Intelligent life prediction and optimization system and method for steam turbine rotor welded joint

Abstract

The invention discloses an intelligent life prediction and optimization system and method for a steam turbine rotor welded joint. The system comprises a multi-source data acquisition module, a digital twin modeling module, a health state evaluation and life prediction module, a risk early warning module and an operation collaborative optimization module. The multi-source data acquisition module acquires the multi-dimensional physical quantity of the rotor welding joint in real time. The digital twin modeling module establishes a high-fidelity virtual model and realizes real-time synchronization and correction of a physical entity and a digital model. And the health state evaluation and life prediction module is used for calculating a damage accumulation rate and a health index based on fusion of a physical model and an LSTM neural network so as to realize residual life estimation. And the risk early warning module performs graded early warning according to the dynamic trend of the health state. And the operation collaborative optimization module adaptively adjusts operation parameters and optimizes the unit efficiency based on a reinforcement learning algorithm. All the modules are interconnected through an industrial network to form a closed loop, and real-time monitoring, intelligent evaluation and active service life management of the rotor welding joint are achieved.

Description

Technical Field

[0001] This invention belongs to the field of steam turbine operation safety and intelligent maintenance technology, and in particular relates to an intelligent life prediction and optimization system and method for steam turbine rotor welded joints. Background Technology

[0002] With the widespread application of high-parameter, large-capacity steam turbine generator sets, their rotors operate under extreme conditions of high temperature, high pressure, and high speed for extended periods. For ease of manufacturing and maintenance, modern large rotors often adopt segmented welded structures. The welded joints, as key connections, bear the complex coupling effects of centrifugal force, thermal stress, vibration stress, and high-temperature creep, making them sensitive areas for the initiation and propagation of fatigue cracks and creep damage.

[0003] Existing turbine condition monitoring systems primarily focus on macroscopic parameters such as overall rotor vibration, eccentricity, or bearing temperature, lacking specialized and effective monitoring of welded joints—a critical weak point in localized environments. Traditional life assessments are mostly based on static calculations during the design phase or offline inspections after operation, failing to reflect the dynamic damage accumulation process of joints under actual variable operating conditions. Furthermore, existing safety monitoring and operation control systems are independent of each other, making it impossible to intelligently adjust operating strategies to extend the service life while ensuring the structural integrity of the rotor.

[0004] Therefore, there is an urgent need in this field for a comprehensive solution that can deeply integrate multi-source monitoring information, combine high-fidelity digital twin models to realize dynamic stress state mapping of rotor welded joints, and achieve accurate life prediction and collaborative optimization of operation through intelligent algorithms. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent life prediction and optimization system and method for steam turbine rotor welded joints, so as to solve the technical problems of insufficient monitoring of steam turbine rotor welded joints, low life prediction accuracy and disconnection of operation control in the prior art.

[0006] To address the aforementioned technical problems, the specific technical solution of the intelligent life prediction and optimization system and method for turbine rotor welded joints of the present invention is as follows:

[0007] A smart life prediction and optimization system for steam turbine rotor welded joints includes: a multi-source data acquisition module, a digital twin modeling module, a health status assessment and life prediction module, a risk warning module, and an operation collaborative optimization module;

[0008] The multi-source data acquisition module is used to collect multi-dimensional physical quantities and operating condition parameters of the turbine rotor welded joint during operation;

[0009] The digital twin modeling module is used to establish a high-fidelity virtual model of the rotor welding joint to achieve real-time synchronization, mutual correction and state inversion between the physical entity and the digital model.

[0010] The health status assessment and life prediction module is based on the integration of physical models and intelligent algorithms to realize the quantification of the health status and the estimation of the remaining life of the welded joint.

[0011] The risk warning module is used to classify the risks and provide safety warnings for the turbine rotor welded joints based on the dynamic changes in health status.

[0012] The operation coordination optimization module, while ensuring safety, adaptively adjusts operating parameters based on reinforcement learning algorithms to achieve optimal control of turbine operating efficiency.

[0013] The modules are interconnected via industrial Ethernet or high-speed data bus to form a complete closed-loop system of "perception-modeling-evaluation-optimization-control" to achieve real-time monitoring, intelligent evaluation and proactive optimization of rotor welding joints.

[0014] Furthermore, the multi-source data acquisition module includes a high-temperature wireless strain sensing unit, an acoustic emission detection unit, a vibration monitoring unit, and a steam parameter acquisition unit arranged in and near the weld joint.

[0015] The strain sensing unit is used to measure the strain signal of the rotor material in real time under rotation, heating and loading conditions.

[0016] The acoustic emission detection unit is used to monitor the transient high-frequency acoustic signals generated by the initiation, propagation, and interaction of microcracks inside metallic materials.

[0017] The vibration monitoring unit is located at the connection between the rotor bearing end and the housing, and is used to measure radial and axial vibration acceleration.

[0018] The steam parameter acquisition unit includes a temperature sensor, a pressure sensor, and a flow sensor, used to acquire operating condition information such as main steam temperature, main steam pressure, and steam flow rate;

[0019] The multi-source data acquisition module achieves stable acquisition and transmission of rotating body signals through wireless transmission or slip ring signal extraction. All sensor signals are synchronized, filtered, and feature extracted by edge computing nodes. The extracted features include strain cycle counting, acoustic emission event energy, vibration amplitude spectrum distribution, and steam temperature fluctuation rate. The extracted feature parameters are transmitted to the digital twin modeling module via industrial Ethernet to achieve real-time linkage between the data layer and the model layer.

[0020] Furthermore, the digital twin modeling module of the present invention includes a structure mapping unit, a virtual-real synchronization unit, and a state inversion unit;

[0021] The structural mapping unit is based on the three-dimensional CAD geometric model of the turbine rotor to establish a finite element simulation model including the welded joint area. The model is divided into a refined mesh to ensure the geometric accuracy of the weld and heat-affected zone. During the modeling process, the material parameters are input in a temperature-dependent form.

[0022] The virtual-real synchronization unit is used to apply the physical quantities collected in real time as boundary conditions to the digital twin, drive the model to perform dynamic simulation calculations, and realize the virtual-real synchronous evolution. The solution content of the model includes temperature field, stress field and creep rate field.

[0023] The state inversion unit is used to compensate for missing sensor data using model results. By fitting and interpolating historical data, it adaptively adjusts material parameters based on operating data and inverts the temperature or stress parameters of missing measurement points to achieve self-recovery of virtual measurement points.

[0024] The governing equations of the model include the thermal equilibrium equation and the stress constitutive equation:

[0025] The heat balance equation is:

[0026]

[0027] in, For material density, For specific heat capacity, For temperature field, Thermal conductivity, For internal heat source items, This is the rotating heat source term.

[0028] Stress constitutive equation:

[0029]

[0030] in, The stress tensor represents the interaction forces within the material per unit area and is a fundamental quantity reflecting the internal mechanical state of a welded joint.

[0031] It is the elastic constant tensor, used to characterize the linear response relationship between stress and elastic strain;

[0032] The total strain is used to describe the overall deformation of the welded joint under various loads.

[0033] Thermal strain is used to describe the free expansion or contraction deformation of a material caused by temperature changes.

[0034] Creep strain reflects the time-dependent plastic deformation of a material under high temperature and long-term load.

[0035] Strain generated by centrifugal force is used to describe structural deformation caused by the body force field generated by high-speed rotation;

[0036] The digital twin model periodically updates the calculation results through a finite element solver and performs bidirectional correction with the measured data. When the measured temperature deviates from the model's predicted value by more than a threshold, the system automatically adjusts the material parameters or boundary conditions to ensure consistency between the virtual and real data.

[0037] The digital twin model supports online operation and batch processing modes. When the system is under high load, a simplified model is used for rapid calculation. When the unit is in steady-state operation, it automatically switches to a high-fidelity model for detailed solution to balance calculation accuracy and speed.

