Method for predicting life migration of hot end fitting of physically constrained network

Abstract

The present application relates to the technical field of combustion engine monitoring, and particularly relates to a method for life transfer prediction of a hot-end accessory of a physically constrained network, which first acquires a full-life operation time sequence of a source unit, and extracts a peak-regulation working condition sample containing start-stop cycle time stamps; secondly, temperature loads and mechanical stress monitoring information corresponding to the time are collected, and are combined with thermal-mechanical fatigue damage parameters and creep physical constitutive model parameters of materials; finally, the physically constrained network is used to fuse the above-mentioned multi-source information and mechanism parameters, to perform life transfer prediction on a target cross-unit first-stage rotating blade, and to generate a residual life output result. The present application solves the problem of poor generalization of a pure data-driven model by introducing a physical constitutive constraint, uses rich data of a source unit to make up for the shortcoming of early data scarcity of a target unit, realizes accurate evaluation of thermal-mechanical coupling damage under complex peak-regulation working conditions, and significantly improves the accuracy and reliability of cross-unit life prediction.

Description

Technical Field

[0001] This invention relates to the field of gas turbine monitoring technology, and in particular to a method for predicting the life migration of hot-end components in a physical constraint network. Background Technology

[0002] With the increasing demand for flexibility and peak-shaving capabilities in power systems, gas turbines frequently participate in start-stop cycles and variable load operations. This results in key hot-end components being subjected to complex thermo-mechanical coupled loads, significantly accelerating the accumulation of fatigue and creep damage in materials. Among them, the first-stage rotor blade, as one of the components in the gas turbine operating in the harshest environment, is subjected to extreme conditions of high temperature, high pressure, and high speed for extended periods. Its lifespan is affected by the combined effects of thermal stress fluctuations, uneven temperature gradients, and cyclic mechanical loads, making it highly susceptible to thermo-mechanical fatigue and creep damage, which seriously affects the safety and reliability of unit operation.

[0003] Traditional life prediction methods are mostly based on empirical models under constant operating conditions or purely data-driven black-box algorithms, which struggle to accurately characterize the impact of transient thermal shocks and nonlinear material responses on damage evolution during peak-shaving operations. This is especially true when applying these methods across different units, where differences in operating history, control strategies, and boundary conditions lead to a significant drop in prediction accuracy. Furthermore, target units often lack sufficient fault and degradation data in the early stages of service, limiting the effective training and generalization capabilities of data-driven models. While physically constrained neural networks have shown potential in integrating mechanistic models with data learning in recent years, they still have shortcomings in multi-source operating condition migration, damage mechanism separation, and interpretability in practical engineering.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method for predicting the life migration of hot-end components in a physical constraint network. This method aims to solve the technical problems of low prediction accuracy and poor generalization ability of the existing technology when facing the peak-shaving operation of gas turbines across units. This is due to the difficulty in integrating the material physical constitutive structure and transient thermo-mechanical coupling mechanism, and the scarcity of degradation data of the target unit.

[0006] To achieve the above objectives, the present invention provides a method for predicting the lifetime migration of hot-end components in a physically constrained network, the method comprising: The full life cycle operation sequence of the target gas turbine power unit is obtained. The full life cycle operation sequence includes multiple peak-shaving operating condition samples, and each operating sample has a corresponding start-stop cycle time stamp. The original monitoring information of temperature load and mechanical stress of the first-stage rotor blade of the target gas turbine at the time stamp of each start-stop cycle is collected, as well as the thermomechanical fatigue coupling damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade. Based on the original monitoring information and the damage parameters and physical constitutive model parameters corresponding to the start-stop cycle timestamps of each cycle, the life migration prediction of the target cross-unit first-stage rotor blade is performed, and the remaining life output result of the thermomechanical fatigue coupling damage of the first-stage rotor blade is generated.

