Well logging instrument reliability analysis method based on digital twinning and dynamic stress boundary

Abstract

This application belongs to the field of oil well logging tool technology, specifically a well logging tool reliability analysis method based on digital twins and dynamic stress boundaries. The method includes: constructing a digital twin model of the well logging tool; establishing a virtual-physical interactive connection; performing virtual accelerated testing using the digital twin model to acquire virtual test data; performing high-acceleration life testing on a physical prototype based on dynamic stress boundaries generated from historical operating data; simultaneously acquiring physical test data; comparing and verifying the physical test data with the virtual test data; adjusting the simulation process based on the comparison and verification results; regenerating virtual test data; fusing the virtual test data and physical test data; using graph neural networks for correlation analysis; identifying coupled failure modes; and constructing a predictive health management model. This forms a complete closed loop from virtual design verification to physical test calibration and then to on-site data feedback optimization, enabling reliability to continuously increase throughout the product's entire lifecycle.

Description

Technical Field

[0001] This application belongs to the field of oil well logging instrument technology, specifically a well logging instrument reliability analysis method based on digital twin and dynamic stress boundary. Background Technology

[0002] As a core tool for obtaining critical information about underground oil and gas resources, logging tools operate for extended periods in extreme downhole environments characterized by high temperatures, high pressures, and strong vibrations. Their reliability directly impacts the success rate and cost of logging operations. Traditional reliability analysis methods for logging tools primarily rely on Failure Mode and Effects Analysis (FMEA), Reliability Block Diagrams (RBD), and lifetime statistical models based on the Weibull distribution. These methods are largely ex post facto and static assessments, exhibiting significant limitations: First, they depend on historical failure data, making it impossible to proactively predict novel failures caused by multi-physics coupling during the design phase; second, accelerated life testing in laboratories typically uses fixed, standardized stress profiles, failing to accurately simulate the complex, variable, and transient impact-laden composite load environments of actual downhole operations, leading to a disconnect between testing and real-world conditions; finally, they lack a complete lifecycle data loop from design and testing to field deployment, hindering the continuous improvement of reliability based on real-world data.

[0003] While existing technologies have demonstrated the application of digital twins in equipment health management, these applications are primarily concentrated in large rotating machinery or rail transportation, focusing on monitoring and diagnosing operational status. In the specific field of well logging tools, there is currently no innovative method that systematically integrates high-fidelity digital twins, big data-based dynamic stress boundary modeling, AI-driven failure detection, and full lifecycle closed-loop optimization, and applies them to the reliability design and analysis phase. Therefore, developing an innovative method that enables reliability analysis to shift from static assessment to dynamic evolution is crucial for improving the design quality of well logging tools, reducing R&D risks, and ensuring operational safety. Summary of the Invention

[0004] This application provides a logging tool reliability analysis method based on digital twins and dynamic stress boundaries, which solves the problems of traditional methods being unable to predict coupling failures in advance, the test conditions not matching the actual situation, and the lack of data closed-loop drive.

[0005] This application provides a method for reliability analysis of logging tools based on digital twins and dynamic stress boundaries, including: Constructing a digital twin model of a logging tool involves creating a three-dimensional geometric model based on engineering drawings, and embedding material properties, dynamic stress boundaries, behavioral logic, and fault rules into the three-dimensional geometric model to form a digital twin model for multiphysics simulation. Establish a virtual-physical interactive connection to achieve bidirectional data synchronization and closed-loop control between the physical entity of the logging tool and the digital twin model; The digital twin model is used to conduct virtual acceleration testing and obtain virtual test data; Based on the dynamic stress boundary generated by historical operating condition big data, high-acceleration life test is performed on the physical prototype, and physical test data is collected simultaneously. The physical test data is compared and verified with the virtual test data, and the simulation process is adjusted according to the comparison and verification results to regenerate the virtual test data. By integrating virtual test data and physical test data, graph neural networks are used for correlation analysis to identify coupled failure modes and build a predictive health management model.

[0006] Furthermore, the generation of dynamic stress boundaries includes: Collect historical operating data, which includes at least depth-temperature-pressure curves, downhole vibration spectrum, tripping and drilling operation parameters, formation characteristics, and instrument status. Identify pressure stress, temperature stress, vibration and shock stress, and cyclic stress as key stress parameters; Statistical analysis was performed on the historical working conditions big data corresponding to each key stress parameter, the probability distribution function of each key stress parameter was fitted, the independent distribution model of each key stress parameter was obtained, and the Monte Carlo method combined with the Copula function was used to generate a stress vector sample set. Based on the stress vector sample set, a static stress envelope containing typical working condition envelopes, severe working condition envelopes, and extreme working condition envelopes is generated, and the static stress envelope is reconstructed into a time-varying dynamic sequence. Multiple stress combinations are defined as simulation scenarios, and each simulation scenario contains a complete time-varying stress sequence.

[0007] Furthermore, the embedding of fault rules includes: Establish a failure mode and impact analysis rule base for logging tools. In the failure mode and impact analysis rule base, high-frequency failure modes in the field are mapped to executable rules. Each executable rule includes a trigger threshold, failure rate function, impact path and a list of associated failures at the next level, and is injected into the digital twin model in the form of XML fragments. High-frequency failure modes in the field include, but are not limited to: high-temperature aging and leakage of sealing rings, abrasive jamming of push arm hinges, and thermal cycling fatigue of circuit board solder joints. Trigger thresholds include one or more combinations of temperature threshold, vibration amplitude threshold, pressure threshold, and cycle number threshold.

[0008] Furthermore, degradation laws are superimposed on the digital twin model, including a wear model based on the Archard formula, a fatigue model based on the Miner linear accumulation rule, and an aging model based on the Arrhenius curve.

[0009] Furthermore, the wear model, fatigue model, and aging model work collaboratively through a multiphysics coupling mechanism of a digital twin model, specifically including: The wear model, fatigue model and aging model share a unified simulation time axis, the same stress field and temperature field, and influence each other through a real-time parameter exchange mechanism; When the leakage rate of the sealing ring calculated by the aging model exceeds the mud intrusion threshold, the mud intrusion flag is activated, and the particle concentration parameter in the mud is transmitted to the wear model to dynamically adjust the wear coefficient and friction coefficient in the wear model. When the wear model calculates that the hinge wear produces wear particles, the particle size and concentration parameters of the wear particles are fed back to the sealing interface to correct the leakage rate of the sealing ring, forming a positive feedback coupling loop between seal aging and hinge wear. When the leakage rate calculated by the aging model exceeds the threshold, the thermal boundary conditions of the electronic compartment are updated. The fatigue model reads the updated temperature field, recalculates the thermal stress cycle amplitude of the solder joint, and adjusts the cumulative damage. When the cumulative damage of the solder joint output by the fatigue model reaches a preset threshold, causing the temperature sensor signal to be interrupted, the aging model switches to virtual sensor interpolation or historical data prediction mode to maintain continuous calculation of the aging degree. System-level failure determination adopts a comprehensive criterion: the system is determined to be in failure when any model output reaches the damage threshold, or when the outputs of two or more models reach the warning threshold, or when the cumulative damage on the coupled propagation path exceeds the threshold.

[0010] Furthermore, virtual acceleration testing is performed using the digital twin model to obtain virtual test data, including: The ultimate stress scanning test is carried out, which includes temperature limit scanning, vibration limit scanning, pressure limit scanning and composite stress limit exploration, and the operational limit and destructive limit under each single stress and composite stress are recorded respectively. Using the operating limit and the failure limit as the upper limit of the accelerated stress, accelerated degradation simulation is performed. Accelerated stress is applied using a stepped stress profile or a cyclic stress profile, and the acceleration factor of each failure mode is calculated. The accelerated failure time under accelerated conditions is extrapolated to the failure time under normal operating conditions through the acceleration factor. Output virtual test data, which includes at least a failure hotspot list, failure time or cycle number prediction, and failure mode distribution.

[0011] Furthermore, the simulation process is adjusted based on the results of the comparison and verification, including: calibrating the material parameters, boundary conditions, or degradation model parameters of the digital twin model, and adjusting the deviation between the virtual test data and the physical test data.

[0012] Furthermore, graph neural networks are used for association analysis, including: Construct a heterogeneous graph with instrument components, stress loads, and failure symptoms as nodes to form a multi-layered correlation network of structure-stress-failure; Define structure-stress edges, stress-failure edges, structure-structure edges, and failure-failure edges in the heterogeneous graph, and introduce physical prior information into the edge weights; A graph neural network is used to propagate and aggregate features from heterogeneous graphs through message passing, graph attention, and hierarchical aggregation strategies. Multidimensional feature vectors are constructed for different types of nodes and used as input to the graph neural network for training. By utilizing the node embedding vectors and attention weights of the trained graph neural network, cross-component cascade failure paths caused by multi-stress coupling can be automatically identified.

[0013] Furthermore, the automatic identification of cross-component cascade failure paths caused by multi-stress coupling includes: employing an attention path tracing method, starting from the initial failure node, tracing the path along edges with high attention weights, examining all outgoing edges at each current node, selecting the edge with the highest attention weight as the propagation path, moving to the next node, and repeating this process until a system-level failure node is reached or the maximum path length is reached; or employing a gradient saliency graph method, calculating the gradient of the graph-level output with respect to the characteristics of each node, with nodes having large gradient magnitudes indicating a significant contribution to system failure, and connecting these high-gradient nodes to form critical failure paths.

[0014] Furthermore, constructing a predictive health management model includes: using the node embedding vector and failure probability output by the graph neural network as input features and inputting them into a lightweight predictive health management model; the predictive health management model includes a remaining lifespan prediction branch, a health index assessment branch, and a failure warning branch, which respectively output the remaining lifespan prediction value, the health index score, and the warning status.