[0038] Furthermore, the health status assessment and life prediction module uses a centrifugal force-thermal-fatigue-creep coupling model as its physical basis. By analyzing the comprehensive damage process of materials under high temperature, high stress, and multi-cycle conditions, it calculates the damage accumulation rate D(t) of the joint.

[0039]

[0040] in, To comprehensively consider the equivalent stress of mechanical stress, thermal stress, and centrifugal stress;

[0041] Fatigue life is the number of cycles a material can withstand under a given stress amplitude and temperature.

[0042] Creep rate is the increment of creep strain per unit time under certain stress and temperature.

[0043] The creep coupling coefficient represents the degree of mutual influence between fatigue and creep damage.

[0044] By accumulating the integral, the total damage D(t) within time t is obtained. Its value increases monotonically with the increase of running time. When D(t)=1, it is defined as the end of the failure life.

[0045] The Health Index (HI) is defined to characterize the remaining health status of materials; a lower value indicates more severe structural deterioration.

[0046]

[0047] An early warning is triggered when HI is less than 0.6, and the system enters protection mode when it is less than 0.4. The early warning signal is received by the control system and used for automatic cooling, load reduction, or steam flow control.

[0048] The system calculates the health index change curve over time in real time, and further calculates the health index change rate ΔHI to describe the rate of change of the health index over time, reflecting the dynamic trend of the health status of the turbine rotor welded joint. Its calculation formula is as follows:

[0049]

[0050] in, The current health index. The health index at the previous moment. The time interval between two consecutive monitoring values ​​reflects the health trend; a negative value indicates deterioration, and a positive value indicates recovery.

[0051] When ΔHI remains negative and its absolute value increases, it indicates that the damage to the welded joint has intensified; if ΔHI approaches zero or a positive value, it indicates that the operating condition has stabilized or has recovered somewhat.

[0052] The lifetime decay rate η represents the proportion of damage increase per unit time and is an important indicator reflecting the rate of lifetime depletion. It is defined as:

[0053]

[0054] The larger η is, the faster the lifespan is consumed, and the more likely the system is to fail;

[0055] A Long Short-Term Memory (LSTM) network model is introduced for prediction correction. The LSTM takes feature variables as time series input and, by capturing their temporal correlation and long-term memory characteristics, outputs a dynamic correction coefficient for the lifetime decay rate η.

[0056] The lifetime degradation rate correction process is expressed as follows:

[0057]

[0058] in, The theoretical life decay rate is calculated based on the centrifugal force-heat-fatigue-creep coupling model.

[0059] These are the dynamic correction coefficients output by the LSTM network. This indicates that the physical model underestimates the damage rate; if This indicates that the model overestimated the damage rate; if This means that the model prediction is consistent with reality and no correction is needed;

[0060] This is the corrected lifetime decay rate result.

[0061] Furthermore, the risk warning module uses health index, health index change rate and lifespan decline rate as input parameters, and combines historical trends and threshold judgment rules to generate different warning levels.

[0062] The system is set to four risk levels: Level I, Level II, Level III and Level IV, which correspond to green, yellow, orange and red status indicators, respectively;

[0063] When the system is in Level I condition, the welded joints are in good health, HI≥0.8, ΔHI<0.01, and the life decay rate η is less than 5%, and the system maintains normal operation.

[0064] When the system enters Level II, if HI is between 0.6 and 0.8 or ΔHI ≥ 0.02 and the life decay rate η ≥ 15%, it indicates that the joint health has slightly declined or fatigue damage has begun to accumulate. The system issues a warning signal and prompts the operators to pay attention to load fluctuations.

[0065] When the system enters Level III, if HI is between 0.4 and 0.6 or ΔHI ≥ 0.05 and the life decay rate η ≥ 30%, it indicates that the welded joint has a significant deterioration trend. The system will activate the audible and visual alarm and record the alarm log.

[0066] When the system enters Level IV state, if HI<0.4 or ΔHI≥0.10 and the life decay rate η≥50%, it indicates that there is a serious risk of fatigue or creep damage. The system immediately issues a load reduction or shutdown command to the unit DCS to prevent structural damage.

[0067] The calculation cycle of the risk warning module is matched with the monitoring frequency; the module adopts a multi-dimensional logic judgment algorithm to comprehensively analyze the changing trends of HI, ΔHI and η, so as to avoid false alarms caused by fluctuations of a single parameter.

[0068] The risk warning module also includes an adaptive threshold adjustment mechanism. During long-term operation, the system automatically calculates the distribution characteristics of HI, ΔHI, and η, and corrects the warning boundary based on the operational stability.

[0069] When the risk level reaches Level III or Level IV, the system will automatically trigger the event reporting module to send a risk event information package to the operation management platform.

[0070] The risk warning module also interacts with the monitoring information system to achieve comprehensive monitoring of the status of all equipment in the plant.

[0071] Furthermore, the operation collaborative optimization module is built based on reinforcement learning algorithm, takes the real-time operating status of the steam turbine as the environment and the control parameters as the actions, and obtains the optimal operation strategy through continuous interaction;

[0072] The module's input state variables include main steam temperature, main steam pressure, flow rate, load, speed, health index HI, life decay rate η, and risk level. ;

[0073] Output control variables include main steam temperature setpoint correction, load adjustment, and speed regulation command;

[0074] Reinforcement learning algorithms update policies through trial and error and feedback, aiming to maximize the cumulative reward function.

[0075]

[0076] in, Unit efficiency is calculated using main steam temperature, pressure, and exhaust steam enthalpy, and is an indicator reflecting the unit's thermal economy.

[0077] These are risk factors, risk cost parameters, and their corresponding relationships with risk levels are as follows: Level I = 0, Level II = 1, Level III = 5, Level IV = 50, used to penalize high-risk states;

[0078] ΔHI represents the rate of change of the health index, which indicates the speed of health deterioration. If ΔHI is negative and its absolute value increases, it means that the equipment is deteriorating faster, and the algorithm will reduce the reward value to encourage the strategy to operate more safely.

[0079] The parameters α, β, and γ are weighting coefficients used to balance the relationship between operational efficiency, risk control, and health degradation.

[0080] In the initial stage of the module, offline training is performed using historical operating data to establish a basic policy network; during actual operation, the system updates the policy online and gradually approaches the optimal solution.

[0081] The training process employs the policy gradient method, whereby the agent adjusts the policy distribution π(a|s) based on real-time feedback to achieve dynamic optimization of action selection.

[0082] When the risk level increases or the health status deteriorates, the algorithm automatically reduces the load setting or the main steam temperature, thereby reducing the stress level and delaying the accumulation of damage.

[0083] When the system is running stably and the risk is low, reinforcement learning algorithms tend to improve efficiency metrics. Economic optimization is achieved by fine-tuning the main steam parameters;

[0084] The operation coordination optimization module has continuous learning capabilities and can automatically adapt to new operating characteristics under different seasons and load conditions.

[0085] The module includes a safety protection constraint layer to limit the range of motion output and ensure that any control adjustment will not cause parameter overruns or operational shocks.

[0086] The system also includes a policy review mechanism, allowing operators to view the current reinforcement learning policy output and corresponding weight parameters, thus achieving human control and interpretability;

[0087] The collaborative optimization module and the risk warning module achieve closed-loop linkage: when the warning level rises, the optimization module automatically enters the safety priority mode; when the risk level falls, the system gradually restores the efficient operation mode.

[0088] The operation collaboration optimization module has a learning transfer function, which can transfer the trained strategy model to similar units, shorten the deployment cycle, and improve the system's versatility and economy;

[0089] The system adopts a unified time synchronization protocol to ensure data consistency between modules. Each module has a built-in timestamp mechanism to ensure that all data sampling points correspond accurately to the simulation calculation time.

