension length, and geometric parameters to assist in fault structure identification and reservoir flow path analysis, particularly suitable for fractured geothermal reservoirs or tight sandstone gas reservoirs.

[0064] The inference results are encapsulated in structured data format, including fields such as classification labels, parameter values, confidence probabilities, and depth coordinates, and are transmitted to the result output unit through a unified interface. The result output unit can generate corresponding visualization charts based on this structured data, such as lithological profiles, fluid type distribution maps, and fracture geometry diagrams, achieving engineering visualization of the model results. This structured approach to model inference results not only improves the accuracy and readability of the interpretation results but also achieves automatic connection between model output and result display, reducing manual intervention and enhancing interpretation efficiency and data application value—a key element in realizing intelligent well logging.

[0065] The distributed private cloud system for well logging data processing proposed in this invention can be applied to well logging data processing in multiple scenarios, thus proposing various possible application scenarios and implementations, but is not limited to these application scenarios.

[0066] Example 1:

[0067] Geothermal logging encompasses conventional logging methods such as temperature logging, resistivity logging, natural gamma logging, and sonic logging, as well as novel logging techniques like electroacoustic imaging and remote sensing. These methods obtain parameters such as temperature, pressure, resistivity, porosity, and permeability of subsurface rocks, enabling the evaluation of geothermal reservoirs' thermal storage capacity, thermal conductivity, and the properties and distribution of fluids within the reservoir. Geothermal reservoirs typically exhibit complex pore structures and mineral compositions, making it difficult to describe the complex response relationships between lithology, physical properties, and oil-bearing characteristics using traditional petrophysical models. Therefore, the geological engineering evaluation of geothermal fields relies heavily on logging technology. Geothermal logging evaluation allows for the assessment and prediction of these risks and problems, providing technical support and risk control measures for the development and utilization of geothermal energy. Based on the task assigned by the geothermal field company, the text was first preprocessed, including punctuation filtering, word segmentation, and vocabulary building. Next, the text was vectorized using a bag-of-words model. Then, the dataset was partitioned. Finally, a deep neural network was established. Finally, a feedforward neural network is used to output the well logging series. The purpose of the geothermal field geological engineering evaluation is to establish the corresponding well logging series and the applicability of the learning algorithm.

[0068] Example 2:

[0069] The quality of well logging data is crucial for the geological engineering evaluation of geothermal fields, especially the fundamental data. The following describes quality control of well logging data for lithology identification in geothermal fields, analyzing unsupervised learning clustering of well logging curves and applying a SOM neural network for data clustering analysis. In the case study, 888 sets of sample data were analyzed using expert knowledge to select 26 columns from 32 data columns as observations for cluster analysis. Based on the characteristic analysis of the well logging data, the mapping size was set to 20. The resulting SOM neural network clustering evaluation index chart shows that the neural network weights are relatively dispersed, resulting in a low clustering hit rate. Therefore, the well logging data quality is not high, and data cleaning is necessary, such as removing out-of-bounds data and correcting missing data.

[0070] Example 3:

[0071] This embodiment uses the Weihe Basin as an example for analysis. Tectontologically, the Weihe Basin is a Cenozoic rift basin formed by the Himalayan orogeny, characterized by deep depressions and relatively new strata. Deep geothermal fields possess enormous exploration and development potential, but their complex reservoir composition makes evaluation increasingly difficult. Currently, well logging interpretation faces challenges in identifying lithology, assessing and monitoring reservoirs, and predicting production. Reservoir identification using only characteristic conventional lithological logging curves is complicated by the presence of gypsum, limestone, and dolomite components. The lithological and mineral composition of rocks directly affects the thermal conductivity, permeability, and chemical stability of reservoirs. Different mineral combinations influence the thermal conductivity of rocks; some minerals may have higher thermal conductivity, thus affecting the transfer efficiency of geothermal resources.