[0007] Optionally, the step of obtaining the full lifecycle operating sequence of the target gas turbine power unit includes multiple peak-shaving operating condition samples, each of which has a corresponding start-stop cycle timestamp, including: Obtain the unit identifier to which the target first-stage rotor blade belongs, and retrieve the corresponding full-life operation database and material performance parameter database based on the unit identifier; Receive the peak load range and prediction assessment time range set by the user, wherein the prediction assessment time range includes the start time of operation and the end time of prediction; Based on the peak load range, the start time, and the termination prediction time, multiple peak load operation samples are selected from the full life cycle operation database, and the start and stop cycle counts of each operation sample are extracted simultaneously as start and stop cycle timestamps for each operation sample.

[0008] Optionally, the acquisition of the original monitoring information of temperature load and mechanical stress corresponding to the time stamps of each start-stop cycle of the target gas turbine first-stage rotor blade, as well as the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade, includes: The design geometric parameters and material crystal orientation type of the first-stage rotor blade of the target gas turbine are obtained. The design geometric parameters include the blade root coordinates and the distribution density of film gas holes. The damage neighborhood calculation window range is determined based on the air film pore distribution density and the material crystal orientation type. Based on the blade root coordinates, the original monitoring information of temperature load and mechanical stress corresponding to each start-stop cycle time stamp is obtained; The coupled damage calculation region is determined based on the blade design geometric parameters and the damage neighborhood calculation window range, and the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters within the calculation region are obtained.

[0009] Optionally, determining the damage neighborhood calculation window range based on the gas film pore distribution density and the material crystal orientation type includes: The equivalent spacing of the air film pores is calculated based on the air film pore distribution density, and the first window radius is obtained from the damage window mapping table based on the equivalent spacing. The second window radius is determined based on the crystal orientation type of the material, and the second window radius is a preset multiple of the influence range of material anisotropy; Compare the first window radius with the second window radius, and take the larger value as the base window radius; The damage neighborhood calculation window range is obtained by adding a fatigue damage margin to the basic window radius.

[0010] Optionally, the step of predicting the life migration of the target cross-unit first-stage rotor blade based on the original monitoring information corresponding to each start-stop cycle time stamp, the damage parameters, and the physical constitutive model parameters, and generating the remaining life output result of the thermomechanical fatigue coupled damage of the first-stage rotor blade, includes: Based on the thermomechanical fatigue coupled damage parameters, creep physical constitutive model parameters and the original monitoring information corresponding to each start-stop cycle time stamp, the actual cumulative damage component corresponding to each start-stop cycle time stamp is calculated. The initial prediction results of the migration model are corrected based on the actual cumulative damage components corresponding to each start-stop cycle time stamp, and the corrected remaining life prediction values ​​corresponding to each start-stop cycle time stamp are obtained, generating the remaining life output results of the thermomechanical fatigue coupling damage of the first-stage blade.

[0011] Optionally, the step of correcting the initial prediction results of the transfer model based on the actual cumulative damage component corresponding to each start-stop cycle time stamp to obtain the corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp includes: Traverse multiple peak-shaving operation samples in the entire lifecycle operation sequence to obtain the first set of peak-shaving operation samples; Extract the first start-stop cycle timestamp of the first group of peak-shaving operation samples, and obtain the first actual damage accumulation component based on the first start-stop cycle timestamp; Based on the first actual damage accumulation component, the first group of peak-shaving operation samples are divided into creep-dominated damage region and thermal fatigue-dominated damage region. For the first group of peak-shaving operation samples, the first damage component is extracted in the creep-dominated damage region, and the second damage component is extracted in the thermal fatigue-dominated damage region. The first damage component and the second damage component are coupled and synthesized to generate the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp. The corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp is obtained by using the same method as obtaining the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp.