[0090] The present invention also discloses a method for implementing the intelligent life prediction and optimization system for turbine rotor welded joints, comprising the following steps:

[0091] Step 1: Real-time monitoring data of the rotor welding joint is collected using a multi-source data acquisition module;

[0092] Step 2: Input the data into the digital twin model and perform virtual-real synchronous calculations to generate the temperature field, stress field, and creep rate field of the rotor welded joint;

[0093] Step 3: Calculate the life decay rate using the centrifugal force-heat-fatigue-creep coupling model, and dynamically correct the prediction results using an LSTM neural network.

[0094] Step 4: Based on the calculated Health Index (HI) and its changing trend, and combined with preset thresholds, classify and assess risk levels.

[0095] Step 5: Run the collaborative optimization module to execute the reinforcement learning optimization strategy, dynamically adjust the operating parameters, and realize the integrated collaborative control of rotor welding joint life and unit operating efficiency.

[0096] Furthermore, step 2 includes the following steps:

[0097] The acquired signals undergo real-time preprocessing at edge computing nodes, including filtering and noise reduction, time alignment, feature extraction, and data compression.

[0098] Filtering and denoising: Multi-level filtering modules are set up in the edge computing nodes. First, a low-pass filter is used to suppress high-frequency pseudo signals and retain low-frequency components that reflect load changes. Second, wavelet threshold filtering algorithm is used to decompose and reconstruct transient impacts and high-frequency noise at multiple scales. When there is significant signal drift, Kalman filtering algorithm is further introduced to dynamically estimate and correct the signal based on the system dynamics equation.

[0099] Time alignment: A unified timestamp mechanism is used to synchronize multi-source signals. Each sampled data is marked with a high-precision clock and aligned on a unified time axis through interpolation and resampling algorithms. For asynchronous signals that are out of sync, the system uses linear interpolation and spline function reconstruction methods to generate equal-time data streams, ensuring that each physical quantity has a consistent reference benchmark at the same time. The time-aligned signals are uniformly input into the virtual-real synchronization calculation module to achieve consistency of multi-physics boundary conditions.

[0100] Feature extraction: For strain signals, strain amplitude, root mean square value and cycle count results are extracted; for acoustic emission signals, event count rate, amplitude distribution and energy density are extracted; for temperature and pressure signals, moving average, rate of change and periodic fluctuation features are extracted. These feature variables are used as input parameters for fatigue-creep coupling model and long short-term memory network, thereby realizing health status assessment and life prediction.

[0101] Data compression: The preprocessed data is compressed and stored by combining adaptive sampling and wavelet compression. When the monitored signal is in a steady state, the system automatically reduces the sampling rate and records only key change points. When a sudden fluctuation or abnormal event is detected, the system instantly increases the sampling frequency to capture the complete waveform.

[0102] The preprocessed data is input into the digital twin model, and the temperature field, stress field, and creep rate field of the welded joint under the current operating condition are obtained through finite element calculation.

[0103] Furthermore, step 3 includes the following steps:

[0104] By establishing a life model that couples centrifugal force, thermal stress, fatigue damage and high-temperature creep, the damage evolution law of materials under different operating conditions is calculated, and the life decay trend of welded joints is obtained.

[0105] Based on the traditional physical model, a Long Short-Term Memory (LSTM) neural network is introduced to dynamically correct the lifetime prediction results. The LSTM model takes multi-source time-series signals as input, learns their time variation patterns, identifies nonlinear damage trends under complex operating conditions, and realizes real-time correction of the physical model prediction results.

[0106] Furthermore, step 4 includes the following steps:

[0107] The system calculates the damage accumulation rate D(t) and derives the health index HI based on the stress and temperature data output by the model; at the same time, it uses an LSTM network to dynamically correct the damage rate, thereby achieving short-term life prediction and health trend prediction.

[0108] In the model, the input features of LSTM include time series variables, and the output is the predicted correction value of the lifetime decay rate η. Based on the real-time calculated parameters HI, ΔHI and η, the system performs multi-condition logic judgment and generates a risk level. When it detects that ΔHI is continuously decreasing or η exceeds the limit, the system immediately triggers the early warning mechanism and issues a prompt message.

[0109] The intelligent life prediction and optimization system and method for turbine rotor welded joints of the present invention has the following advantages:

[0110] A systematic intelligent life management solution for turbine rotor welded joints was proposed, which solved the problems of difficult monitoring and prediction of this key component.

[0111] By integrating multi-source real-time data with a high-fidelity digital twin model and using algorithms for dynamic correction, the accuracy and timeliness of lifespan prediction have been significantly improved.

[0112] It breaks down the barriers between monitoring and control, and achieves proactive operation optimization based on life state through reinforcement learning, realizing the transformation from passive protection to proactive protection. It can effectively extend the overhaul cycle, reduce the risk of unplanned shutdowns, and improve the operating efficiency of power plants. Attached Figure Description

[0113] Figure 1 This is a schematic diagram of the framework of the present invention;

[0114] Figure 2 This is a flowchart illustrating the present invention. Detailed Implementation

[0115] To better understand the purpose, structure, and function of this invention, the following detailed description of a smart life prediction and optimization system and method for turbine rotor welding joints, in conjunction with the accompanying drawings, is provided.

[0116] like Figure 1 As shown, the present invention provides an intelligent life prediction and optimization system for turbine rotor welding joints. The system mainly includes: a multi-source data acquisition module, a digital twin modeling module, a health status assessment and life prediction module, a risk warning module, and an operation collaborative optimization module.

[0117] The modules are interconnected via industrial Ethernet or high-speed data bus to form a complete closed-loop system of "perception-modeling-evaluation-optimization-control" to achieve real-time monitoring, intelligent evaluation and proactive optimization of rotor welding joints.

[0118] The multi-source data acquisition module described in this invention is used to collect multi-dimensional physical quantities and operating condition parameters of the turbine rotor welded joint during operation, such as multi-dimensional information on rotor welded joint stress, strain, acoustic emission, temperature and vibration, and is the basic data source of the entire system.

[0119] The module includes a high-temperature wireless strain sensing unit, an acoustic emission detection unit, a vibration monitoring unit, and a steam parameter acquisition unit, all arranged in the weld joint and its adjacent area.

[0120] The strain sensing unit is used to measure the strain signal of the rotor material in real time under rotation, heating, and loading conditions. Strain is a fundamental physical quantity that reflects the deformation of a metal structure under stress. The strain signal can reflect the elastoplastic deformation of the welded joint area under centrifugal force, thermal expansion, and structural constraints, and is an important basis for judging fatigue cycles and stress concentration changes.

[0121] The acoustic emission detection unit is used to monitor transient high-frequency acoustic signals generated by the initiation, propagation, and interaction of microcracks within metallic materials. The amplitude, energy, and count rate of acoustic emission events are closely related to the material damage process and can be used to identify early signs of failure in localized welds.

[0122] The vibration monitoring unit is located at the connection between the rotor bearing end and the housing, and is used to measure radial and axial vibration acceleration. The vibration parameters can reflect the influence of rotor system imbalance, alignment error, or crack propagation on dynamic characteristics.

[0123] The steam parameter acquisition unit includes a temperature sensor, a pressure sensor, and a flow sensor, used to acquire operating condition information such as main steam temperature, main steam pressure, and steam flow rate. Main steam temperature is a key factor affecting the thermal stress and creep rate of the rotor material, main steam pressure reflects the intensity of the operating load, and flow rate reflects the energy conversion load level.

[0124] The multi-source data acquisition module achieves stable acquisition and transmission of rotating body signals through wireless transmission or slip ring signal extraction. The wireless method has the advantages of strong anti-interference capability and flexible deployment, and is suitable for high-speed operating environments.