[0015] Compared with the prior art, the advantages of this application are as follows: This application utilizes digital twin virtual accelerated testing to expose potential failures caused by multi-physics coupling that are difficult to detect using traditional methods during the design phase, significantly reducing the cost of later design modifications. Dynamic stress boundaries generated from real-world operating data enable laboratory HALT (High Accelerated Life Testing) tests to more accurately reproduce extreme downhole composite loads, shortening the testing cycle and increasing test coverage. Artificial intelligence algorithms analyze multi-source data to reveal implicit coupled failure modes and cascading paths under the combined effects of vibration, heat, and shock. This forms a complete closed loop from virtual design verification to physical testing calibration and then to on-site data feedback optimization, ensuring that reliability continuously increases throughout the product's entire lifecycle. Attached Figure Description

[0016] Figure 1 The overall flowchart of the reliability analysis method based on digital twin and dynamic stress boundary provided in the embodiments of this application is shown. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] like Figure 1 As shown, this application provides a logging tool reliability analysis method based on digital twins and dynamic stress boundaries, which specifically includes: S101, Construct a digital twin model of the logging tool, including building a three-dimensional geometric model based on engineering drawings, and embedding material properties, dynamic stress boundaries, behavioral logic and fault rules into the three-dimensional geometric model to form a digital twin model for multi-physics simulation; S102, Establish a virtual-physical interactive connection to achieve bidirectional data synchronization and closed-loop control between the physical entity of the logging tool and the digital twin model; S103, Use the digital twin model to perform virtual acceleration testing and obtain virtual test data; S104, based on the dynamic stress boundary generated by historical working condition big data, performs high-acceleration life test on the physical prototype and collects physical test data simultaneously; S105, compare and verify the physical test data with the virtual test data, and adjust the simulation process according to the comparison and verification results to regenerate the virtual test data; S106 integrates virtual test data and physical test data, uses graph neural networks for correlation analysis, identifies coupled failure modes, and constructs a predictive health management model.

[0019] In step S101, a digital twin model of the logging tool is constructed, including establishing a three-dimensional geometric model based on engineering drawings, and embedding material properties, dynamic stress boundaries, behavioral logic, and fault rules into the three-dimensional geometric model to form a multiphysics simulation model; collecting engineering drawings and technical specifications of the logging tool, creating a three-dimensional geometric model for each component, ensuring that the key mating dimensions are accurate, strictly defining the assembly relationship, and finally assembling the components into a complete downhole logging tool string model. The simulation platform includes, but is not limited to, SolidWorks three-dimensional modeling software, and can also use CAD software with equivalent functions such as CATIA, NX, and Creo.

[0020] For example, SolidWorks 3D modeling software can be used to create 3D solid models for each individual component of the logging tool. During the modeling process, it is necessary to ensure the dimensional accuracy of key mating parts, including but not limited to: the interference fit between the sealing groove and the sealing ring, the clearance fit tolerance of the push arm hinge, and the engagement parameters of the threaded connection pair.

[0021] After completing the 3D modeling of each independent component, the components are sequentially assembled into a complete downhole logging tool string model based on predefined assembly relationships. The core of the assembly process lies in strictly defining the spatial constraints of each component to ensure that the digital twin model can accurately reflect the actual assembly state and motion characteristics of the physical prototype in subsequent kinematic and dynamic simulations. These constraints include, but are not limited to: Coaxial constraints are used to ensure that two or more components with circular cross-sections are aligned along the same central axis. They are the most basic and widely used type of constraint in downhole logging tool strings.

[0022] For example, the positioning of the outer shell and the internal electronic compartment: The pressure-bearing outer shell of the downhole logging tool and the internal electronic compartment must maintain a strict coaxial relationship to ensure that the electronic compartment is centered after assembly, avoiding local stress concentration or uneven heat dissipation caused by eccentricity. In the assembly model, this is achieved by defining the coaxial fit between the inner cylindrical surface of the outer shell and the outer cylindrical surface of the electronic compartment.

[0023] Concentric positioning of the sealing ring and sealing groove: After installation, the sealing ring must remain concentric with the sealing groove to ensure uniform circumferential contact pressure on the sealing surface. During modeling, the central axis of the sealing ring and the central axis of the sealing groove are bound together as a coaxial constraint.

[0024] Connection between logging tool sections: A downhole logging tool string is typically composed of multiple logging tool sections connected end-to-end via threads or bayonet joints. The housing axes of each logging tool section must be strictly coaxial to form a straight logging tool string. During assembly, the center axis of the tail end face of the previous logging tool section is defined to coincide with the center axis of the head end face of the next logging tool section.

[0025] Bearing and shaft fit: For rotating parts (such as the shaft of a rotary push arm), the coaxial constraint between the inner ring of the bearing and the shaft needs to be defined to ensure the smoothness of the rotational motion.

[0026] Fitting constraint: Used to define the contact relationship between the surfaces of two components, specifying that they are positioned in a way that the surfaces are in contact with each other after assembly. It is usually accompanied by the definition of normal contact force and friction characteristics.

[0027] For example, the contact relationship between the sealing ring and the sealing groove: After being compressed, the outer surface of the sealing ring fits tightly against the inner wall of the sealing groove. In the assembly model, the fit between the surface of the sealing ring and the bottom and side walls of the sealing groove needs to be defined, and the contact type can be further specified as "friction contact," setting the friction coefficient to simulate the contact behavior between rubber and metal.

[0028] End face sealing contact: The metal end face sealing structure between each section of the logging tool needs to define the "fit" relationship between the two end faces, and the initial contact pressure or interference (such as 0.05mm~0.10mm interference fit) can be specified in the advanced settings to simulate the sealing effect under preload.

[0029] End face contact in threaded connections: After tightening, the lower end face of the bolt head or nut of a threaded fastener forms a close contact with the surface of the connected parts. This contact surface needs to be defined during modeling, and a preload load can be applied in subsequent simulations.

[0030] Contact stop between the push arm and the base: When the push arm is opened to its maximum angle, its limiting surface forms a tight contact with the corresponding stop surface of the base to limit excessive opening. This contact constraint needs to be defined and specified as rigid contact to simulate the limiting effect of a hard stop.

[0031] Distance constraints are used to define the minimum allowable distance or maximum allowable travel range between two components, and are used to limit the movement boundaries of moving parts or ensure assembly clearance.

[0032] For example, the stroke range of moving parts: During the opening and closing process of the push arm, its stroke needs to be limited by the geometry. In the assembly model, the minimum distance (e.g., a gap of 2mm in the closed state) and the maximum distance (e.g., an extension of 150mm in the open state) between the end effector of the push arm and the housing of the logging tool are defined to constrain the movement range of the push arm and avoid non-physical movements that exceed the stroke in the virtual simulation.

[0033] Rotation angle limitations for rotating components: For hinge or pivot components, upper and lower limits for their rotation angle must be defined. For example, define the rotation angle range of the push arm relative to the base as 0°~60°, where 0° is the retracted state and 60° is the fully open state.

[0034] Safety clearance between moving parts: Inside the logging tool, a minimum safety clearance must be maintained between moving parts (such as sliders and connecting rods) and stationary parts (such as housings and circuit board supports) to prevent motion interference. Define the minimum distance constraint between them during assembly (e.g., not less than 0.5 mm) and perform collision detection during motion simulation.

[0035] Spring compression stroke constraint: For preloaded springs or return springs, a distance constraint between their free length and maximum compression length needs to be defined to simulate the working stroke range of the spring.

[0036] Sensor mounting angle positioning: Some directional sensors (such as acoustic transducers and gamma detectors) need to be mounted on the logging tool housing at a specific angle. During assembly, a fixed angular constraint (such as 45° or 90°) between the sensor mounting surface and the logging tool reference surface must be defined to ensure the accuracy of the measurement direction.

[0037] Phase angle of eccentric mechanism: For logging tools that use eccentric mechanisms to achieve the pushing function, it is necessary to define the phase angle constraint between the eccentric block and the logging tool axis, and drive the angle change according to the time function in the simulation to simulate the pushing force generated by the eccentric rotation.

[0038] Thread engagement angle alignment: After tightening, threaded connections usually require the relative angles of the two components to meet specific requirements (such as wrench groove alignment and locating pin hole alignment).

[0039] The swing angle of the hinge structure: For hinged flip structures or swing arms, it is necessary to define the swing angle range around the hinge axis, and a function of the angle changing with time (such as sinusoidal swing) can be applied in motion drive to simulate fatigue accumulation under dynamic load.

[0040] Exporting the complete assembly model while preserving the assembly tree, solid geometry, and basic coordinate information provides ideal input for subsequent processes. To improve simulation efficiency while maintaining details of key components, the model is optimized through the following steps: For key objects in reliability analysis, such as stress concentration areas, moving friction pairs, and sealing structures, their geometric details are preserved or even enhanced; for exterior shells and large support structures, appropriate polygon simplification is performed, removing minor fillets and chamfers that do not affect the mechanical analysis; for complex internal components such as circuit boards and wiring harnesses, simplified solid representations are used, focusing on their placement and attributes rather than geometric details.

[0041] Integrating information beyond geometry, a comprehensive model capable of simulating real behavior and physical response is constructed. The specific steps are as follows: Define density, elastic modulus, Poisson's ratio, yield strength, coefficient of thermal expansion, and fatigue characteristic curves for each key component; collect a large amount of historical data from oilfield databases and well logging field acquisition systems, and use the Monte Carlo method to randomly combine different stress parameters to define the confining pressure and internal pressure, temperature field, vibration, and shock of the logging tool downhole, generating diverse dynamic stress boundaries; model behavioral logic to simulate the complete working cycle of the logging tool from power-on, self-test, push-in, data acquisition, retraction, to shutdown.

[0042] In this embodiment, to construct a digital twin model that can realistically reflect the complex downhole service environment, it is necessary to define the dynamic stress boundary that the logging tool experiences during operation. This step specifically includes four steps: historical operating condition big data acquisition, stress parameter statistical modeling, Monte Carlo random combination, and dynamic stress boundary generation. These will be described in detail below.