[0012] Optionally, the step of extracting a first damage component in the creep-dominated damage region and a second damage component in the thermal fatigue-dominated damage region for the first group of peak-shaving operation samples includes: Obtain the coupling damage separation threshold; In the creep-dominated damage region, the first group of peak-shaving operation samples are subjected to damage component separation processing based on the coupled damage separation threshold, and the cumulative damage introduced by Larson-Miller creep is separated and extracted to obtain the first damage component. In the thermal fatigue-dominated damage region, after creep coupling correction is performed on the first group of peak-shaving operation samples, damage component separation processing is performed in combination with the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

[0013] Optionally, in the thermal fatigue-dominated damage region, after performing creep coupling correction on the first group of peak-shaving operation samples, damage component separation processing is performed in conjunction with the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component, including: In the first group of peak-shaving operation samples, the original monitoring values ​​and creep coupling components of each sampling point in the thermal fatigue-dominated damage area are obtained; The original monitoring value is replaced by the creep coupling correction value after creep coupling correction is obtained by performing creep coupling correction on each sampling point in the thermal fatigue-dominated damage area. Based on the corrected damage value and the coupled damage separation threshold, the first group of peak-shaving operation samples are subjected to damage component separation processing, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

[0014] Optionally, the construction steps of the creep coupling corrector include: Collect a sample creep coupling record set and a sample original damage set, and annotate based on the sample creep coupling record set and the sample original damage set to obtain a sample corrected damage set; Using the sample creep coupling record set and the sample original damage set as inputs, and the sample corrected damage set as a supervision signal, the creep coupling corrector is trained and generated by combining the pre-trained weights of the physical constraint network for transfer fine-tuning.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a hot-end component lifetime migration prediction program for a physical constraint network, wherein when the hot-end component lifetime migration prediction program for a physical constraint network is executed by a processor, the program implements the steps of the hot-end component lifetime migration prediction method for a physical constraint network as described above.

[0016] This invention provides a method for predicting the life migration of hot-end components using a physical constraint network. The method integrates the full-life-cycle operating sequence data of the target gas turbine unit with the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the first-stage rotor material to construct a life migration prediction framework based on a physical constraint network. This effectively overcomes the problem of scarce degradation data across target units in the early stages of service, enabling accurate modeling of damage evolution of hot-end components under start-stop cycles and peak-shaving conditions. By introducing start-stop cycle time stamps to align key loads and damage parameters temporally, the method enhances the response capability to the coupling effect of transient thermal stress and mechanical loads, improving the accuracy and generalization ability of predicting the remaining life of the first-stage rotor under different operating boundary conditions. This provides highly reliable technical support for the condition assessment, maintenance decision-making, and life management of key gas turbine components. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment of the hot-end component lifetime migration prediction method of the physical constraint network of the present invention. Figure 2 This is a flowchart illustrating another embodiment of the hot-end component lifetime migration prediction method of the physical constraint network of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the hot-end component lifetime migration prediction method for physically constrained networks according to the present invention. An embodiment of the hot-end component lifetime migration prediction method for physically constrained networks according to the present invention is presented.

[0023] In one embodiment, the method for predicting the lifetime migration of hot-end components in the physical constraint network includes: Step S100: Obtain the full life cycle operating sequence of the target gas turbine power unit. The full life cycle operating sequence includes multiple peak-shaving operating sample samples, and each operating sample has a corresponding start-stop cycle time stamp.

[0024] The target gas turbine source unit can be a gas turbine entity with a complete service history, serving as a training data source for the life prediction model. It can provide a benchmark dataset containing rich failure modes and degradation trajectories to support transfer learning. In this embodiment, the target gas turbine source unit can access the historical operating database of a specific unit through the power plant's distributed control system or remote diagnostic platform. The full life cycle operating sequence can be a continuous time series data set covering all start-up, shutdown, and load change processes of the gas turbine from commissioning to the current moment. It can be used to fully characterize the cumulative damage evolution process of components under long-term complex operating conditions. Furthermore, the full life cycle operating sequence can clean, resample, and stitch together the sensor sampling data in the distributed control system. In an exemplary embodiment, the full life cycle operating sequence can use thermal parameter time series, vibration monitoring time series, performance index time series, etc. The peak-shaving operating condition sample can be a data segment from the full life cycle operating sequence reflecting the non-steady-state operating characteristics of the unit, such as frequent load changes and rapid start-up and shutdown. It can be used to specifically characterize the transient thermal shock loads borne by hot-end components under flexible operating modes. In one specific embodiment, peak-shaving operation samples can be automatically segmented and extracted from the entire lifecycle time series based on load change rate thresholds or scheduling instruction tags. Furthermore, peak-shaving operation samples may include, but are not limited to, cold-start samples, warm-start samples, and samples of rapid load increases and decreases.