[0072] The geological structure of the W22 geothermal well is mainly controlled by faults on the north bank of the Wei River. The reinjection section of this well consists of Neogene Lantian-Bahe Formation and Gaoling Group, and Paleogene Ganhe Formation strata, with sandstone as the reservoir lithology. In geothermal fields with complex geology, the well logging interpretation classification problem involves more complex nonlinear functional relationships. Using 2840 sets of geophysical well logging sample data, 27 feature values ​​were selected, and eight machine learning algorithms—fine-grained tree, coarse-grained tree, weighted KNN, linear SVM, medium-grained tree, fine-grained KNN, logistic regression, and wide neural network—were applied to accurately identify lithology (see Table 10). The best algorithm was fine-grained KNN, with an accuracy of 93.8%; followed by wide neural network, with an accuracy of 84.9%. Table 3 shows that the longest training time required by existing machine learning platforms was for the wide neural network algorithm, at 900.23 seconds, while the GPU cloud platform required 17.668 seconds. The shortest training time required by existing machine learning training platforms is 12.737 seconds for the fine-grained tree algorithm, while the GPU cloud platform requires only 0.212 seconds. This clearly demonstrates that, at the same training accuracy, the GPU cloud platform is on average more than 50 times faster than standalone machine learning training software, exhibiting a significant advantage.

[0073] Table 3 Comparison of accuracy and time of unified logging learning model for lithology identification on CPU and GPU platforms

[0074]

[0075] The Zhangjiapo Formation, containing gas-bearing geothermal reservoirs, is composed of limestone. The Neogene (N) strata consist of three groups. The upper part of the Upper Neogene Zhangjiapo Formation (N2z) consists of alternating layers of brownish-red and light brownish-red mudstone, sandy mudstone, and gray, grayish-brown, and brownish-red siltstone and fine-grained argillaceous sandstone of roughly equal thickness. The middle part consists of grayish-white medium-grained sandstone, light brownish-red mudstone, and sandy mudstone; the lower part consists of brownish-red and light brownish-red mudstone and grayish-white fine-grained sandstone. The Zhangjiapo Formation of the Neogene Yongledian Group is a argillaceous reservoir. The Upper Neogene Lantian Bahe Formation is 957.1 m thick. The upper part consists of alternating layers of light brown and brownish-red mudstone and grayish-white fine sandstone, mudstone, and fine-grained sandstone of roughly equal thickness; the middle part consists of brownish-red and light brownish-red mudstone interbedded with grayish-white fine-grained sandstone; and the lower part consists of alternating layers of light brownish-red and brownish-red mudstone and grayish-white fine-grained sandstone and fine-grained argillaceous sandstone of roughly equal thickness. The average well temperature is 83.87℃. The Neogene Lantian-Bahe Formation is a muddy stratum. The Miocene Gaoling Group has a apparent thickness of 443.3 m, with lithology consisting of light brownish-red to dark red mudstone, silty mudstone interbedded with gray to grayish-white and grayish-green fine siltstone, grayish-white medium-coarse sandstone, and conglomerate. The grain size is coarse at the bottom and fine at the top, with a relatively dense structure. This section has few sandstone reservoirs, a short cumulative thickness, high mud content, and poor porosity and permeability. The average well temperature is 100.45℃. The well temperature of the Neogene Gaoling Group strata is 101.18℃. The Paleogene (E) Ganhe Formation (E3gh) is 126.5 m thick. The lithology is mainly composed of grayish-green and grayish-white medium sandstone, gravelly medium sandstone, conglomerate, and alternating layers of brownish-brown and brown mudstone and sandy mudstone. The average well temperature is 104.72℃. The well temperature of the Paleogene Ganhe Formation muddy strata is 105.12℃.