[0043] Historical operating data collection: First, a large amount of historical operational data was collected from the database and the well logging field acquisition system. This historical operational data includes at least the following five categories of information: depth-temperature-pressure curves, specifically including hydrostatic pressure, formation temperature, and geothermal gradient at different well depths; downhole vibration spectrum, specifically including vibration acceleration power spectral density, dominant frequency distribution, and vibration amplitude range; tripping and pulling operation parameters, specifically including tripping and pulling speed, dwell time, drill string rotation speed, and circulation rate; formation characteristics, specifically including rock hardness, formation dip angle, fault distribution, and formation pressure coefficient; and instrument status, specifically including number of pushes, cumulative working time, and temperature records of each component.

[0044] The time span of the collected data should cover at least one complete well workover cycle, such as twelve to twenty-four months, to ensure that it includes a complete stress spectrum under different seasons, different blocks, and different operating conditions.

[0045] Stress Parameter Identification and Classification: Preprocessing and feature extraction are performed on the collected historical operating data to identify key stress parameters affecting the reliability of the logging tool. These parameters are categorized according to their nature, including but not limited to: pressure stresses, which include four parameters: confining pressure, internal pressure, differential pressure, and pressure change rate. Specifically, confining pressure refers to the external hydrostatic pressure borne by the logging tool's pressure-bearing shell, typically ranging from 0 to 140 MPa; internal pressure refers to the pressure or gas pressure within the logging tool's internal cavities, typically ranging from 0 to 60 MPa; differential pressure refers to the pressure difference between the inside and outside, determining the leakage risk of the sealing structure, typically ranging from -20 to +140 MPa; and pressure change rate refers to the transient rate of pressure change during tripping in and out of the well, typically ranging from 0 to 10 MPa per second.

[0046] Temperature-related stresses include four parameters: ambient temperature, temperature gradient, rate of temperature change, and self-heating temperature. Ambient temperature refers to the temperature of the formation where the logging tool is located, typically ranging from 20 to 200 degrees Celsius. Temperature gradient refers to the rate of temperature change along the length of the tool, typically ranging from 0 to 50 degrees Celsius per meter. Rate of temperature change refers to the transient rate of temperature change during tripping in and out of the hole, typically ranging from 0 to 5 degrees Celsius per second. Self-heating temperature refers to the additional temperature rise generated by the operation of the tool's internal electronic components, typically ranging from 0 to 30 degrees Celsius.

[0047] Vibration and shock stresses include four parameters: vibration acceleration, vibration spectrum, peak impact acceleration, and impact duration. Vibration acceleration refers to the root mean square amplitude of continuous downhole vibration; the vibration spectrum refers to the energy distribution of different frequency components; peak impact acceleration refers to the maximum acceleration of a transient impact; and impact duration refers to the duration of the impact pulse.

[0048] Cyclic stress includes three parameters: push-in cycle count, pressure cycle count, and temperature cycle count. Push-in cycle count refers to the cumulative number of times the push arm opens and closes; pressure cycle count refers to the number of pressure cycles during the logging tool's deployment and tripping; and temperature cycle count refers to the number of high and low temperature cycles the instrument experiences.

[0049] Statistical analysis was performed on historical data of the aforementioned key stress parameters to fit their probability distribution functions. Different distribution types were adopted based on the different characteristics of the data.

[0050] For example, for confining pressure parameters, a normal or log-normal distribution is used for fitting. For temperature parameters, a normal distribution is used. For vibration amplitude parameters, a Rayleigh or Weibull distribution is used. For peak impact parameters, an extreme value distribution is used, specifically the Gumbel distribution. For the number of impacts, a Poisson distribution is used. For the rate of pressure change parameters, an exponential distribution is used.

[0051] After completing the independent distribution modeling of each key stress parameter, a Monte Carlo method combined with the Copula function is used to generate a stress vector sample set through random combination. The specific steps are as follows.

[0052] Random sampling. For each stress parameter, independent random sampling is performed according to its probability distribution function. Suppose there are N stress parameters in total, then the i-th Monte Carlo simulation generates an N-dimensional stress vector, where each element represents the i-th sampled value of the corresponding stress parameter.

[0053] Correlation handling. Since some stress parameters exhibit physical correlations—for example, an increase in depth simultaneously leads to increases in both confining pressure and temperature—copula functions are needed to handle the dependencies between parameters. Commonly used copula types include: Gaussian Copula, suitable for linear correlations; t-Copula, suitable for scenarios with strong tail correlations; and Clayton Copula, suitable for lower tail correlations, i.e., scenarios where extreme pressures and extreme temperatures occur simultaneously. Copula functions transform independent sampling into joint sampling that reflects actual correlations.

[0054] The sampling is repeated M times, and the value of M must meet the statistical convergence requirement. In one example, for basic analysis, M is between 10,000 and 50,000; for high-precision requirements, M is between 100,000 and 500,000; for extreme condition coverage, M is over one million, which can be accelerated by Latin hypercube sampling.

[0055] Based on the M stress vector sample sets generated by Monte Carlo simulation, dynamic stress boundaries that can be used for digital twin simulation are further generated. Specifically, this includes: Generate static stress envelopes. Perform statistical analysis on all stress vector samples to generate the following three static stress envelope curves: Typical working condition envelope: calculated as the mean of each stress parameter plus or minus one standard deviation. Severe working condition envelope: calculated as the mean of each stress parameter plus or minus two standard deviations. Extreme working condition envelope: calculated as the mean of each stress parameter plus or minus three standard deviations.

[0056] Time series reconstruction. The static stress envelope is converted into a time-varying stress sequence for transient simulation. Three reconstruction methods are used: Random walk method to simulate continuous random fluctuations in stress over time; Markov chain method to simulate discrete jumps in stress state, such as during drilling operations; and measured data stitching method to stitch together real-world segments from historical data.

[0057] Multiple stress combinations are defined as simulation scenarios, each containing a complete time-varying stress sequence. These simulation scenarios serve as input conditions for virtual accelerated testing of the digital twin model. A complete simulation scenario requires defining the complete time-varying history of various stresses such as temperature, vibration, and pressure, simulating the actual stress conditions experienced by the logging tool throughout its entire operational cycle from well access and measurement to well tripping. Each simulation scenario corresponds to a specific operational condition, such as conventional vertical well logging, deep well high-temperature logging, complex horizontal well conditions, extreme shock conditions, or high-frequency cyclic conditions, ensuring that virtual and physical test data cover the complete stress spectrum from typical operations to extreme conditions.

[0058] The embedding of fault rules includes: establishing an FMEA (Failure Mode and Effects Analysis) rule base for logging tools, mapping high-frequency failure modes in the field to executable rules, each executable rule containing a trigger threshold, failure rate function, impact path and a list of associated faults at the next level, and injecting it into the digital twin model in the form of XML fragments; high-frequency failure modes in the field include but are not limited to: high-temperature aging leakage of sealing rings, abrasive jamming of push arm hinges and thermal cycling fatigue of circuit board solder joints, and trigger thresholds including one or more combinations of temperature threshold, vibration amplitude threshold, pressure threshold and cycle number threshold.

[0059] The degradation laws are superimposed in the digital twin model. These degradation laws include: a wear model based on the Archard formula, which updates the clearance and friction coefficient in real time according to the contact pressure distribution and sliding distance of the push-arm kinematic pair; a fatigue model that deducts the remaining life of components subjected to alternating loads, such as electronic connectors and wire rope suspension points, according to the Miner linear cumulative damage law; and an aging model that fits the permanent compression deformation and elastic modulus decay of rubber seals under high temperature environment into an Arrhenius curve, dynamically correcting the sealing force and leakage rate.

[0060] Among them, the Archard formula is a classic model in the field of tribology for predicting the amount of material wear. The core formula is: , Wear volume : Dimensionless wear coefficient Contact pressure, Sliding distance, Brinell hardness for softer materials; In this embodiment, the Archard formula is applied as follows: First, the contact pressure distribution cloud map of the push arm hinge during the movement is obtained through finite element analysis, and the maximum contact pressure value and contact area are extracted.

[0061] Secondly, based on the opening and closing motion of the push arm, calculate the sliding distance for each cycle. The sliding distance is equal to the arc length of the contact point relative to the trajectory.

[0062] Then, an initial value for the wear coefficient is set, which is calibrated based on material friction and wear test data. For the wear coefficient of steel-to-steel friction pairs in a sandy slurry environment, the typical range is from 10⁻⁵ to 10⁻³.

[0063] Finally, after each simulation time step or each push-pull cycle, the Arcard formula is called to calculate the cumulative wear volume and convert it into the hinge clearance increment and friction coefficient increment. When the clearance increment exceeds the maximum allowable value (e.g., 0.2 mm) or the friction coefficient exceeds the threshold (e.g., 0.6), wear failure is determined to have occurred, triggering the corresponding fault rule.

[0064] The Miner Linear Cumulative Damage Rule is a classic theory for predicting the fatigue life of materials. It is used to quantify the cumulative damage of materials under alternating loads. When a material is subjected to cyclic loads, each cycle will cause a certain degree of damage. The Miner rule assumes that the damage is linearly cumulative, and the material fails when the total damage reaches 1. In this embodiment, the Miner's rule is applied as follows: The stress-life curves (SN curves) of each component are obtained from material fatigue test data and then parameterized and embedded into a digital twin model.

[0065] During the simulation, the alternating stress time history of each key component is monitored in real time, and the stress amplitude and mean stress of the stress cycle are extracted using the rainflow counting method.

[0066] For each extracted stress cycle, the fatigue life at that stress level is obtained by interpolation from the SN curve based on its stress amplitude, and the damage contribution is accumulated.

[0067] When the cumulative damage degree D reaches the preset warning threshold, a fatigue life warning is issued; when the cumulative damage degree D reaches 1, fatigue failure is determined to have occurred, and the corresponding fault rules are triggered.

[0068] The Arrhenius curve is a mathematical model describing the relationship between the aging rate of a material and temperature, based on the Arrhenius equation: , For the reaction rate, For frequency factors, For activation energy, The gas constant is This refers to absolute temperature.