[0125] All sensor signals undergo synchronization, noise reduction filtering, and feature extraction via edge computing nodes. These edge computing nodes are equipped with data caching, short-term prediction, and self-diagnostic functions. When signal anomalies or interruptions are detected, interpolation repair based on historical trends can be performed to ensure data continuity.

[0126] Feature extraction includes strain cycle counts, acoustic emission event energy, vibration amplitude spectrum distribution, and steam temperature fluctuation rate. The extracted feature parameters are transmitted to the digital twin modeling module via industrial Ethernet, enabling real-time linkage between the data layer and the model layer.

[0127] The digital twin modeling module described in this invention is used to establish a high-fidelity virtual model of a rotor welding joint that integrates centrifugal force, thermal and mechanical coupling effects, so as to realize real-time synchronization, mutual correction and state inversion between the physical entity and the digital model.

[0128] This module consists of a structure mapping unit, a virtual-real synchronization unit, and a state inversion unit.

[0129] The structural mapping unit establishes a finite element simulation model including the welded joint region based on the three-dimensional CAD geometric model of the turbine rotor. The model is meshed with refined meshes to ensure the geometric accuracy of the weld and heat-affected zone. During the modeling process, material parameters are input in a temperature-dependent form, including parameters such as elastic modulus, yield strength, thermal conductivity, and specific heat capacity.

[0130] The virtual-real synchronization unit is used to apply real-time acquired physical quantities such as temperature, pressure, and rotational speed as boundary conditions to the digital twin, driving the model to perform dynamic simulation calculations and achieving synchronous evolution between the virtual and real worlds. The model's solution includes a temperature field, a stress field, and a creep rate field. The temperature field reflects the heat distribution patterns in each region. The stress field reflects the superposition effect of thermal stress and centrifugal force. The creep rate field characterizes the slow plastic deformation behavior at high temperatures.

[0131] The state inversion unit is used to compensate for missing sensor data using model results. By fitting and interpolating historical data, it adaptively adjusts material parameters based on operational data to invert the temperature or stress parameters of missing measurement points, thus achieving self-recovery of virtual measurement points.

[0132] The governing equations of the model include the thermal equilibrium equation and the stress constitutive equation.

[0133] The heat balance equation is:

[0134]

[0135] in, For material density, For specific heat capacity, For temperature field, Thermal conductivity, For internal heat source items, This is the rotating heat source term.

[0136] Stress constitutive equation:

[0137]

[0138] in, Stress tensor represents the interaction forces within a unit area of ​​the material and is a fundamental quantity reflecting the internal mechanical state of a welded joint.

[0139] It is the elastic constant tensor, used to characterize the linear response relationship between stress and elastic strain.

[0140] The total strain is used to describe the overall deformation of the welded joint under various loads.

[0141] Thermal strain is used to describe the free expansion or contraction deformation of a material caused by temperature changes.

[0142] Creep strain reflects the time-dependent plastic deformation of a material under high temperature and long-term load.

[0143] Strain generated by centrifugal force is used to describe structural deformation caused by the body force field generated by high-speed rotation.

[0144] The digital twin model receives boundary conditions from multiple sources of sensors in real time, such as main steam temperature, steam pressure and rotation speed, and drives the virtual model to calculate synchronously, outputting the stress field, temperature field and creep rate distribution of the welded joint area.

[0145] The digital twin model periodically updates the calculation results through a finite element solver and performs bidirectional correction with the measured data. When the measured temperature deviates from the model's predicted value by more than a threshold, the system automatically adjusts the material parameters or boundary conditions to ensure consistency between the virtual and actual data.

[0146] This module supports both online operation and batch processing modes. When the system is under high load, a simplified model is used for rapid calculations; when the unit is in steady-state operation, it automatically switches to a high-fidelity model for detailed solutions to balance calculation accuracy and speed.

[0147] Using the modeling methods described above, the digital twin can output the thermal stress distribution, equivalent stress over time curve, and creep strain accumulation in key areas of the welded joint in real time, providing reliable input for subsequent health assessment and life prediction.

[0148] The health status assessment and lifespan prediction module of this invention is based on the fusion of physical models and intelligent algorithms to quantify the health status and estimate the remaining lifespan of welded joints. This module has a self-learning function, which can automatically correct model parameters based on actual operating data, thereby continuously improving prediction accuracy.

[0149] This module uses a centrifugal force-thermal-fatigue-creep coupled model as its physical basis. By analyzing the comprehensive damage process of materials under high temperature, high stress, and multi-cycle conditions, it calculates the damage accumulation rate D(t) of the joint.

[0150]

[0151] in, To comprehensively consider the equivalent stress of mechanical stress, thermal stress and centrifugal stress.

[0152] Fatigue life is the number of cycles a material can withstand under a given stress amplitude and temperature.

[0153] Creep rate is the increment of creep strain per unit time under certain stress and temperature.

[0154] The creep coupling coefficient represents the degree of mutual influence between fatigue and creep damage.

[0155] By integrating and accumulating, the total damage D(t) within time t can be obtained, and its value increases monotonically with the increase of running time. When D(t) = 1, it is defined as the end of the failure life.

[0156] To quantify the health status, a health index HI is defined to characterize the remaining health status of materials; the lower the value, the more severe the structural deterioration.

[0157]

[0158] An early warning is triggered when HI is less than 0.6, and the system enters protection mode when it is less than 0.4. The early warning signal is received by the control system and used for automatic cooling, load reduction, or steam flow control.

[0159] The system calculates the health index change curve over time in real time, and further calculates the health index change rate ΔHI to describe the rate of change of the health index over time, reflecting the dynamic trend of the health status of the turbine rotor welded joint. Its calculation formula is as follows:

[0160]

[0161] in, The current health index. The health index at the previous moment. This represents the time interval between two consecutive monitoring sessions. Its sign reflects the health trend, with negative values ​​indicating deterioration and positive values ​​indicating recovery.

[0162] When ΔHI remains negative and its absolute value increases, it indicates that the damage to the welded joint has intensified; if ΔHI approaches zero or a positive value, it indicates that the operating condition has stabilized or has recovered to some extent.

[0163] The lifetime decay rate η represents the proportion of damage increase per unit time and is an important indicator reflecting the rate of lifetime depletion. It is defined as:

[0164]

[0165] The larger η is, the faster the lifespan is consumed, and the more likely the system is to fail.

[0166] This module further introduces a Long Short-Term Memory (LSTM) network model for prediction correction. The LSTM takes time series inputs such as equivalent stress, temperature, rotational speed, load, and creep rate as inputs, and by capturing their temporal correlation and long-term memory characteristics, outputs a dynamic correction coefficient for the lifetime decay rate η.

[0167] The lifetime degradation rate correction process can be expressed as:

[0168]

[0169] in, The theoretical life decay rate is calculated based on the centrifugal force-heat-fatigue-creep coupling model.

[0170] These are the dynamic correction coefficients for the output of the LSTM network, and their values ​​typically range from... .like This indicates that the physical model underestimates the damage rate (η should be increased); if This indicates that the model overestimates the damage rate (η should be reduced); if This means that the model prediction is consistent with reality and no correction is needed.

[0171] This is the corrected lifetime decay rate result.

[0172] This network can adaptively learn the nonlinear damage characteristics of rotor welded joints under complex and variable working conditions, and realize real-time correction of life prediction results and short-term trend prediction.

[0173] By comprehensively analyzing the changing trends of HI, ΔHI, and η, the system can determine the health status of the welded joint in real time, predict potential risks, and provide a basis for subsequent optimization control.

[0174] The calculation results of this module are transmitted to the risk warning module and the operation coordination optimization module in the form of a data stream, so as to achieve coordinated adjustment of safety and efficiency.

[0175] The risk warning module described in this invention is used to classify risks and provide safety warnings for turbine rotor welded joints based on the dynamic trend of health status changes. It is a key unit for achieving active protection and operational safety control.