[0025] Start-stop cycle timestamps can serve as time index identifiers marking key time nodes throughout the entire process of a gas turbine's single start-up, grid connection, and shutdown. They can be used as benchmark anchors for aligning the timing of multi-source heterogeneous data to ensure causal consistency in transient processes. In this embodiment, the start-stop cycle timestamps can be automatically generated based on the unit's sequential control logic signals or speed / load state machine jump points. For example, the start-stop cycle timestamps can be coordinated with the original monitoring information of temperature load and mechanical stress to map discrete physical parameters to a unified cyclic phase coordinate system. Furthermore, start-stop cycle timestamps can include, but are not limited to, ignition timestamps, start-up timestamps, and full-load timestamps. Obtaining the full lifecycle operating sequence of the target gas turbine source unit can be achieved by extracting all relevant sensor timing records of the specified unit since its commissioning from a historical database. Furthermore, obtaining the full lifecycle operating sequence of the target gas turbine source unit can be achieved by subscribing to and caching historical data in real time via the OPCUA interface, and by batch importing CSV format data from offline backup tapes, thereby constructing a complete source domain knowledge base covering multiple degradation modes.

[0026] Step S200: Collect the original monitoring information of temperature load and mechanical stress corresponding to the time stamps of each start-stop cycle of the target gas turbine first-stage rotor blade, as well as the thermomechanical fatigue coupling damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade.

[0027] The first-stage rotor blade can be the rotating blade component of the first stage of a gas turbine that bears the highest gas temperature and centrifugal load, and can be used as a specific application object and damage assessment carrier for life migration prediction methods. In this embodiment, the first-stage rotor blade can include, but is not limited to, film-cooled blades, impact-cooled blades, and composite-cooled blades. The temperature load can be a thermal input quantity characterizing the temperature field distribution on the surface and inside of the first-stage rotor blade and its gradient over time, and can be used to quantify the thermodynamic boundary conditions driving the generation of thermal stress and the accumulation of creep damage. For example, the temperature load can be reconstructed from the three-dimensional temperature field of the blade by combining thermocouple measurements with a heat transfer inversion algorithm. Further, the temperature load can include, but is not limited to, blade root temperature, blade mid-section temperature, and blade tip temperature. The original mechanical stress monitoring information can be sensor data that directly reflects the mechanical response state of the first-stage rotor blade under rotation and aerodynamic loads, and can be used to provide the basic data of the true stress amplitude and mean value required to calculate low-cycle fatigue damage. In an exemplary embodiment, the original mechanical stress monitoring information can be collected and transmitted in real time using strain gauges, fiber optic grating sensors, or blade tip timing systems. In one specific embodiment, the raw monitoring information of mechanical stress can be obtained from resistance strain monitoring data, fiber Bragg grating monitoring data, piezoelectric sensing monitoring data, etc.

[0028] Thermomechanical fatigue coupled damage parameters can be physical characteristic quantities describing the damage evolution of materials under the combined effects of high-temperature alternating stress and temperature cycling. They can be used to provide prior knowledge that conforms to the material failure mechanism for neural networks to constrain the model output range. In this embodiment, the thermomechanical fatigue coupled damage parameters can be obtained based on fitting from standard specimen bench tests and include variables such as strain amplitude, holding time, and temperature phase difference. For example, the thermomechanical fatigue coupled damage parameters can be coordinated with creep physical constitutive model parameters to jointly constitute the regularization term in the loss function of the physical constraint network. Further, the thermomechanical fatigue coupled damage parameters can include, but are not limited to, oxidation-induced damage parameters, plastic strain accumulation parameters, and load-holding crack propagation parameters. Creep physical constitutive model parameters can be mathematical model coefficients characterizing the time-dependent plastic deformation behavior of materials under constant high-temperature stress. They can be used to embed into a prediction framework to accurately describe the steady-state and accelerated creep stages during long-term high-temperature service. In an exemplary embodiment, the creep physical constitutive model parameters can be determined by Norton-Bailey or Theta projection method parameters through endurance strength tests and creep curve regression analysis. In one specific embodiment, the parameters of the creep physical constitutive model may include, but are not limited to, Norton power-law parameters, Kachanov-Rabotnov damage parameters, Unified viscoplasticity parameters, etc.