[0069] In this embodiment, the Arrhenius curve is applied as follows: The activation energy and frequency factor of the material are obtained through accelerated aging tests.

[0070] The temperature time history at the location of the sealing ring is monitored in real time in a digital twin model. Based on the Arrhenius equation, the aging rate k at different temperatures is calculated, and the cumulative aging degree is obtained by integrating over time.

[0071] Establish a mapping relationship between aging degree and material performance parameters. Specifically, fit the compressive set rate, elastic modulus decay rate, and elongation at break decrease rate as functions of aging degree. For example, the compressive set rate increases exponentially with aging time.

[0072] The sealing force (reduction in compression rebound force) and leakage rate (increased cross-sectional area of ​​leakage channels) of the sealing ring are dynamically adjusted based on the current aging level. When the leakage rate exceeds the allowable threshold (e.g., one milliliter per minute) or the sealing force is lower than the minimum sealing force threshold, it is determined that the seal has failed due to aging, triggering the corresponding fault rules.

[0073] In this embodiment, the three degradation models do not operate independently, but rather work together through the multiphysics coupling mechanism of the digital twin model.

[0074] For example, under high temperature and high pressure combined environments, the aging model continuously calculates the material property degradation of the sealing ring, while the wear model calculates the increase in hinge clearance; the two evolve independently. When the aging of the sealing ring leads to an increase in leakage rate, slurry intrudes into the hinge region, altering lubrication conditions and increasing the wear coefficient K in the wear model, thereby accelerating hinge wear. Conversely, wear debris generated by hinge wear may enter the sealing interface with the slurry circulation, further accelerating seal failure.

[0075] In one embodiment, the wear model, fatigue model, and aging model work collaboratively through a multiphysics coupling mechanism of a digital twin model, specifically including: The wear model, fatigue model and aging model share a unified simulation time axis, the same stress field and temperature field, and influence each other through a real-time parameter exchange mechanism; When the leakage rate of the sealing ring calculated by the aging model exceeds the mud intrusion threshold, the mud intrusion flag is activated, and the particle concentration parameter in the mud is transmitted to the wear model to dynamically adjust the wear coefficient and friction coefficient in the wear model, thereby accelerating the hinge wear process. When the wear model calculates that the hinge wear produces wear particles, the particle size and concentration parameters of the wear particles are fed back to the sealing interface to correct the leakage rate of the sealing ring, forming a positive feedback coupling loop between seal aging and hinge wear. When the leakage rate calculated by the aging model exceeds the threshold, the thermal boundary conditions of the electronic compartment are updated. The fatigue model reads the updated temperature field, recalculates the thermal stress cycle amplitude of the weld joint, and adjusts the cumulative damage. When the cumulative damage degree of the weld joint output by the fatigue model reaches a preset threshold, causing the temperature sensor signal to be interrupted, the aging model switches to virtual sensor interpolation or historical data prediction mode to maintain continuous calculation of the aging degree. The virtual sensor interpolation mode refers to the method of using the measurement data of other related position sensors in the same system to calculate the theoretical measurement value of the failed sensor position through interpolation algorithm when the physical sensor fails or the signal is interrupted, thereby replacing the real sensor data. The historical data prediction mode refers to the method of using the data pattern of the physical sensor in its historical normal operation, combined with the current operating condition parameters, to predict the theoretical measurement value at the current moment when the physical sensor fails and effective interpolation cannot be performed through adjacent sensors.

[0076] System-level failure determination adopts a comprehensive criterion: the system is determined to be in failure when any model output reaches the damage threshold, or when the outputs of two or more models reach the warning threshold, or when the cumulative damage on the coupled propagation path exceeds the threshold.

[0077] This coupling effect is captured through the real-time data exchange mechanism of the digital twin model, enabling virtual accelerated testing to truly reflect the complex degradation behavior under the synergistic effect of multiple failure modes.

[0078] In step S102, a virtual-physical interaction connection is established to achieve bidirectional data synchronization and closed-loop control between the physical entity of the logging tool and the digital twin model. In one example, the virtual-physical interaction connection is based on OPC UA (OPC Unified Architecture, a service-oriented industrial communication protocol). The logging tool uploads the sensed data information to the OPC UA server through an industrial switch using industrial Ethernet technology. The OPC UA server mainly provides services for two functional layers, driving the virtual layer of the digital twin model and the application layer of the twin data. The virtual layer supports the direct connection between the embedded OPC UA client and the real-time data of the field production equipment connected to the OPC UA server, thereby realizing the mapping function of the twin model to the production behavior of the field equipment. In one example, the OPC UA server is built using KEPServerEX6 software to construct the OPC UA communication system for the manufacturing production line. This software is deployed on the central control computer of the logging tool. Each device installs the corresponding hardware driver according to the communication interface protocol specified by the manufacturer and is assigned a unique IP address, thereby enabling real-time acquisition of multi-source heterogeneous data from field devices, control logging tools, and sensors. Through the OPC UA communication protocol, efficient data upload and download are achieved between field devices and the OPC UA server. Leveraging the connection mechanism between KEPServerEX6 software and the MySQL database, as well as the real-time data transmission and storage methods, the integrity and consistency of data during virtual-physical interaction are effectively guaranteed. The digital twin model establishes a high-speed bidirectional connection with the KEPServerEX6 software on a third-party server via an OPC UA communication architecture. It pushes real-time operational data and status information from the downhole logging tool to the KEPServerEX6 software and receives feedback and control commands, achieving a closed-loop interaction between the virtual and physical systems. While exchanging data, the KEPServerEX6 software pre-configures the system data source name and enters the TCP / IP address, server name, and access password for the MySQL database. After verifying the connection, it creates a new data logger internally, selecting the data source name as the target database. It then re-enters the MySQL database authentication information to ensure accurate data entry and binds each data item to be collected. These data items represent various real-time operational parameters and status information of the logging tool. Through this step, data mapping and storage from the KEPServerEX6 software to the MySQL database are achieved.

[0079] In step S103, the digital twin model is used to perform virtual acceleration testing to obtain virtual test data. By inputting stresses that far exceed the limits of physical equipment, the design boundaries are quickly explored and preliminary weak points are identified. This includes: performing limit stress scanning tests, which include temperature limit scanning, vibration limit scanning, pressure limit scanning, and composite stress limit exploration. The operational limits and destructive limits under each single stress and composite stress are recorded respectively. Using the operational limit and failure limit as the upper limit of the accelerating stress, accelerated degradation simulations are performed. Accelerated stress is applied using a stepped stress profile or a cyclic stress profile, and the acceleration factor for each failure mode is calculated. The simulation results under accelerated conditions are extrapolated to the predicted values ​​under normal operating conditions using the acceleration factor. Here, the simulation results under accelerated conditions represent the failure time observed under accelerated conditions, and the acceleration factor is the ratio of the time required to reach the preset degradation level under normal operating conditions to the time required to reach the same degradation level under accelerated conditions. The accelerated failure time under accelerated conditions is then extrapolated to the failure time under normal operating conditions using the acceleration factor.

[0080] The system outputs virtual test data, which includes at least a list of failure hotspots, predicted failure times or cycles, and a failure mode distribution. The failure hotspot list is sorted from lowest to highest stress level at the time of failure; a lower stress level indicates a weaker component. The predicted failure times or cycles provide the estimated failure time or number of failure cycles for each critical component under normal operating conditions. The failure mode distribution statistically analyzes the frequency of various failure modes and identifies the dominant failure mode.

[0081] In this embodiment, after completing the virtual accelerated testing and identifying initial weak points, a high-accelerated life test (HALT) is performed on the physical prototype based on dynamic stress boundaries. This refers to applying composite stresses far exceeding normal operating conditions to a physical prototype of an actual well logging tool in a laboratory environment to quickly induce potential defects and verify the reliability of the virtual test results. The specific process includes: A physical prototype of the logging tool, manufactured according to the final design drawings, was selected as the test object. This physical prototype should be completely identical to the design version corresponding to the digital twin model, including the same materials, manufacturing processes, and assembly tolerances. Before testing, a comprehensive functional check and performance calibration of the physical prototype was performed, and initial state data was recorded as a benchmark for subsequent degradation assessment. To comprehensively collect the response data of the physical prototype during testing, a multi-type sensor network was deployed on the physical prototype. The sensor deployment followed these principles: key components were strategically placed; redundant measurements ensured reliability; and sensor size was kept as small as possible to minimize the impact on the original state of the physical prototype.

[0082] The specific sensor types and their arrangement are as follows: A triaxial high-frequency accelerometer is installed in the middle of the physical prototype to simultaneously acquire time-domain acceleration signals and spectral data in the X, Y, and Z orthogonal directions. This triaxial high-frequency accelerometer is used to monitor the dynamic response of the physical prototype under vibration, including resonant frequency drift, changes in vibration transmission characteristics, and the capture of abnormal impact events. The sampling frequency is set to 5,000 to 20,000 Hz to meet the requirements for acquiring high-frequency vibration components.

[0083] High-g impact sensors are deployed at key locations on the physical prototype body and the outer shell of the electronics compartment, including near the push arm hinges, at the connection points of the pressure-bearing outer shell end caps, and at the electronics compartment mounting supports. These high-g impact sensors are used to record the peak acceleration and duration of impact events.

[0084] Multi-channel temperature sensors are deployed at multiple points along the outer wall of the physical prototype, the internal electronics compartment, and near the detectors. The outer wall temperature sensor measures the temperature of the interface between the instrument and the well fluid; the internal electronics compartment sensor measures the ambient temperature of the electronic components; and the sensor near the detector measures the operating temperature of the core detection element. Through the coordinated operation of these multi-channel temperature sensors, the three-dimensional temperature field distribution of the outer wall and interior of the physical prototype is acquired in real time.