[0176] This module uses the Health Index (HI), Health Index Change Rate (ΔHI), and Lifespan Decline Rate (η) as the main input parameters, and combines historical trends and threshold judgment rules to generate different levels of early warning.

[0177] The system is set to four risk levels: Level I (Normal), Level II (Caution), Level III (Warning), and Level IV (Danger), corresponding to green, yellow, orange, and red status indicators, respectively.

[0178] When the system is in Level I condition, the welded joints are in good health, HI≥0.8, ΔHI<0.01, and the life decay rate η is less than 5%, and the system maintains normal operation.

[0179] When the system enters Level II, if HI is between 0.6 and 0.8 or ΔHI ≥ 0.02 and the life decay rate η ≥ 15%, it indicates that the joint health has slightly declined or fatigue damage has begun to accumulate. The system issues a warning signal and prompts the operators to pay attention to load fluctuations.

[0180] When the system enters Level III, if HI is between 0.4 and 0.6 or ΔHI ≥ 0.05 and the life decay rate η ≥ 30%, it indicates that the welded joint has a significant deterioration trend. The system will then activate an audible and visual alarm and record the alarm log.

[0181] When the system enters Level IV state, if HI < 0.4 or ΔHI ≥ 0.10 and the life decay rate η ≥ 50%, it indicates a serious risk of fatigue or creep damage. The system immediately issues a load reduction or shutdown command to the unit DCS to prevent structural damage.

[0182] The calculation cycle of the risk warning module is matched with the monitoring frequency, usually once every 1 to 5 seconds, to ensure the real-time performance and accuracy of the warning response.

[0183] The module employs a multi-dimensional logic judgment algorithm to comprehensively analyze the changing trends of HI, ΔHI, and η, thus avoiding false alarms caused by fluctuations in a single parameter.

[0184] The system simultaneously saves historical warning records and parameter trajectories for later fault diagnosis and lifespan tracing analysis.

[0185] To further enhance reliability, the risk warning module also includes an adaptive threshold adjustment mechanism. During long-term operation, the system automatically calculates the distribution characteristics of HI, ΔHI, and η, and corrects the warning boundaries based on operational stability. This mechanism avoids misjudgments under different unit conditions and operating circumstances, enabling personalized and intelligent risk identification.

[0186] Risk warning signal output includes not only numerical alarms, but also textual descriptions and operational suggestions, such as "It is recommended to reduce the main steam temperature by 5°C" or "It is recommended to maintain a stable load for 30 minutes".

[0187] When the risk level reaches Level III or IV, the system will automatically trigger the event reporting module to send a risk event information package to the operation management platform, which includes the warning level, the trend of changes in key parameters, and the HI curve and ΔHI curve for the most recent 10 minutes.

[0188] This module can also interact with the plant-level SIS (Monitoring Information System) to achieve comprehensive monitoring of the status of all equipment in the plant.

[0189] Through the above design, the risk warning module of the present invention has achieved a technological leap from single threshold alarm to multi-parameter fusion intelligent early warning, which greatly improves the early identification capability of welded joint failure.

[0190] The collaborative optimization module is the core of the intelligent decision-making system of this invention. It aims to achieve optimal control of the turbine's operating efficiency by adaptively adjusting operating parameters based on reinforcement learning algorithms, while ensuring safety.

[0191] This module is built on a reinforcement learning algorithm. It takes the real-time operating status of the steam turbine as the environment and control parameters (such as steam temperature setting, load distribution, speed adjustment, etc.) as actions. It obtains the optimal operating strategy through continuous interaction.

[0192] The module's input state variables include main steam temperature, main steam pressure, flow rate, load, speed, health index HI, life decay rate η, and risk level. wait.

[0193] Output control variables include main steam temperature setpoint correction, load adjustment, and speed regulation commands.

[0194] Reinforcement learning algorithms update policies through trial and error and feedback, aiming to maximize the cumulative reward function.

[0195]

[0196] in, Unit efficiency, which can be calculated from main steam temperature, pressure, and exhaust steam enthalpy, is an indicator reflecting the unit's thermal economy.

[0197] The risk factor and risk cost parameter correspond to the risk level as follows: Level I = 0, Level II = 1, Level III = 5, Level IV = 50, used to penalize high-risk states.

[0198] ΔHI represents the rate of change of the health index, indicating the speed of health deterioration. If ΔHI is negative and its absolute value increases, it indicates that the equipment is deteriorating faster, and the algorithm will reduce the reward value to encourage the strategy to operate more safely.

[0199] The parameters α, β, and γ are weighting coefficients used to balance the relationship between operational efficiency, risk control, and health degradation.

[0200] In the initial stage of the module, offline training is conducted using historical operating data to establish a basic policy network; during actual operation, the system updates the policy online and gradually approaches the optimal solution.

[0201] The training process employs the policy gradient method, where the agent adjusts the policy distribution π(a|s) based on real-time feedback to achieve dynamic optimization of action selection.

[0202] When the risk level increases or the health status deteriorates, the algorithm will automatically reduce the load setting or reduce the main steam temperature, thereby reducing the stress level and delaying the accumulation of damage.

[0203] When the system is running stably and the risk is low, reinforcement learning algorithms tend to improve efficiency metrics. Economic optimization is achieved by fine-tuning the main steam parameters.

[0204] This module has continuous learning capabilities and can automatically adapt to new operating characteristics under different seasons and load conditions.

[0205] The module includes a safety protection constraint layer to limit the range of action output and ensure that any control adjustment will not cause parameter over-limits or operational shocks.

[0206] The system also includes a policy review mechanism, allowing operators to view the current reinforcement learning policy output and corresponding weight parameters, thus achieving human control and interpretability.

[0207] The collaborative optimization module and the risk warning module achieve closed-loop linkage: when the warning level rises, the optimization module automatically enters the safety priority mode; when the risk level falls, the system gradually restores the efficient operation mode.

[0208] Through this mechanism, the present invention achieves synergistic optimization of operating strategy and structural health status, truly realizing intelligent operation control "with lifespan constraints as the core".

[0209] In addition, the module has a learning transfer function, which can transfer the trained policy model to similar units, shorten the deployment cycle, and improve the system's versatility and economy.

[0210] The system employs a unified time synchronization protocol to ensure data consistency across modules. Each module has a built-in timestamp mechanism to guarantee that all data sampling points correspond precisely to the simulation calculation time.

[0211] like Figure 2 As shown, the intelligent life prediction and collaborative optimization method for turbine rotor welded joints of the present invention includes the following steps:

[0212] Step 1: Real-time monitoring data of the rotor welding joint is collected using a multi-source data acquisition module;

[0213] Various operational data are collected in real time by deploying high-temperature strain sensors, acoustic emission sensors, temperature sensors, vibration sensors, and steam parameter monitoring units in and around the welded joint and key areas. The acquisition cycle is typically 100ms to 1s, enabling the capture of high-frequency dynamic stress changes and acoustic emission event characteristics.

[0214] The collected data includes temperature signals, strain signals, acoustic emission event rate, vibration acceleration, steam pressure, flow rate, and rotational speed.

[0215] The temperature signal represents the change in temperature of the metal material in the welded area over time, indicating the actual temperature of the metal material at the weld joint. Temperature affects thermal stress and creep rate, and is one of the most sensitive variables in life prediction.

[0216] The strain signal indicates the degree of deformation of the local weld under rotation and thermal load; it represents the total stress borne by the welded joint, which includes mechanical stress, thermal stress and centrifugal stress.

[0217] Acoustic emission event rate represents the intensity of microcrack activity within a material per unit time;

[0218] Vibration acceleration reflects the rotor's operational balance and structural integrity;

[0219] Steam pressure, flow rate, and rotational speed reflect the operating load and heat input.