[0076] Example 4:

[0077] Helium in geothermal fields originates from the decay of radioactive U and TH gases. Another source is the release of various gases, including helium, from the mantle during magmatic processes. These gases, through tectonic movements, enter the geothermal fluid system of the Weihe Basin. The Lantian Bahe Formation is a group of Cenozoic sandstone porous geothermal reservoirs. Vertically, the resource content varies significantly across different strata in the Weihe Basin, with the Lantian Bahe Formation and Gaoling Group being the most concentrated. The Zhangjiapo Formation in the Weihe Basin is a deep-lacustrine deposit with a large thickness and rapid depositional velocity, exhibiting high organic matter content and hydrocarbon generation potential. The section at depths of 3608-3612 meters consists of calcareous sandstone, analyzed to be a water- and gas-bearing layer. The section at depths of 3663-3665 meters consists of fine sandstone, 2 meters thick, with gas anomalies and a helium content of 0.151%.

[0078] Example 5:

[0079] The application of a remote logging control module in a cloud platform enables cementing quality evaluation in geothermal fields. Due to the remote location, inconvenient transportation, and dispersed well sites in geothermal fields, on-site personnel often lack sufficient interpretation experience. Logging companies utilize remote logging controlled by a big data cloud platform to complete cementing quality inspection operations, achieving logging data quality control, data acquisition, and rapid data processing and interpretation. This solves the problem of acquiring and transmitting multi-dimensional and multi-type logging data signals. The logging big data cloud platform, through its remote logging module, realizes functions such as logging instrument operation control, data processing, and interpretation. Remote logging acquisition establishes a technical bridge between the central hub and the well site via the cloud platform, enabling the assembly of logging instruments and equipment at the well site, remotely achieving standardized logging operations, winch instrument operation, and quality, safety, and environmental management. Through remote communication, control over on-site operations, monitoring of logging process quality, and data acquisition are achieved at the central hub. Currently, the long update times for geothermal field management and static / dynamic models increase development risks and can potentially lead to significant economic losses. The monitoring effect of geothermal reservoir cementing quality obtained by remote logging is achieved by analyzing casing waves, formation waves, and mud waves. It can be seen that the casing waves are very weak and the formation waves are very strong, which indicates that the first and second cementing surfaces are well cemented.

[0080] Example 6:

[0081] Remote logging modules in a cloud platform are used to control remote-detection reflected shear wave logging to detect fracture zones near the wellbore. The principle of remote-detection reflected shear wave logging is to generate high-energy seismic waves using transducers placed along the wellbore. These seismic waves propagate through the formation until they reach the interface between two different lithologies. Due to the discontinuity of the elastic modulus, the propagation direction of the seismic waves suddenly changes. Then, using an array acoustic remote-detection acoustic logging tool, monopole wave trains and orthogonal dipole wave trains are acquired, and reflected wave information is extracted from the wave trains. The process involves loading, decoding, instrument type identification, and data quality checks of the raw remote-detection acoustic data. For dipole shear wave reflection imaging processing, the process includes data preprocessing, velocity analysis, wavelet preprocessing, four-component multi-angle imaging, image enhancement, structural occurrence analysis, direct wave suppression and filtering, reflected wave extraction, uplink and downlink wave combination, and image repositioning. Locating the geological structure of a geothermal field that generates reflected waves allows for the determination of the presence of fractures, fissures, reservoir boundaries, faults, and other geological structures and paleogeographic sedimentary information within tens of meters of the geothermal well. This enables geophysicists to determine various characteristics of the subsurface environment, including the presence of fault zones, fractures, and folds, all of which can potentially affect geothermal field production. Based on the unidirectional axis of the reflected shear wave, multiple fractures at different dip angles and azimuths can be identified, along with multi-angle imaging of four-component dipole shear waves.