[0085] High-precision pressure sensors are installed on both the inner and outer sides of the pressure-bearing housing of the physical prototype. The external pressure sensor measures the hydrostatic column pressure applied by the test chamber, simulating the downhole confining pressure environment; the internal pressure sensor monitors pressure changes in the internal cavity of the prototype to determine whether the sealing structure has failed. The pressure difference between the inside and outside is a key indicator for evaluating sealing performance.

[0086] Fiber Bragg grating (FBG) sensors are deployed along the length of the physical prototype at various connection points, including threaded connections between instrument sections, fixed connections between the electronic compartment and the outer shell, and hinged connections between the push arm and the base. FBG sensors are used to monitor bending strain and local deformation, offering advantages such as resistance to electromagnetic interference, high temperature resistance, and the ability to be connected in series at multiple points. They are particularly suitable for distributed strain measurement in long, strip-shaped structures like logging tools.

[0087] High-accelerated life testing utilizes a comprehensive environmental test chamber, configured to handle rapid temperature changes, multi-axis random vibration, and pressure control. The chamber's control system should support externally input dynamic stress profiles and be able to adjust the load in real time based on sensor feedback. The generated dynamic stress boundary is then tested in a high-accelerated life comprehensive environment, dynamically adjusting various stress parameters to simulate the transient composite stress environment in the wellbore.

[0088] Historical operating data is used to generate dynamic stress boundaries, and various stress parameters are dynamically adjusted on a comprehensive environmental test chamber to simulate the transient composite stress environment downhole.

[0089] The physical test data is compared and verified with the virtual test data, and the simulation process is adjusted according to the comparison and verification results to regenerate the virtual test data, including: time domain comparison: the sensor time history curves collected by the physical test are superimposed and displayed with the corresponding curves output by the virtual test, and the root mean square error, peak error and phase error are calculated. Frequency domain comparison: Perform spectral analysis on the vibration response data from physical and virtual tests, and compare the power spectral density curves and dominant frequency distribution; Failure mode comparison: Compare the location, mode, and stress conditions of failure when the physical prototype fails with the weak points and failure conditions predicted by the virtual test; Model calibration: When significant deviations are found in the virtual-to-physical comparison, the material parameters, boundary conditions, or degenerate model parameters of the digital twin model are calibrated using the backpropagation algorithm or Bayesian parameter identification method to adjust the deviation between the virtual test data and the physical test data.

[0090] The process of integrating virtual and physical test data includes: preprocessing both virtual and physical test data, including but not limited to time alignment and resampling, data cleaning, normalization and standardization, and feature extraction and dimensionality reduction. Due to time alignment and resampling, the time axes of the virtual and physical test data may differ (virtual test data typically uses an accelerated time axis, while physical test data uses a true time axis), necessitating time alignment and resampling. The time alignment method uses the time axis of the physical test data as a reference, mapping the virtual time axis to the true time axis based on the acceleration factor of the virtual test data. The resampling method unifies both types of data to the same sampling frequency. Linear interpolation or cubic spline interpolation methods are used to upsample low-sampling-rate data and downsample high-sampling-rate data.

[0091] In one embodiment, a virtual test data-driven and physical test data-calibrated strategy is adopted to fuse preprocessed virtual test data and physical test data into a fused dataset; The strategy of virtual test data dominance and physical test data calibration involves using virtual test data as the main dataset, calculating the average deviation under the same or similar working conditions using physical test data, applying the deviation correction to all virtual test data, and then merging the corrected virtual test data with the physical test data to obtain a fused dataset.

[0092] In step S106, a graph neural network (GNN) is used to perform deep learning on the fused dataset to identify coupled failure paths such as high temperature softening of sealing materials and vibration at a specific frequency causing connector fretting wear, both of which jointly accelerate seal failure. Based on these insights, a lightweight predictive health management model is trained to estimate remaining lifespan.

[0093] Association analysis is performed using graph neural networks, including: Construct a heterogeneous graph with instrument components, stress loads, and failure symptoms as nodes to form a multi-layered correlation network of structure-stress-failure; Define structure-stress edges, stress-failure edges, structure-structure edges, and failure-failure edges in the heterogeneous graph, and introduce physical prior information into the edge weights; A graph neural network is used to propagate and aggregate features in heterogeneous graphs, and node embeddings are learned through message passing mechanism, graph attention mechanism and hierarchical aggregation strategy. Multidimensional feature vectors are constructed for different types of nodes and used as input to the graph neural network for training. By utilizing the node embedding vectors and attention weights of the trained graph neural network, cross-component cascade failure paths caused by multi-stress coupling can be automatically identified.

[0094] The following is a detailed description: Constructing a heterogeneous diagram with instrument components, stress loads, and failure symptoms as nodes includes: Based on the structural composition, stress load types, and failure mechanisms of the logging tool, a heterogeneous graph is constructed. Key elements in the system are abstracted into nodes of different types, forming a multi-layered network of structure-stress-failure relationships. Node types specifically include: Structural nodes: Key components and subsystems of logging tools, such as outer shell, electronic compartment, circuit board, sealing ring, push arm, hinge structure, connector, etc.; these nodes carry the geometric information, material properties and assembly relationships of the components, and are the carriers of failure propagation.

[0095] Stress nodes: The main load factors acting on the structure during downhole and testing processes, including high temperature, temperature gradient, hydrostatic column pressure, impact load, vibration spectrum characteristics, number of push-and-pull cycles, etc. These nodes describe the external excitations and environmental conditions that the logging tool is subjected to during service and are the driving force for failure initiation.

[0096] Failure symptom nodes: These represent abnormal phenomena and signs of degradation characterized by sensor readings or simulation results. Specifically, they include increased seal leakage rate, increased frictional resistance, abnormal strain accumulation, and functional performance degradation. These nodes are observable degradation indicators and are precursor signals for failure.

[0097] Through collaborative modeling of the above three types of nodes, the physical system of the logging tool is completely mapped into the graph space, enabling the subsequent graph neural network to simultaneously process structural information, stress information, and failure information, and capture the complex interaction relationship between the three.

[0098] Various directed or undirected edges are defined in heterogeneous graphs to describe the physical, logical, or statistical relationships between nodes, mainly including: Structural-stress edges: These represent structural components subjected to specific stress loads. For example, an edge between the push arm and vibration indicates that the push arm experiences random vibrations transmitted by the drill string during operation; an edge between the sealing ring and high temperature and pressure indicates that the sealing ring is exposed to a high temperature and high pressure formation environment; an edge between the circuit board and thermal cycling indicates that the circuit board experiences temperature cycling during drilling; and an edge between the outer shell and hydrostatic pressure indicates that the outer shell is subjected to external confining pressure. The existence of these edges allows stress information to flow from stress nodes to structural nodes.

[0099] Stress-failure edges: These represent the driving relationship between stress and failure symptoms. For example, high temperature indicates seal aging, meaning that high temperature causes an aging reaction in the sealing material; thermal cycling indicates weld fatigue, meaning that temperature cycling leads to cumulative fatigue damage at the weld; random vibration indicates hinge wear, meaning that vibration excitation causes abrasive wear on the hinge contact surface; impact loads indicate structural deformation, meaning that impact causes localized plastic deformation of the structure; and push-and-pull cycles indicate actuator degradation, meaning that repeated push-and-pull cycles lead to a decrease in actuator performance. The existence of these edges allows the driving effect of stress on failure to be expressed in heterogeneous diagrams.

[0100] Structural edges: These indicate assembly relationships, force transmission paths, or functional coupling relationships. For example, an edge between a push arm and a hinge indicates a kinematic pair connection; an edge between the outer shell and the electronic compartment indicates the shell encloses the compartment; an edge between a sealing ring and the outer shell indicates the ring is installed in a sealing groove within the shell; an edge between a circuit board and a connector indicates an electrical connection; and an edge between a detector and a circuit board indicates a signal connection. The presence of these edges allows failures to propagate between different components. Failure-failure edge: This indicates a cascaded or common-cause failure relationship. For example, an increase in the seal leakage rate indicates a decrease in insulation resistance, meaning that seal leakage leads to mud intrusion, which in turn reduces insulation resistance; a decrease in insulation resistance indicates increased signal noise, meaning that insulation deterioration leads to a decrease in signal quality; increased frictional resistance indicates insufficient pushing force, meaning that increased friction prevents the effective transmission of pushing force; abnormal strain accumulation indicates a decrease in structural stiffness, meaning that cumulative deformation leads to a reduction in structural stiffness; increased signal noise indicates functional performance degradation, meaning that decreased signal quality leads to a decrease in measurement accuracy. The existence of such edges allows failures to propagate cascading along the causal chain.

[0101] By incorporating prior physical information, such as stress concentration factors, fatigue damage rates, and thermal coupling strength obtained from finite element method (FEM) calculations, the learning process of the Graph Neural Network (GNN) is constrained by engineering mechanisms, avoiding the physical uninterpretability issues arising from purely data-driven approaches. Specifically, the weights of structure-stress edges can be determined based on physical quantities such as stress concentration factors, heat transfer coefficients, or contact pressures obtained from FEM calculations; the weights of stress-failure edges can be based on sensitivity analysis or statistical correlation calculations; the weights of structure-structure edges can be determined based on connection stiffness, force transmission coefficients, or signal coupling coefficients; and the weights of failure-failure edges can be obtained based on fault tree analysis or historical failure data statistics. In this way, the initial knowledge of the GNN is not learned entirely from scratch, but rather undergoes limited adjustments under the guidance of physical priors, thereby ensuring the physical rationality of the learning results.

[0102] For different types of nodes, multi-dimensional feature vectors are constructed as input to the graph neural network. The quality of feature encoding directly affects the learning performance of the graph neural network. Structural node characteristics: Each structural node encodes material parameters, geometric features, historical cumulative damage, and design safety margin. Material parameters include elastic modulus, density, Poisson's ratio, coefficient of thermal expansion, and yield strength; geometric features include characteristic dimensions, surface area, and volume; historical cumulative damage is the cumulative damage value calculated based on wear models, fatigue models, or aging models; and the design safety margin is the ratio of material strength to design stress.