[0220] Step 2: Input the data into the digital twin model and perform virtual-real synchronous calculations to generate the temperature field, stress field, and creep rate field of the rotor welded joint;

[0221] The acquired signals undergo real-time preprocessing at edge computing nodes to ensure the stability and synchronization of the input digital twin model, including filtering and denoising, time alignment, feature extraction, and data compression.

[0222] Filtering and Noise Reduction: To eliminate electromagnetic interference, environmental vibration, and high-frequency random noise in the measurement signal, this invention sets up a multi-level filtering module in the edge computing node. First, a low-pass filter is used to suppress high-frequency pseudo-signals while retaining low-frequency components reflecting load changes. Second, a wavelet threshold filtering algorithm is used to perform multi-scale decomposition and reconstruction of transient impacts and high-frequency noise. When significant signal drift occurs, a Kalman filtering algorithm is further introduced to dynamically estimate and correct the signal based on the system dynamics equations. Through these methods, the interference of noise on stress, strain, and acoustic emission characteristics can be effectively reduced, ensuring that the input data reflects the actual working conditions.

[0223] Time Alignment: Due to the different sampling frequencies of the temperature, strain, acoustic emission, vibration, and steam parameter monitoring units, time deviations exist. This invention employs a unified timestamp mechanism to achieve multi-source signal synchronization. Each sampled data is appended with a high-precision clock identifier and aligned on a unified time axis through interpolation and resampling algorithms. For asynchronous signals that are out of sync, the system uses linear interpolation and spline function reconstruction methods to generate an equal-time-series data stream, ensuring that each physical quantity has a consistent reference benchmark at the same time. The time-aligned signals are uniformly input into the virtual-real synchronization calculation module to achieve consistency of multi-physics boundary conditions.

[0224] Feature Extraction: The preprocessed raw signals are converted into key indicators describing the physical state by the feature extraction module. For strain signals, strain amplitude, root mean square value, and cycle count results are extracted; for acoustic emission signals, event count rate, amplitude distribution, and energy density are extracted; for temperature and pressure signals, moving average, rate of change, and periodic fluctuation characteristics are extracted. These feature variables serve as input parameters for the fatigue-creep coupling model and long short-term memory network, thereby achieving health status assessment and life prediction.

[0225] Data Compression: To reduce communication bandwidth usage and storage load, this invention employs a combination of adaptive sampling and wavelet compression to compress and store preprocessed data. When the monitored signal is in a steady state, the system automatically reduces the sampling rate, recording only key change points; when sudden fluctuations or abnormal events are detected, the system instantaneously increases the sampling frequency to capture the complete waveform. The compressed data retains key features with high information density while meeting real-time requirements, providing high-quality input for subsequent digital twin model solving and training.

[0226] The preprocessed data is input into the digital twin model, and the temperature field, stress field, and creep rate field of the welded joint under the current operating condition are obtained through finite element calculation.

[0227] Step 3: Calculate the life decay rate using the centrifugal force-heat-fatigue-creep coupling model, and dynamically correct the prediction results using an LSTM neural network.

[0228] This step comprehensively considers the complex stress state of the turbine rotor welded joint under high-speed rotation, high temperature and high pressure, and cyclic loads. By establishing a life model that couples centrifugal force, thermal stress, fatigue damage, and high-temperature creep, the damage evolution law of the material under different operating conditions is calculated, and the life decay trend of the welded joint is obtained.

[0229] The centrifugal force-thermal-fatigue-creep coupled model can simultaneously reflect the thermal expansion and contraction effect caused by temperature changes, the centrifugal stress caused by rotation, the creep deformation under long-term loading, and the fatigue accumulation caused by stress cycles, thus comprehensively depicting the actual stress state of the welded joint. The life decay rate is used to measure the rate at which the material's life is consumed under current operating conditions and is a key indicator for judging the structural health status.

[0230] Based on traditional physical models, this invention introduces a Long Short-Term Memory (LSTM) neural network to dynamically correct lifetime prediction results. The LSTM model takes multi-source time-series signals such as equivalent stress, temperature, creep rate, rotational speed, steam pressure, and flow rate as input. By learning the time variation patterns of these signals, it identifies nonlinear damage trends under complex operating conditions, thereby achieving real-time correction of the physical model's prediction results.

[0231] LSTM networks can utilize their memory properties to capture historical operating information, compensating for the shortcomings of traditional physical models in accurately reflecting short-term load fluctuations and material nonlinear degradation. Their output corrects for lifetime decay rates, making predictions more closely reflect the actual operating conditions of the equipment.

[0232] By integrating the physical mechanism model with the data-driven model, the system achieves high-precision estimation of the life decay rate of welded joints, providing reliable input for subsequent health index calculation, risk warning and remaining life prediction, thereby realizing dynamic adaptation and intelligence in life prediction.

[0233] Step 4: Based on the calculated Health Index (HI) and its changing trend, and combined with preset thresholds, classify and assess risk levels.

[0234] The system calculates the damage accumulation rate D(t) and derives the health index HI based on the stress and temperature data output by the model.

[0235] Simultaneously, the damage rate is dynamically corrected using an LSTM network to achieve short-term life prediction and health trend prediction.

[0236] In the model, the input features of LSTM include time series variables such as equivalent stress, temperature, creep strain rate, rotational speed, steam pressure, and steam flow rate, and the output is the predicted correction value of the lifetime decay rate η, which further improves the accuracy of the assessment.

[0237] Based on the real-time calculated parameters HI, ΔHI, and η, the system performs multi-condition logical judgments and generates a risk level. When a continuous decrease in ΔHI or an excess of η is detected, the system immediately triggers an early warning mechanism and issues a prompt message.

[0238] Step 5: Run the collaborative optimization module to execute the reinforcement learning optimization strategy, dynamically adjust the operating parameters, and realize the integrated collaborative control of rotor welding joint life and unit operating efficiency.

[0239] The reinforcement learning module automatically adjusts operating parameters based on risk level and operational efficiency feedback to achieve optimal energy efficiency under lifespan constraints.

[0240] For example, when When the risk level decreases and the risk level is low, the system will automatically increase the main steam temperature setpoint to improve economy; when ΔHI rises sharply, the system will automatically reduce the load to protect the welded joint.

[0241] The system archives all monitoring data, calculation results, and control commands from all operating cycles for model retraining and algorithm optimization.

[0242] After long-term operation, the system will form a knowledge base based on multi-condition experience, supporting policy migration and intelligent management of unit groups.

[0243] Compared to traditional steam turbine life management systems, this invention has the following significant advantages:

[0244] Real-time capture of microscopic damage in welded joints is achieved through high-temperature wireless sensing and multi-source data fusion.

[0245] Virtual-real synchronous simulation based on digital twins can accurately reproduce the multi-field coupling effect of thermo-mechanical-centrifugal forces;

[0246] By introducing LSTM networks and reinforcement learning algorithms, it possesses self-learning and self-correction capabilities;

[0247] By using multi-parameter fusion for judgment, false alarms caused by a single threshold are avoided;

[0248] To achieve a shift from passive maintenance to proactive lifecycle management;

[0249] The system has a modular structure that can adapt to different unit sizes and plant-level platforms.

[0250] This invention, through intelligent data analysis and automated control strategies, can respond to problems in steam turbine units more quickly and accurately, reducing the need for human intervention and improving system efficiency and stability.