[0082] Example 7:

[0083] The evaluation module in the cloud platform can be used to evaluate geothermal field reinjection. Geothermal field water reinjection refers to the process of utilizing hot water or steam from underground reservoirs and then reinjecting the resulting water, typically low-temperature wastewater or condensate, back into the ground to maintain groundwater pressure balance and protect the environment. Geothermal water reinjection can achieve effects such as energy recovery and reduction of environmental pollution. Before and during geothermal water reinjection, reservoir pressure data is monitored in real time using logging instruments or sensors to evaluate the physical and mechanical anisotropic properties of the reservoir rocks, including permeability. High permeability is necessary for the effective conduction and storage of geothermal water within the formation. Sandstone, conglomerate, and clastic rocks, due to their high porosity and permeability, are widely used in geothermal water reinjection. For geothermal reservoirs, higher porosity means a greater capacity for geothermal water storage. Thermal conductivity is also an important indicator, as it determines the efficiency of geothermal water conduction and storage.

[0084] The specific process is as follows: First, the thermal conductivity of geothermal water within the rock is determined, which can be obtained through experimental measurements or literature. Next, the volume fraction of solid phase and geothermal water in the rock is determined through exploration or core analysis; the volume fraction can be calculated from the density and porosity of the core sample. Finally, the effective thermal conductivity of the reservoir is calculated. Higher thermal conductivity means that the rock can better conduct and store heat. Rock types with high thermal conductivity include igneous rocks such as granite and basalt, as well as sedimentary rocks with high water-bearing mineral content.

[0085] Geothermal water reinjection requires favorable groundwater circulation conditions to ensure sufficient heat transfer to the surrounding rocks and groundwater during the reinjection process. Ideally, the reservoir should consist of highly porosity and permeable rocks, such as sandstone and conglomerate, which offer good permeability and facilitate better heat diffusion and transfer. Several horizontal and high-angle fractures exist below 7260m. Considering the effect of horizontal in-situ stress, their difference coefficient is considered small. The reservoir's overall quality is good, suitable for fracturing to form a sufficient fracture network, creating good fluid pathways and a high-flow-rate, long-path geothermal reservoir fracture system, which is beneficial for geothermal water reinjection and can effectively support the effective development of geothermal energy storage. There are two fractures extending upwards and downwards, with a defined fracture effect.

[0086] Figure 2 This is a flowchart illustrating a control method for a distributed private cloud system used for well logging data processing, provided by one embodiment of the present invention. Figure 2 As shown, this invention provides a control method for a distributed private cloud system for well logging data processing, the method comprising:

[0087] Step S10: Collect structured and unstructured data from downhole logging instruments, and perform format conversion and preprocessing on the data to generate a standardized dataset.

[0088] Step S20: Store the standardized dataset in an object-oriented distributed storage structure using a time-deep indexing method.

[0089] Step S30: Construct a well logging learning model based on the standardized dataset and perform model training to generate a trained model structure.

[0090] Step S40: Deploy the trained model structure to multiple GPU computing nodes and perform dynamic training and inference task scheduling according to the node resource status.

[0091] Step S50: Generate a well logging report based on the model inference results and well logging feedback data, and push it to the user terminal through the front-end interface.

[0092] Preferably, dynamic training and inference task scheduling is performed based on node resource status, including: obtaining the dominant resource utilization rate of each GPU node, wherein the dominant resource is the type of resource with the highest proportion in the current GPU node resources; calculating the dominant resource consumption ratio of each task to be scheduled on each GPU node and constructing a resource competition matrix; and allocating the task to the target GPU node that minimizes the maximum dominant resource ratio.

[0093] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the control method of the distributed private cloud system described above.

[0094] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0095] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0096] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A distributed private cloud system for processing well logging data, characterized in that, The system comprises: a data access unit for collecting structured data and unstructured data generated by downhole logging instruments, and performing format conversion and preprocessing on the collected data to generate a standardized data set; a distributed storage unit for receiving the standardized data set and storing it in an objectified distributed storage structure according to time-depth indexing; a learning modeling unit for reading the standardized data set from the distributed storage module to build a logging learning model and perform training of the logging learning model to generate a trained model structure; a model deployment and computing power scheduling unit for containerizing the trained model structure and deploying it to multiple GPU computing nodes, and dynamically assigning training and inference tasks according to node load; a remote collaboration unit for using model inference results to control wellsite operation processes based on a central control communication link, receiving real-time feedback data from downhole and returning it to the data access module; a result output unit for receiving model inference results, generating a logging report, and pushing it to the user end through a front-end interface.