[0103] Stress node characteristics: Each stress node encodes the stress amplitude, spectral characteristics, rate of change, load duration, and number of cycles. The stress amplitude includes the mean, maximum, minimum, and standard deviation; the spectral characteristics include the dominant frequency, bandwidth, and power spectral density; the rate of change describes the rate of change of stress over time; the load duration describes the cumulative duration of stress application; and the number of cycles describes the cycle count of alternating stress.

[0104] Failure Node Characteristics: Each failure node is encoded with a degradation index value, a trend of change, an anomaly threshold offset, and a failure probability estimate. The degradation index value is a quantitative representation of the current degree of degradation; the trend of change describes the direction (increasing, decreasing, or remaining stable) and rate of change of the degradation index over time; the anomaly threshold offset describes the degree of deviation of the current value from the normal threshold or warning threshold; and the failure probability estimate is the likelihood of failure based on historical data or physical test data.

[0105] Edge characteristics: Each edge encodes the intensity and temporal relationship of the action. The intensity of the action includes physical quantities such as the magnitude of contact pressure, thermal conductivity, vibration energy transfer ratio, stress concentration factor, fatigue damage rate, or thermal coupling strength; the temporal relationship includes the time delay of failure propagation and the directionality of causal relationships.

[0106] A graph neural network is used to propagate and aggregate features of the heterogeneous graphs mentioned above. The core design includes a message passing mechanism, a graph attention mechanism, and a hierarchical aggregation strategy.

[0107] Message Passing Mechanism: In each layer of the graph neural network, nodes exchange information through message passing. In the message generation phase, the message passed from the source node to the target node is determined by the features of the source node, the target node, and the edge features. The message function is typically implemented as a multilayer perceptron or a linear transformation. In the message aggregation phase, all neighbor messages received by the target node are aggregated. The aggregation function can employ summation, averaging, maximum value calculation, or attention weighting. In the node update phase, the node features are updated by combining the target node's own features and the aggregated messages. The update function is typically implemented as a gated recurrent unit or a multilayer perceptron.

[0108] Using a message passing mechanism, stress information can propagate along the structural path and accumulate its impact among multiple components. The core formula for message passing is as follows: The first-level message generation formula is that the message passed from source node u to target node v is equal to the result of the message function applied to the initial features of the source node, the initial features of the target node, and the edge features. The first-level node update formula is that the new features of the target node are equal to the result of the update function applied to the initial features of the target node and the aggregated message. This can be expressed as: , , in, For the source node The initial feature vector, For the target node The initial feature vector, For the source node To the target node Edge features, For the first layer message function, For the first-level aggregation function, For the source node Transmitted to the target node The news For the update function of the first layer, For the target node In the feature vectors of the first layer, For the first layer from the source node Transmitted to the target node The news.

[0109] Graph Attention Mechanism: To enable the model to dynamically learn the importance of different neighboring nodes, a graph attention network mechanism is adopted. The calculation of attention weights considers the features of the source and target nodes. The input features are transformed to a higher-dimensional space through learnable attention vectors and weight matrices, and then the matching degree between the two is calculated. After LeakyReLU activation and Softmax normalization, the attention coefficients are obtained.

[0110] The core formula of the graph attention mechanism is as follows: In the second layer, the attention weight of node v to its neighbor node u is equal to the dot product of the attention vector activated by LeakyReLU and the concatenated result of the transformed features, after Softmax normalization on the set of neighbor nodes. The new features of node v in the second layer are equal to the weighted sum of the features of each neighbor node after processing by the non-linear activation function according to the attention weights. Using the graph attention network mechanism, the importance weights of neighbors are dynamically learned, as expressed by the following formula: , , in, Indicates attention weights, This indicates that LeakyReLU is activated. This represents the transpose of the attention vector. This represents the learnable weight matrix. For the source node In the feature vectors of the first layer, For the target node In the feature vectors of the first layer, For nodes In the feature vectors of the first layer, For the target node The set of neighboring nodes, For the target node In the new feature vector of the second layer, This is a learnable attention vector.

[0111] Hierarchical Aggregation Strategy: Considering the hierarchical structure of the logging tool system, a hierarchical aggregation strategy is adopted to infer failure risk from the local component level to the system level. The first layer performs message passing at the component level, capturing the initiation process of local failures, targeting individual seals, solder joints, and hinges. The second layer aggregates the features of component nodes belonging to the same component to obtain component-level features; aggregation functions can employ max pooling, average pooling, or attention pooling. The third layer aggregates the features of component nodes belonging to the same subsystem to obtain subsystem-level features. The fourth layer aggregates all subsystem features to obtain a system-level failure risk score. Through this hierarchical aggregation, the complete propagation path from component-level failure to system-level fault can be identified.

[0112] The graph neural network used is a multi-layered structure. For example, in one embodiment, a three-layer structure is used, including: an input layer that receives the initial features of the heterogeneous graph. The initial features of structural nodes include material parameters, geometric features, cumulative damage, and design safety margin; the initial features of stress nodes include stress amplitude, spectral features, rate of change, and cycle count; and the initial features of failure nodes include degradation indices, trends, anomaly threshold offsets, and failure probability estimates. Edge features include prior physical information such as contact pressure magnitude, thermal conductivity, and vibration energy transfer ratio.

[0113] Layer 1 (Local Perception Layer): In the first layer, message passing is restricted to directly connected nodes. Specifically, information from stress nodes is passed along structure-stress edges to structure nodes, enabling them to perceive the various stresses they experience; information from structure nodes is passed along structure-structure edges to adjacent structure nodes, enabling components to perceive their directly assembled neighboring components; and information from structure nodes is passed along structure-failure edges to failure nodes, enabling failure symptoms to perceive their corresponding structural damage. The output of the first layer is a node embedding vector that incorporates one-hop neighbor information.

[0114] Layer 2 (Component-Level Propagation Layer): In Layer 2, message passing extends to two hops. After processing in Layer 1, stress information has been passed to structural nodes. In Layer 2, these structural nodes carrying stress information propagate the information along structure-to-structure edges to more distant structural nodes. For example, the outer shell node passes its pressure information to the internal electronics compartment node, allowing the electronics compartment to perceive the indirect impact of external pressure. Simultaneously, information from failed nodes begins to propagate along failure-to-failure edges to subsequent failed nodes. For example, information from a node with increasing seal leakage rate is passed to a node with decreasing insulation resistance. The output of Layer 2 is a node embedding that incorporates two-hop neighbor information.

[0115] The third layer (subsystem-level propagation layer): In the third layer, message passing extends to three hops. Stress and failure information propagates over a wider range. For example, high-temperature stress nodes pass through sealing ring nodes, leakage rate increase nodes, and insulation degradation nodes, ultimately transmitting their impact to circuit board functional failure nodes. This layer achieves complete information transmission from local failures to system-level failures. The output of the third layer is a node embedding that incorporates three-hop neighbor information; each node is now able to perceive the status information of the entire subsystem and even the entire machine.

[0116] Output Layer: The node embedding vectors output from the last layer are used as the final result. These node embedding vectors contain complete contextual information for each node within its K-hop range. Depending on the specific task, the output layer can connect to different prediction heads. For node classification tasks, a fully connected layer and a softmax activation function are added after the node embeddings; for node regression tasks (such as remaining lifetime prediction), a fully connected layer and a linear activation function are added after the node embeddings; for graph-level tasks (such as overall system health assessment), all node embeddings are first globally pooled (summed, averaged, or maximized) before being connected to a fully connected layer.

[0117] Each layer includes: an input layer that receives data; and a convolutional layer that uses multiple learnable convolutional kernels (also called filters) to slide across the input data, performing convolution operations to extract local features. The convolution operation involves multiplying the kernel element-wise with a local region of the input data and summing the results to produce a value in the output feature map. The kernel then slides to the next position, repeating this process until the entire input has been traversed. Each convolutional kernel extracts a specific feature, and multiple kernels work in parallel to produce multi-channel output feature maps.

[0118] Key parameters of a convolutional layer include: kernel size (e.g., 3×3 or 5×5), stride (the distance the kernel slides each time), padding (padding zeros around the input boundaries to control the output size), and number of kernels (which determines the number of channels in the output feature map).

[0119] Activation layers introduce nonlinear transformations after convolutional layers, enabling graph neural networks to learn nonlinear relationships. The most commonly used activation function is ReLU.

[0120] Pooling layers are used to reduce the spatial size of feature maps, decrease the number of parameters and computational cost, and enhance the translation invariance of features. Pooling operations perform downsampling within local regions.

[0121] Max pooling takes the maximum value within a local region as the output, preserving the most salient features. Average pooling takes the average value within a local region as the output, preserving the overall trend. Typical parameters of a pooling layer include the pooling window size (e.g., 2×2) and the stride (usually equal to the window size, effectively halving the size).

[0122] Fully connected layers, typically located at the end, combine features extracted by convolutional and pooling layers for the final classification or regression task. In a fully connected layer, each neuron is connected to all neurons in the previous layer.

[0123] Before entering a fully connected layer, the multidimensional feature map of the previous layer is typically flattened into a one-dimensional vector. The output dimension of a fully connected layer depends on the task objective; for example, in a classification task, the output dimension equals the number of classes.

[0124] The fused dataset is mapped to graph samples for training graph neural networks.

[0125] Training data preparation: Each graph sample contains a node feature matrix, an edge index matrix, an edge feature matrix, node labels, and graph-level labels. Node labels include the failure time or failure status of each node; graph-level labels include the overall failure time or health status of the system.

[0126] A multi-task loss function is employed to simultaneously optimize node-level prediction, edge-level prediction, and graph-level prediction. Node-level loss is used to predict the failure probability or remaining lifetime of each node, employing mean squared error or cross-entropy loss; edge-level loss is used to predict the probability or intensity of failure propagation, employing binary cross-entropy loss; graph-level loss is used to predict the overall health status or failure time of the system, employing mean squared error loss; each loss is weighted and summed using weighting coefficients.