[0251] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A steam turbine rotor welded joint intelligent life prediction and optimization system, characterized in that, Comprise: Multi-source data acquisition module, digital twin modeling module, health state evaluation and life prediction module, risk early warning module and operation collaborative optimization module; The multi-source data acquisition module is used for collecting multi-dimensional physical quantities and operating condition parameters of the turbine rotor welded joint in the running process; The digital twin modeling module is used for establishing a high-fidelity virtual model of the rotor welded joint to realize real-time synchronization, mutual correction and state inversion of the physical entity and the digital model; The health state evaluation and life prediction module realizes the quantification of the health state and the estimation of the remaining life of the welded joint based on the fusion of physical models and intelligent algorithms; The risk early warning module is used for risk grading and safety warning of the turbine rotor welded joint based on the dynamic change trend of the health state; The running collaborative optimization module adaptively adjusts the operating parameters based on the reinforcement learning algorithm to realize the optimal control of the turbine operation efficiency under the premise of ensuring safety; Each module is interconnected through industrial Ethernet or high-speed data bus to form a complete "perception-modeling-evaluation-optimization-control" closed loop system to realize real-time monitoring, intelligent evaluation and active optimization of the rotor welded joint.

2. The steam turbine rotor welded joint intelligent life prediction and optimization system of claim 1, wherein, The multi-source data acquisition module comprises a high-temperature wireless strain sensing unit, an acoustic emission detection unit, a vibration monitoring unit and a steam parameter acquisition unit arranged at the welded joint and its adjacent area; The strain sensing unit is used for real-time measurement of strain signals of the rotor material in the rotating, heated and loaded state; The acoustic emission detection unit is used for monitoring the transient high-frequency acoustic signals generated by the initiation, expansion and interaction of micro-cracks in the metal material; The vibration monitoring unit is arranged at the connection between the rotor bearing end and the shell for measuring radial and axial vibration acceleration; The steam parameter acquisition unit comprises a temperature sensor, a pressure sensor and a flow sensor for acquiring operating condition information such as main steam temperature, main steam pressure and steam flow; The multi-source data acquisition module realizes stable acquisition and transmission of the rotating body signal through wireless transmission or slip ring signal lead-out mode; all sensor signals are synchronized and time-stamped, denoised and filtered and feature extracted by the edge computing node, the feature extraction content includes strain cycle count, acoustic emission event energy, vibration amplitude spectrum distribution and steam temperature fluctuation rate, and the extracted feature parameters are transmitted to the digital twin modeling module through industrial Ethernet to realize real-time linkage of the data layer and the model layer.

3. The intelligent life prediction and optimization system for steam turbine rotor welded joints of claim 1, wherein, The digital twin modeling module comprises a structure mapping unit, a virtual-real synchronization unit and a state inversion unit; The structure mapping unit establishes a finite element simulation model containing the welded joint area based on the three-dimensional CAD geometric model of the turbine rotor, the model is divided into refined grids to ensure the geometric accuracy of the weld and the heat affected zone, and the material parameters are input in a temperature-dependent form during modeling; The virtual-real synchronization unit is used for applying the real-time collected physical quantities as boundary conditions to the digital twin body to drive the model to perform dynamic simulation calculation, realize virtual-real synchronous evolution, and the solution of the model includes temperature field, stress field and creep rate field; The state inversion unit is used to compensate for missing sensor data by using model results, to calculate missing temperature or stress parameters of the virtual measuring points by historical data fitting and interpolation, and to adaptively adjust material parameters according to operation data, so as to realize self-recovery of the virtual measuring points; The control equation of the model includes a heat balance equation and a stress constitutive equation: The heat balance equation is: ， wherein, is the density of the material, is the specific heat capacity, is the temperature field, is the thermal conductivity, is the internal heat source term, is the rotational heat source term; The stress constitutive equation is: ， wherein, is the stress tensor, representing the interaction force within the material per unit area, and is a basic quantity reflecting the internal mechanical state of the welded joint; Cijklis the elastic constant tensor, which characterizes the linear response relation between stress and elastic strain; Total strain is used to describe the overall deformation of the welded joint under various loads; Thermal strain is used to describe the free expansion or contraction deformation of a material caused by a change in temperature. For creep strain, reflecting the time-dependent plastic deformation of the material under the action of high temperature and long-term load; Centrifugal strain. Strain due to the body force field resulting from high speed rotation. The digital twin model periodically updates the calculation results through a finite element solver, and is bidirectionally corrected with the measured data; when the measured temperature deviates from the model prediction value by more than a threshold value, the system automatically adjusts the material parameters or the boundary conditions to ensure consistency between the virtual and real. The digital twin model supports online operation and batch processing modes; when the system is in a high-load operation stage, a simplified model is preferentially used for rapid calculation; when the unit is in a steady-state working condition, the system is automatically switched to a high-fidelity model for detailed solution, so as to balance the calculation accuracy and speed.

4. The intelligent life prediction and optimization system for steam turbine rotor welded joints of claim 1, wherein, The health state evaluation and life prediction module uses a centrifugal force-heat-fatigue-creep coupling model as a physical basis, analyzes the comprehensive damage process of the material under high temperature, high stress and multi-cycle working conditions, and calculates the damage accumulation rate D(t) of the joint: ， wherein, is the equivalent stress considering the mechanical stress, thermal stress and centrifugal stress. Fatigue life, i.e. the number of cycles a material can withstand at a given stress amplitude and temperature; The creep rate is the increment of creep strain per unit time at a certain stress and temperature; C is the creep coupling coefficient, which represents the degree of interaction between fatigue and creep damage; Through integral accumulation, the total damage amount D(t) within time t is obtained, which monotonously increases with the increase of the operation time; when D(t)=1, the failure life end is defined; A health index HI is defined to represent the remaining health state of the material, and the lower the value, the more serious the structure degradation, ， When HI is less than 0.6, a warning is triggered, and when it is less than 0.4, the system enters a protection mode. The warning signal is received by the control system and is used for automatic cooling, load reduction or control of steam flow; The system calculates the change curve of the health index with time in real time, and further calculates the health index change rate ΔHI for describing the change rate of the health index with time, reflecting the dynamic change trend of the health state of the turbine rotor welded joint, and the calculation formula is: ， wherein, Htis the health index at the current time instant, Ht-1is the health index at the previous time instant, is the time interval between two consecutive monitoring, whose sign reflects the health trend, negative value indicates deterioration, positive value indicates recovery; When ΔHI is continuously negative and the absolute value increases, it indicates that the damage of the welded joint is aggravated; if ΔHI tends to zero or a positive value, it indicates that the operation state is stable or has recovered; The life attenuation rate η represents the proportion of damage growth per unit time, and is an important index reflecting the life consumption rate, and is defined as: ， The larger η is, the faster the life consumption is, and the more the system tends to fail; A long short-term memory network (LSTM) model is introduced for prediction correction. The LSTM inputs the feature variables in the form of time series, captures the time correlation and long-term memory characteristics, and outputs the dynamic correction coefficient of the life attenuation rate η, The life attenuation rate correction process is represented as: ， wherein, is the theoretical life decay rate calculated based on the centrifugal force-heat-fatigue-creep coupling model; is the dynamic correction factor output by the LSTM network, if it means that the physical model underestimates the damage rate; if it means that the model overestimates the damage rate; if it means that the model prediction is consistent with the actual, no correction is needed; The corrected lifetime decay rate result.