[0029] Collecting raw monitoring information on temperature load and mechanical stress corresponding to each start-stop cycle timepoint of the first-stage rotor blade can be achieved by using the start-stop cycle timepoint as an index window to extract thermal and mechanical sensing data within the corresponding time period. Furthermore, collecting this raw monitoring information can be achieved by using a sliding time window to match trigger signals for data slicing and by automatically recording data before and after key nodes based on an event-driven mechanism, thus enabling precise synchronization and correlation of multi-physics data under transient conditions. Obtaining the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade can be achieved by querying a material performance database or loading a parameter set from a pre-stored constitutive model configuration file. Further, obtaining these parameters can be achieved by reading JSON-formatted material parameter cards or calling a cloud-based materials genome engineering database via an API interface, thus providing non-learnable deterministic physical priors for the physical constraint network.

[0030] Step S300: Based on the original monitoring information, damage parameters, and physical constitutive model parameters corresponding to the start-stop cycle timestamps, perform life migration prediction on the target cross-unit first-stage rotor blades and generate the output result of the remaining life of the first-stage rotor blades due to thermomechanical fatigue coupling damage.

[0031] The physical constraint network can be a neural network architecture that embeds the material physics constitutive equations and damage evolution laws as hard constraints or soft penalties into the loss function. It can be used to ensure that the prediction results conform to physical laws and improve generalization ability under small sample conditions of the target unit. In this embodiment, the physical constraint network can simultaneously minimize data prediction error and physical equation residuals during backpropagation. For example, the physical constraint network can be coordinated with start-stop cycle time stamps to ensure that the transient response characteristics learned by the network are consistent with the time scale of the actual physical process. Furthermore, the physical constraint network can include, but is not limited to, loss function weighted, network structure embedded, and Lagrange multiplier types. The remaining life output of the first-stage rotor's thermomechanical fatigue coupled damage can be an estimate of the safe service life or cycle count of the first-stage rotor after comprehensively considering the interaction between fatigue and creep. This can be used to provide quantitative basis for gas turbine maintenance planning and spare parts replacement decisions. In an exemplary embodiment, the remaining life output of the first-stage rotor's thermomechanical fatigue coupled damage can include, but is not limited to, the remaining cycle count, remaining operating hours, and confidence interval life values.

[0032] Time-series alignment of original monitoring information with damage parameters and physical constitutive model parameters based on start-stop cycle time stamps can be achieved by interpolating and resampling monitoring data at different sampling rates to a unified time axis and aligning it with broadcast static physical parameters. Furthermore, time-series alignment of original monitoring information with damage parameters and physical constitutive model parameters based on start-stop cycle time stamps can be achieved by using linear interpolation to align dynamic data with piecewise constant physical parameters and employing a dynamic time warping algorithm to correct phase deviations, thereby eliminating pseudo-correlation interference caused by differences in time bases of multi-source data. Lifetime migration prediction based on a physical constraint network that integrates full-lifetime runtime timing data and material physical parameters can be achieved by inputting the aligned time-series data into the network for forward propagation, while simultaneously using physical parameters to calculate constraint gradients to guide weight updates. In a specific embodiment, lifetime migration prediction based on a physical constraint network that integrates full-lifetime runtime timing data and material physical parameters can be achieved by using an adaptive weight strategy to dynamically balance data loss and physical loss, and by using course learning to gradually increase the strength of physical constraints, thereby enabling high-precision lifetime inference with small samples in cross-unit scenarios. The remaining life output of the first-stage rotor blade thermomechanical fatigue coupled damage can be obtained by analyzing the final layer output of the analytical network and then performing inverse normalization to obtain an engineering-usable life index. Furthermore, the remaining life output of the first-stage rotor blade thermomechanical fatigue coupled damage can be generated by estimating the output points and adding uncertainty bands generated by Monte Carlo Dropout, converting them into discrete health levels that conform to maintenance procedures. This transforms the abstract network output into actionable operational decision support information.