[0127] The tag data comes from three sources. First, known failure event times or failure locations, primarily derived from actual failure records observed in physical test data. Second, estimated remaining life of critical components, extrapolated from degradation trajectories in virtual test data. Third, health status levels or failure risk scores, derived from expert knowledge.

[0128] Batch training is employed, grouping multiple small images into a single batch for training. An early stopping mechanism is used: training is stopped and the best model is saved when the validation set loss no longer decreases after several consecutive training epochs. Random deactivation and L2 regularization are used to prevent overfitting.

[0129] After training, the node embedding vectors and attention weights of the graph neural network are used to automatically identify typical coupling failure paths.

[0130] For different types of nodes, the node embedding vector contains different semantic information. For example, the embedding of structural nodes, taking a sealing ring node as an example, includes the following information: First, the sealing ring's own attribute information, including material type (e.g., nitrile rubber), geometric dimensions (e.g., cross-sectional diameter, inner diameter), elastic modulus, hardness, etc. Second, the sealing ring's current state information, including cumulative aging degree, compression set rate, current leakage rate, etc. Third, the stress information borne by the sealing ring, obtained from the stress node through message passing, including current temperature, pressure amplitude, vibration level, etc. Fourth, the connection relationship information between the sealing ring and adjacent components, including the assembly relationship with the outer shell and the positional relationship with the electronic compartment, etc. Fifth, the impact information of sealing ring failure on downstream components, obtained through multi-layer message passing back propagation from the failure node, including the impact on insulation resistance after leakage and the potential threat to the circuit board, etc.

[0131] Embedding of stress nodes: Taking high-temperature stress nodes as an example, their embedding vector contains the following information: First, the properties of the high-temperature stress itself, including temperature amplitude, duration, and rate of change. Second, the information transmitted from the structural components subjected to the high-temperature stress to structural nodes such as seals, circuit boards, and shells via message passing, and then back. Third, the failure symptoms caused by high-temperature stress, including seal aging, material softening, and thermal expansion.

[0132] Embedding of Failure Nodes: Taking a node with rising leakage rate as an example, its embedding vector contains the following information: First, the attributes of the failure symptom itself, including the current leakage rate value, its trend, and its distance from the threshold. Second, the stress causing the failure symptom, obtained from the stress node through message passing, including high temperature and pressure. Third, the subsequent failures affected by this failure symptom, including insulation degradation and increased signal noise.

[0133] Attention weights assign an importance score to each edge, representing the strength of information propagation along that edge. In failure path identification, starting from the initial node (such as a high-temperature stress node), path tracing is performed along edges with high attention weights. Specifically, at the current node, all its outgoing edges are examined, and the edge with the highest attention weight is selected as the next propagation direction. This process is repeated until a system-level failure node is reached or the maximum path length is reached.

[0134] Node embedding vectors and attention weights work together in failure path identification. For example, taking the coupled failure path "high temperature → performance degradation of sealing material → vibration causing micro-leakage → internal circuit moisture → functional failure" as an example, we can illustrate the role of node embedding vectors and attention weights in the failure path identification process.

[0135] The process includes: First, after training, the graph neural network contains temperature amplitude and duration information in the embedding vector of the high-temperature stress node; the embedding vector of the sealing ring structure node contains material properties, current aging degree and high temperature information; the embedding vector of the leakage rate increase node contains the current leakage value and change trend; the embedding vector of the insulation decrease node contains the resistance value and decrease rate; and the embedding vector of the functional failure node contains the system availability status information.

[0136] The second step is to calculate the attention weights. Edges from high-temperature nodes to sealing ring nodes receive a high attention weight (e.g., 0.85) because high temperature is a major factor leading to seal aging. Edges from sealing ring nodes to nodes with increasing leakage rates receive a high attention weight (e.g., 0.91) because seal aging directly causes leakage. Edges from vibration nodes to nodes with increasing leakage rates also receive a relatively high weight (e.g., 0.68) because vibration accelerates the formation of microleakage. Edges from nodes with increasing leakage rates to nodes with decreasing insulation receive a relatively high weight (e.g., 0.76), and edges from nodes with decreasing insulation to nodes with functional failure receive a relatively high weight (e.g., 0.95).

[0137] The third step involves verifying the path using node embedding. The consistency of the node embeddings along the path is examined: a high similarity exists between the high-temperature node embedding and the sealing ring node embedding, indicating a close correlation between the two; a high similarity exists between the sealing ring node embedding and the leakage rate increase node embedding, indicating a direct correlation between the sealing condition and leakage symptoms; a high similarity exists between the leakage rate increase node embedding and the insulation degradation node embedding, indicating that leakage does indeed lead to insulation deterioration. This consistency in the embedding space verifies the physical rationality of the path.

[0138] The fourth step is to output the recognition results. The high-attention weighted edges are concatenated and combined with the semantic coherence verification of node embeddings to finally output the complete coupling failure path, and each node on the path is given key dimension information of the embedding vector.

[0139] In one embodiment, an attention path tracing method is employed, starting from the initial failure node and tracing the path along edges with high attention weights. At each current node, all outgoing edges are examined, and the edge with the highest attention weight is selected as the propagation path. This process is repeated until a system-level failure node is reached or the maximum path length is achieved. Alternatively, a gradient saliency graph method can be used to calculate the gradient of the graph-level output with respect to the features of each node. Nodes with large gradient magnitudes indicate a significant contribution to system failure; these high-gradient nodes are connected to form critical failure paths. System-level failure nodes refer to nodes representing the overall failure state of the logging tool. The graph-level output is a predicted value or a vector for the entire graph. The graph-level output focuses on the global attributes of the entire graph, representing the system-level state determined by all nodes and edges.

[0140] For example, the above method can identify coupled failure paths such as high temperature causing performance degradation of sealing materials, vibration causing micro-leakage, which in turn leads to moisture absorption of internal circuitry, ultimately resulting in functional failure. This coupled failure path reveals how the stresses of high temperature and vibration work synergistically to accelerate seal failure and subsequent electrical faults.

[0141] By analyzing the high-weighted edges and critical nodes in the diagram, the complete evolution process of failure from "stress source - structural weak point - symptom manifestation" can be intuitively revealed. Specifically, the following information can be identified: what kind of stress or initial defect is the starting node of the failure path; which component or failure symptom does each intermediate node in the path represent; the attention weight on each edge indicates the importance of that propagation step; and the cumulative length or cumulative weight of the path reflects the overall risk from initial failure to system failure.

[0142] In one embodiment, the node embedding vectors and failure probabilities output by the graph neural network are used as input features to provide a subsequent lightweight predictive health management model. This enables real-time prediction of the remaining useful life (RUL) of critical components; dynamic assessment of system-level health indices; and early warning of potential cascading failure risks.

[0143] The predictive health management model includes a remaining life expectancy prediction branch, a health index assessment branch, and a failure warning branch, which respectively output the remaining life expectancy prediction value, health index score, and warning status.

[0144] The training objective of the predictive health management model is to minimize the errors in remaining life expectancy prediction and health status assessment. The model employs a lightweight design for eventual deployment on the edge computing unit of a logging tool. The overall architecture of the predictive health management model consists of four parts: an input layer, a feature processing layer, a prediction layer, and an output layer. Specifically: Input Layer: Receives two types of information from the graph neural network output. The first type is the node embedding vector. In one example, each node's embedding dimension is 128, containing the node's contextual information within the graph structure, including its own attributes, stress, assembly connections, and failure propagation path information. The second type is the node failure probability, ranging from 0 to 1, representing the likelihood of each node failing under the current stress conditions.

[0145] Feature processing layer: This layer further extracts and reduces the dimensionality of the input node embedding vectors. Since the input features have already been sufficiently processed by the graph neural network, the feature processing layer adopts a simple structure, such as a fully connected layer to reduce the 128-dimensional embedding to 64 dimensions, followed by a batch normalization layer to accelerate training convergence, and finally a random deactivation layer to prevent overfitting. The random deactivation rate is set to 0.2.

[0146] Prediction Layer: Depending on the task type, the prediction layer employs different network structures. For remaining lifetime prediction tasks, a regression network outputs a continuous numerical value representing the remaining lifetime in hours or cycles. For health index assessment tasks, a regression network outputs a continuous value between 0 and 100, representing the system's health level. For failure warning tasks, a classification network outputs a binary signal indicating whether a warning has been triggered.

[0147] Output layer: Outputs corresponding results based on the specific task. Remaining life prediction outputs a value in hours or cycles, health index assessment outputs a score between 0 and 100, and failure warning outputs a binary status of normal or warning.

[0148] Remaining life prediction is one of the core tasks of this application, which aims to predict the remaining working time or number of working cycles of each key component of the logging tool from the current moment to the time of failure.

[0149] In one embodiment, the input to the remaining lifetime prediction branch is the node embedding vector output by the graph neural network, which is processed through two fully connected layers. The first fully connected layer maps the 128-dimensional embedding to 64 dimensions, followed by a ReLU activation function and a random deactivation layer. The second fully connected layer maps the 64-dimensional features to 32 dimensions, followed by a ReLU activation function. The third fully connected layer maps the 32-dimensional features to 1 dimension, outputting the remaining lifetime prediction value.

[0150] The health index assessment branch takes as input the aggregated result of all node embedding vectors. First, global pooling is performed on all node embedding vectors output by the graph neural network; average pooling, max pooling, or attention pooling can be used. This embodiment uses average pooling, averaging the embedding vectors of all nodes by dimension to obtain a 128-dimensional global feature vector. Then, the global feature vector passes through two fully connected layers. The first fully connected layer maps the 128 dimensions to 64 dimensions, followed by a ReLU activation function and a random deactivation layer. The second fully connected layer maps the 64 dimensions to 32 dimensions, followed by a ReLU activation function. The third fully connected layer maps the 32 dimensions to 1 dimension, outputting the health index. The output range of the health index is mapped to 0 to 1 using a sigmoid activation function, then multiplied by 100 to convert it into a score of 0 to 100. A higher score indicates a better health status, while a lower score indicates closer to failure.