2. The system of claim 1, wherein, The data access unit comprises: a structured data collection module for logging curve data as a first collection data set; an unstructured data analysis module for extracting text information from logging images and / or documents as a second collection data set; a data preprocessing module for sequentially performing normalization, missing data completion, and format conversion on the first and second collection data sets to generate a standardized data set with well number, well section, and depth indexing.

3. The system of claim 1, wherein, The distributed storage unit comprises: an object storage module for storing the standardized data set in a key-value manner; a data indexing module for constructing a multi-level indexing structure according to well number and depth; a data synchronization module for automatically synchronizing write operations between multiple storage nodes.

4. The system of claim 1, wherein, The learning modeling unit comprises: a feature mapping module for mapping the standardized data set to a high-dimensional feature space; a parameter adjustment module for adjusting learnable parameters in the model to adapt to different logging data characteristics; a residual learning module for introducing data simulated based on a physical model to construct a loss function structure containing physical priors.

5. The system of claim 1, wherein, The model scheduling unit comprises: a model container packaging module for packaging the trained model into a deployable image file; a GPU node scheduling module for assigning training or inference tasks based on resource load of each node; wherein, assigning training or inference tasks based on resource load of each node comprises: obtaining the dominant resource usage rate of each GPU node, the dominant resource being the type of resource with the highest proportion in the current GPU node resources; calculating the dominant resource consumption proportion of each task on each GPU node, and constructing a resource competition matrix; assigning tasks to the target GPU node with the smallest maximum dominant resource proportion; a dynamic migration module for automatically migrating tasks to idle nodes when GPU resources are insufficient.

6. The system of claim 1, wherein, The collaborative control unit comprises: a central control instruction generation module for generating remote job control instructions based on model inference results; an edge agent communication module for establishing an encrypted communication link with a wellsite logging master control device; A data backhaul module is configured to receive real-time feedback of downhole operation status and logging data, and write the feedback data into the data access unit.

7. The system of claim 1, wherein, The model inference result includes: A lithology identification label for distinguishing reservoir lithology types; A reservoir fluid type identification result for identifying oil layers, water layers, gas layers or dry layers; A fracture strike and dip estimation result for fault structure interpretation and recharge path analysis; The model inference result is transmitted to the result output unit in a structured data form, so that the result output unit generates corresponding visual charts.

8. A control method of a distributed private cloud system for processing well logging data, characterized in that, The method is applied to the distributed private cloud system for logging data processing according to any one of claims 1-7, and the method comprises: Collecting structured data and unstructured data from downhole logging instruments, performing format conversion and preprocessing on the data, and generating a standardized data set; Storing the standardized data set in an objectified distributed storage structure according to time-depth indexing; Building a logging learning model based on the standardized data set, and performing model training to generate a trained model structure; Deploying the trained model structure to multiple GPU computing nodes, and performing dynamic training and inference task scheduling according to node resource status; Generating a logging report based on the model inference result and logging feedback data, and pushing the logging report to the user end through the front-end interface.

9. The method of claim 8, wherein, The dynamic training and inference task scheduling according to the node resource status comprises: Obtaining the dominant resource usage rate of each GPU node, wherein the dominant resource is the type of resource with the highest proportion in the current GPU node resources; Calculating the dominant resource consumption proportion of all tasks on each GPU node, and constructing a resource competition matrix; Assigning the task to the target GPU node with the minimum maximum dominant resource proportion.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores instructions, which when executed on a computer, cause the computer to perform the control method of the distributed private cloud system for logging data processing according to any one of claims 8 or 9. The computer readable storage medium stores instructions, which when executed on a computer, cause the computer to perform the control method of the distributed private cloud system for logging data processing according to any one of claims 8 or 9.

Images