[0033] Taking the initial life management of a new type of F-class gas turbine peak-shaving power plant as an example, the hot-end component life migration prediction method of the physical constraint network in this embodiment can be used when a certain F-class gas turbine power plant has only been in commercial operation for six months and its own fault data is extremely scarce. The system automatically obtains the full life operation sequence of the source unit of the same model that has been in operation for five years, and extracts hundreds of peak-shaving operation samples from it. It uses the time stamp of each start-stop cycle to accurately capture the original monitoring information of temperature load and mechanical stress of the first-stage rotor blade, and completes millisecond-level time sequence alignment with the thermomechanical fatigue coupling damage parameters and creep physical constitutive model parameters of the alloy material measured in the laboratory. Then, these aligned data are input into the pre-constructed physical constraint network. The network is always constrained by the material constitutive equation during the training process, and can output predictions that conform to physical laws even when facing extreme peak-shaving rates that the new unit has never seen. Finally, the remaining life of the first-stage rotor blade under the current peak-shaving strategy is generated, which guides the power plant to reasonably arrange the first hot channel inspection time and avoid economic losses caused by blind conservative maintenance or non-stop risks caused by overestimating life.

[0034] In one embodiment, the full lifecycle operating sequence of the target gas turbine power unit is obtained. The full lifecycle operating sequence includes multiple peak-shaving operating sample samples, each with a corresponding start-stop cycle timestamp, including: Obtain the unit identifier to which the target first-stage rotor blade belongs, and retrieve the corresponding full-life-cycle operation database and material performance parameter library based on the unit identifier.

[0035] The unit identifier can be a unique numerical code or character sequence representing the gas turbine entity to which the target first-stage rotor belongs. It can be used as an index key to accurately associate the operating data and material properties of a specific unit, avoiding cross-unit data confusion. In this embodiment, the unit identifier can be automatically obtained by reading equipment nameplate information, asset management system ledgers, or sensor metadata tags. For example, the unit identifier may include, but is not limited to, KKS functional location codes, equipment serial numbers, and asset tag codes. The full-lifecycle operation database can be a dedicated data warehouse that structurally stores all operating parameters, event logs, and maintenance records of a specific gas turbine since commissioning. It can be used to provide a source domain sample set covering the entire service life to support transfer learning model training. Furthermore, the full-lifecycle operation database can retrieve and load corresponding partitioned data tables from a distributed historical database cluster based on the unit identifier. In an exemplary embodiment, the full-lifecycle operation database can collaborate with a material performance parameter library and automatically bind the operating load spectrum and material constitutive parameters through the unit identifier. In a specific embodiment, the full-lifecycle operation database can employ time-series databases, relational databases, data lake storage, etc.

[0036] The material performance parameter library can be a standardized set of parameters centrally managed for the mechanical response and damage characteristics of materials in the hot-end components of gas turbines under different temperature and stress conditions. This can be used to provide non-learnable deterministic prior knowledge for physical constraint networks to enhance small-sample generalization capabilities. Furthermore, the material performance parameter library can retrieve corresponding constitutive equation coefficients and fatigue curve data from the knowledge base after matching the material grade based on the unit identifier. For example, the material performance parameter library may include, but is not limited to, elastic modulus parameters, creep activation energy parameters, and fatigue life curve parameters. Obtaining the unit identifier of the target first-stage rotor blade can be achieved by parsing a unique identifier string from the equipment asset management system or sensor configuration files. Further, obtaining the unit identifier of the target first-stage rotor blade can be achieved through automatic identification by scanning on-site QR code labels, querying the asset tree of the enterprise asset management system via the application programming interface, or extracting the measurement point description field from the distributed control system. This allows for precise binding of data sources and physical entities and prevents cross-unit data misuse. Retrieving the corresponding full-lifecycle operation database and material performance parameter library based on the unit identifier can be achieved by performing a joint database query with the unit identifier as the primary key and loading the runtime sequence table and material parameter records. In one exemplary embodiment, retrieving the corresponding full-life-cycle operation database and material performance parameter library based on the unit identifier can be achieved by using a caching preloading mechanism to reduce the overhead of repeated input and output or by aggregating data across heterogeneous data sources through a federated query interface. This can automatically complete the contextual association between source domain data and mechanism parameters and improve data preparation efficiency.