[0151] The failure warning system aims to monitor the status of the logging tool in real time and issue an alarm signal in time before a potential failure occurs.

[0152] The input to the failure warning branch is the same as that to the health index evaluation branch, which is the global pooling result of the embedding vectors of all nodes. The warning branch adopts a binary classification network structure and outputs two states: normal or warning.

[0153] In one example, the classification network consists of two fully connected layers. The first fully connected layer maps the 128-dimensional global features to 32 dimensions, followed by a ReLU activation function and a random deactivation layer. The second fully connected layer maps the 32 dimensions to 2 dimensions, corresponding to the two categories of "normal" and "alert". The output layer uses a Softmax activation function to convert the outputs of the two categories into probability distributions, and takes the category with the higher probability as the final prediction.

[0154] The loss function used is cross-entropy loss, which is suitable for classification tasks.

[0155] The training process for a predictive health management model is as follows: Training data preparation: The node embedding vectors and node failure probabilities output by the graph neural network are used as input features. Training samples are derived from a fused dataset, including virtual test data and physical test data. Each sample contains input features and corresponding labels, including the true values ​​of remaining lifespan, health index, and failure status.

[0156] Multi-task joint training is employed, with simultaneous optimization of all three prediction branches. The optimizer is Adam, with an initial learning rate of 0.001, dynamically adjusted using cosine annealing. The batch size is set to 32 or 64, determined based on available GPU memory. The training epochs are set to 100, employing an early stopping mechanism: training ceases when the validation set loss no longer decreases for 20 consecutive epochs.

[0157] To prevent overfitting, random deactivation was employed with a random deactivation rate of 0.2. L2 regularization was used with a weight decay coefficient of 0.0001. An early stopping mechanism was employed to select the best-performing predictive health management model on the validation set.

[0158] The trained and optimized predictive health management model is converted into code suitable for the logging tool's main control chip by a compiler and embedded in the firmware. When the logging tool is working downhole, it can use local sensor data to calculate its own health status in real time and report early warnings when abnormalities occur, realizing the transformation from scheduled maintenance to predictive maintenance.

[0159] In summary, this application, based on digital twin technology, constructs an innovative method for logging tool reliability. It enables deep involvement of reliability work in the early design stages, achieving more accurate forward-looking predictions of logging tool reliability and more effective assurance enhancements.

[0160] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for reliability analysis of logging tools based on digital twins and dynamic stress boundaries, characterized in that, include: Constructing a digital twin model of a logging tool involves establishing a three-dimensional geometric model based on engineering drawings, embedding material properties, dynamic stress boundaries, behavioral logic, and fault rules generated based on historical operating data into the three-dimensional geometric model, forming a multi-physics simulation digital twin model, and superimposing degradation laws into the digital twin model. The degradation laws include wear models, fatigue models, and aging models. The wear models, fatigue models, and aging models work collaboratively through the multi-physics coupling mechanism of the digital twin model. The wear models, fatigue models, and aging models share a unified simulation time axis, the same stress field and temperature field, and influence each other through a real-time parameter exchange mechanism. Establishing a virtual-physical interactive connection to achieve bidirectional data synchronization and closed-loop control between the physical entity of the logging tool and the digital twin model includes: the physical entity of the logging tool uploading the sensed data information to the OPC UA server; the OPC UA server driving the virtual layer of the digital twin model and the application layer of the twin data; and the virtual layer of the digital twin model establishing a connection with the field equipment connected to the OPC UA server using the OPC UA client. The digital twin model is used to conduct virtual acceleration testing and obtain virtual test data; Using the aforementioned dynamic stress boundary, high-acceleration life testing was performed on the physical prototype, and physical test data was collected simultaneously. The physical test data is compared and verified with the virtual test data, and the simulation process is adjusted according to the comparison and verification results to regenerate the virtual test data. By integrating virtual test data and physical test data, graph neural networks are used for correlation analysis to identify coupled failure modes and build a predictive health management model.

2. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, The generation of dynamic stress boundaries includes: Collect historical operating data, which includes at least depth-temperature-pressure curves, downhole vibration spectrum, tripping and drilling operation parameters, formation characteristics, and instrument status. Identify pressure stress, temperature stress, vibration and shock stress, and cyclic stress as key stress parameters; Statistical analysis was performed on the historical working conditions big data corresponding to each key stress parameter, the probability distribution function of each key stress parameter was fitted, the independent distribution model of each key stress parameter was obtained, and the Monte Carlo method combined with the Copula function was used to generate a stress vector sample set. Based on the stress vector sample set, a static stress envelope containing typical working condition envelopes, severe working condition envelopes, and extreme working condition envelopes is generated, and the static stress envelope is reconstructed into a time-varying dynamic sequence. Multiple stress combinations are defined as simulation scenarios, and each simulation scenario contains a complete time-varying stress sequence.

3. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, The embedding of fault rules includes: Establish a failure mode and impact analysis rule base for logging tools. In the failure mode and impact analysis rule base, high-frequency failure modes in the field are mapped to executable rules. Each executable rule includes a trigger threshold, failure rate function, impact path and a list of associated failures at the next level, and is injected into the digital twin model in the form of XML fragments. High-frequency failure modes in the field include, but are not limited to: high-temperature aging and leakage of sealing rings, abrasive jamming of push arm hinges, and thermal cycling fatigue of circuit board solder joints. Trigger thresholds include one or more combinations of temperature threshold, vibration amplitude threshold, pressure threshold, and cycle number threshold.

4. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, The wear model, fatigue model, and aging model work collaboratively through a multiphysics coupling mechanism of a digital twin model, specifically including: When the leakage rate of the sealing ring calculated by the aging model exceeds the mud intrusion threshold, the mud intrusion flag is activated, and the particle concentration parameter in the mud is transmitted to the wear model to dynamically adjust the wear coefficient and friction coefficient in the wear model. When the wear model calculates that the hinge wear produces wear particles, the particle size and concentration parameters of the wear particles are fed back to the sealing interface to correct the leakage rate of the sealing ring, forming a positive feedback coupling loop between seal aging and hinge wear. When the leakage rate calculated by the aging model exceeds the threshold, the thermal boundary conditions of the electronic compartment are updated. The fatigue model reads the updated temperature field, recalculates the thermal stress cycle amplitude of the solder joint, and adjusts the cumulative damage. When the cumulative damage degree of the weld joint output by the fatigue model reaches a preset threshold, causing the temperature sensor signal to be interrupted, the aging model switches to virtual sensor interpolation mode or historical data prediction mode to maintain continuous calculation of the aging degree. Virtual sensor interpolation mode refers to using the measurement data of other related position sensors in the same system to calculate the theoretical measurement value of the failed sensor position through interpolation algorithm when the physical sensor fails or the signal is interrupted, thereby replacing the real sensor data. Historical data prediction mode refers to using the data pattern of the physical sensor in its historical normal operation, combined with the current operating parameters, to predict the theoretical measurement value at the current moment when the physical sensor fails and effective interpolation cannot be performed through adjacent sensors. System-level failure determination adopts a comprehensive criterion: the system is determined to be in failure when any model output reaches the damage threshold, or when the outputs of two or more models reach the warning threshold, or when the cumulative damage on the coupled propagation path exceeds the threshold.

5. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, Virtual acceleration testing is performed using the digital twin model to obtain virtual test data, including: The ultimate stress scanning test is carried out, which includes temperature limit scanning, vibration limit scanning, pressure limit scanning and composite stress limit exploration, and the operational limit and destructive limit under each single stress and composite stress are recorded respectively. Using the operating limit and the failure limit as the upper limit of the accelerated stress, accelerated degradation simulation is performed. Accelerated stress is applied using a stepped stress profile or a cyclic stress profile, and the acceleration factor of each failure mode is calculated. The accelerated failure time under accelerated conditions is extrapolated to the failure time under normal operating conditions through the acceleration factor. Output virtual test data, which includes at least a failure hotspot list, failure time or cycle number prediction, and failure mode distribution.

6. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, The simulation process is adjusted based on the results of the comparison and verification, including: calibrating the material parameters, boundary conditions, or degradation model parameters of the digital twin model, and adjusting the deviation between the virtual test data and the physical test data.

7. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, Association analysis is performed using graph neural networks, including: Construct a heterogeneous graph with instrument components, stress loads, and failure symptoms as nodes to form a multi-layered correlation network of structure-stress-failure; Define structure-stress edges, stress-failure edges, structure-structure edges, and failure-failure edges in the heterogeneous graph, and introduce physical prior information into the edge weights; A graph neural network is used to propagate and aggregate features from heterogeneous graphs through message passing, graph attention, and hierarchical aggregation strategies. Multidimensional feature vectors are constructed for different types of nodes and used as input to the graph neural network for training. By utilizing the node embedding vectors and attention weights of the trained graph neural network, cross-component cascade failure paths caused by multi-stress coupling can be automatically identified.

8. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, Automatic identification of cross-component cascade failure paths caused by multi-stress coupling includes: using the attention path tracing method, starting from the initial failure node, tracing the path along the edges with high attention weights, examining all outgoing edges at each current node, selecting the edge with the highest attention weight as the propagation path, moving to the next node, and repeating this process until the system-level failure node is reached or the maximum path length is reached; or using the gradient saliency graph method, calculating the gradient of the graph-level output with respect to the characteristics of each node, with nodes having large gradient magnitudes indicating that they contribute significantly to system failure, and connecting these high-gradient nodes to form critical failure paths.

9. The logging tool reliability analysis method based on digital twin and dynamic stress boundary as described in claim 1, characterized in that, The construction of a predictive health management model includes: using the node embedding vector and failure probability output by the graph neural network as input features and inputting them into the predictive health management model; the predictive health management model includes a remaining life prediction branch, a health index assessment branch, and a failure warning branch, which respectively output the remaining life prediction value, health index score, and warning status.

Images