5. The intelligent life prediction and optimization system for steam turbine rotor welded joints of claim 1, wherein, The risk warning module takes the health index, the health index change rate and the life attenuation rate as input parameters, combines historical trends and threshold judgment rules to generate different levels of warning levels; The system sets four levels of risk levels: level I, level II, level III and level IV, corresponding to green, yellow, orange and red state indications respectively; When the system is in level I, the welded joint is in good health, HI≥0.8, ΔHI<0.01, and the life attenuation rate η is less than 5%, and the system maintains normal operation. When the system enters the II level state, HI is between 0.6 and 0.8 or ΔHI≥0.02, and the life attenuation rate η≥15%, indicating that the joint health is slightly reduced or the fatigue damage begins to accumulate, the system sends an attention signal and prompts the operator to pay attention to the load fluctuation; When the system enters the III level state, HI is between 0.4 and 0.6 or ΔHI≥0.05, and the life attenuation rate η≥30%, indicating that the welded joint appears a significant deterioration trend, the system starts the audible and visual alarm and records the alarm log; When the system enters the IV level state, HI<0.4 or ΔHI≥0.10, and the life attenuation rate η≥50%, indicating that there is a serious fatigue or creep damage risk, the system immediately sends a load reduction or shutdown instruction to the unit DCS to prevent structural damage; The calculation period of the risk warning module matches the monitoring frequency; the multi-dimensional logic decision algorithm is used in the module to comprehensively analyze the change trend of HI, ΔHI and η, and to avoid false alarms caused by single parameter fluctuation; The risk warning module also includes an adaptive threshold adjustment mechanism, which automatically statistics the distribution characteristics of HI, ΔHI and η during long-term operation, and corrects the warning boundary based on the operation stability; When the risk level reaches III or IV, the system will automatically trigger the event reporting module to send risk event information package to the operation management platform; The risk warning module also interacts with the monitoring information system to realize comprehensive monitoring of the state of all plant equipment.

6. The steam turbine rotor welded joint intelligent life prediction and optimization system of claim 1, wherein, The operation collaborative optimization module is constructed based on the reinforcement learning algorithm, takes the real-time operation state of the steam turbine as the environment, and takes the control parameters as the action, to obtain the optimal operation strategy through continuous interaction; The input state variables of the module include main steam temperature, main steam pressure, flow, load, rotating speed, health index HI, life attenuation rate η and risk level ; The output control variables include the main steam temperature setting correction amount, the load adjustment amount and the speed regulation instruction; The reinforcement learning algorithm updates the strategy through exploration and feedback, aiming to maximize the cumulative reward function: ， wherein, is the unit efficiency, which is calculated by the main steam temperature, pressure and exhaust enthalpy, and is an index reflecting the thermal economy of the unit; The risk factor is the risk cost parameter, and the risk level corresponds as follows: I level = 0, II level = 1, III level = 5, IV level = 50, which is used to punish high-risk state; ΔHI represents the health index change rate, representing the health deterioration speed, if ΔHI is negative and the absolute value increases, indicating that the equipment deterioration is accelerated, the algorithm will reduce the reward value to promote the strategy to safe operation; The parameters α, β and γ are weight coefficients, used to balance the relationship among operation efficiency, risk control and health attenuation; In the initial stage, the module is trained offline through historical operation data to establish a basic strategy network; in the actual operation process, the system updates the strategy online to gradually approach the optimal solution; The training process uses the policy gradient method, and the agent adjusts the strategy distribution π(a|s) according to the real-time feedback to realize dynamic optimization of action selection; When the risk level rises or the health state deteriorates, the algorithm automatically reduces the load setting or the main steam temperature, thereby reducing the stress level and delaying damage accumulation; When the system is in stable operation and the risk is low, the reinforcement learning algorithm tends to improve the efficiency index , and achieve economic optimization by fine-tuning the main steam parameters; The operation collaborative optimization module has continuous learning ability and can automatically adapt to new operation characteristics under different seasons and different load conditions; A safety protection constraint layer is provided in the module to limit the action output range, ensuring that any control adjustment will not cause parameter overrun or operation impact; The system also includes a strategy review mechanism, which allows the operator to view the current reinforcement learning strategy output and the corresponding weight parameters to realize human controllable and interpretable; The operation coordination optimization module and the risk early warning module realize closed-loop linkage: when the early warning level rises, the optimization module automatically enters the safety priority mode; when the risk level decreases, the system gradually recovers the high-efficiency operation mode; The operation coordination optimization module has a learning migration function, can migrate the trained strategy model to similar units, shorten the deployment cycle, and improve the universality and economy of the system; The system uses a unified time synchronization protocol to ensure the consistency of data between modules, and each module is internally provided with a time stamp mechanism to ensure that all data sampling points and simulation calculation times correspond accurately.

7. A method for implementing the intelligent life prediction and optimization system for steam turbine rotor welded joints as claimed in any one of claims 1-6, characterized in that, Comprise the following steps: Step 1: Real-time acquisition of monitoring data of the rotor welded joint through the multi-source data acquisition module; Step 2: Input the data into the digital twin model for virtual-real synchronous calculation to generate the temperature field, stress field and creep rate field of the rotor welded joint; Step 3: Calculate the life decay rate using the centrifugal force-thermal-fatigue-creep coupling model, and dynamically correct the prediction results using the LSTM neural network; Step 4: According to the calculated health index HI and its trend, combined with the preset threshold, the risk is judged by classification; Step 5: The operation coordination optimization module executes the reinforcement learning optimization strategy to dynamically adjust the operation parameters, realizing the integrated collaborative control of the rotor welded joint life and the unit operation efficiency.

8. The method of claim 7, wherein, The step 2 comprises the following steps: The collected signals are preprocessed in real time by the edge computing node, including filtering and denoising, time alignment, feature extraction and data compression; Filtering and denoising: A multi-stage filtering module is set in the edge computing node, first, a low-pass filter is used to suppress high-frequency false signals and retain low-frequency components reflecting load changes; second, a wavelet threshold filtering algorithm is used for multi-scale decomposition and reconstruction of transient impact and high-frequency noise; when the signal has obvious drift, further introduce Kalman filtering algorithm to dynamically estimate and correct the signal according to the system dynamics equation; Time alignment: A unified timestamp mechanism is used to realize the synchronization of multi-source signals, each sampling data is attached with a high-precision clock identifier, and interpolation and resampling algorithms are used to align on the unified time axis. For asynchronous signals, the system uses linear interpolation and spline function reconstruction method to generate isochronous data stream, ensuring that each physical quantity has a consistent reference benchmark at the same time. The time-aligned signals are uniformly input into the virtual-real synchronous calculation module to realize the consistency of the boundary conditions of multi-physical fields; Feature extraction: For strain signals, strain amplitude, root mean square value and cycle count results are extracted; for acoustic emission signals, event count rate, amplitude distribution and energy density are extracted; for temperature and pressure signals, moving average, change rate and periodic fluctuation characteristics are extracted. These characteristic variables are used as input parameters for fatigue-creep coupling model and long short-term memory network, so as to realize health state assessment and life prediction; Data compression: Adaptive sampling and wavelet compression are combined to compress and store the preprocessed data. When the monitoring signal is in a steady state, the system automatically reduces the sampling rate and only records the key change points. When a sudden fluctuation or abnormal event is detected, the system instantaneously increases the sampling frequency to capture the complete waveform. ​ The preprocessed data is input into the digital twin model, and the temperature field, stress field and creep rate field of the welded joint under the current operating state are obtained through finite element calculation.

9. The method of claim 7, wherein, The step 3 comprises the following steps: By establishing a life model coupling centrifugal force, thermal stress, fatigue damage and high-temperature creep, the damage evolution law of the material under different operating conditions is calculated, and the life attenuation trend of the welded joint is obtained; On the basis of the traditional physical model, a long short-term memory neural network LSTM is introduced to dynamically correct the life prediction results, the LSTM model takes multi-source time series signals as input, learns the time variation law, identifies the nonlinear damage trend under complex operating conditions, and realizes real-time correction of the prediction results of the physical model.

10. The method of claim 7, wherein, The step 4 comprises the following steps: According to the stress and temperature data output by the model, the system calculates the damage accumulation rate D(t) and obtains the health index HI; meanwhile, the LSTM network is used to dynamically correct the damage rate, realizing short-term life prediction and health trend prediction; In the model, the input features of LSTM include time series variables, and the output is the predicted correction value of the life attenuation rate η; according to the real-time calculated HI, ΔHI and η parameters, the system performs multi-condition logical judgment and generates a risk level, and when ΔHI continuously decreases or η exceeds the limit, the system immediately triggers the early warning mechanism and sends a prompt message.

Images