[0037] It receives the peak load range and prediction assessment time range set by the user, where the prediction assessment time range includes the start time of operation and the end time of prediction.

[0038] The peak load range can be a user-specified upper and lower limit threshold range of active power used to filter samples of variable operating conditions. It can be used to define the boundary of flexible operation to exclude interference from base load steady-state data on transient damage modeling. In this embodiment, the peak load range may include, but is not limited to, the rated power percentage range, the absolute megawatt value range, and the relative output change rate range. The prediction and evaluation time range can be a time window of interest for lifetime migration prediction, including two boundary points: the start time and the end prediction time. It can be used to clarify the time span of the model input data and the timeliness constraints of the output results. For example, the prediction and evaluation time range can be a calendar time range, a cumulative operating hour range, or an equivalent start-stop count range. The start time can be the left boundary time point of the prediction and evaluation time range and mark the starting point of data collection and analysis. It can be used to determine the initial position of the full life cycle operation sequence to avoid introducing irrelevant historical noise. Furthermore, the start time can include, but is not limited to, the time of completion of the last major overhaul, the time of repair of a specific fault, or the start time of the natural calendar.

[0039] The termination prediction time can be the right boundary of the prediction and assessment time range and mark the target endpoint of the remaining life calculation. It can be used to set the cutoff condition for damage accumulation calculation to output the remaining available time with engineering significance. In a specific embodiment, the termination prediction time can include, but is not limited to, a future fixed date, the next planned maintenance date, or a dynamic health threshold trigger point. Receiving the peak load range and prediction and assessment time range set by the user can be achieved by parsing the numerical boundaries and timestamps input by the user through a human-computer interaction interface or configuration file. Furthermore, receiving the peak load range and prediction and assessment time range set by the user can be achieved by providing a drop-down menu to select a preset operating condition template or by supporting the parsing of natural language commands into structured parameters, thereby giving users the flexibility to customize the analysis granularity as needed and adapting to diverse operation and maintenance decision-making scenarios.

[0040] Based on the peak load range, start time, and end prediction time, multiple peak load operation samples are selected from the full life cycle operation database, and the start and stop cycle counts of each operation sample are extracted simultaneously as start and stop cycle timestamps for each operation sample.

[0041] The start-stop cycle count can be an integer count of the gas turbine's complete start-up, operation, and shutdown process, which can be used as a discretized timestamp to replace continuous absolute time and establish an alignment benchmark directly corresponding to the accumulation of physical damage. In this embodiment, the start-stop cycle count can be obtained by edge detection and cumulative counting of sequential control logic signals or speed state transitions. For example, the start-stop cycle count can be coordinated with peak-shaving operating condition samples to anchor continuous time-series data to damage event nodes with clear physical meaning. Filtering peak-shaving operating condition samples based on peak load range, start time, and end prediction time can be achieved by applying multiple condition filtering to the full-lifecycle operating database and extracting data segments that simultaneously meet load and time constraints. Furthermore, filtering peak-shaving operating condition samples based on peak load range, start time, and end prediction time can be achieved by using structured query language conditional clauses for combined queries or by using a memory computing framework for vectorized Boolean mask filtering, thereby accurately extracting effective samples reflecting the target peak-shaving characteristics and eliminating redundant steady-state data. Synchronously extracting the start-stop cycle count of each operating sample as the start-stop cycle timestamp can be achieved by reading the corresponding cumulative start-stop count value from each selected peak-shaving operating sample and assigning it to the timestamp field. In an exemplary embodiment, synchronously extracting the start-stop cycle count of each operating sample as the start-stop cycle timestamp can be achieved by obtaining the count through a window function associated with the event log table or by using a state machine backtracking algorithm to calculate the cycle number in real time. This allows for the establishment of an aligned coordinate system based on physical damage events and enhances the model's ability to capture transient nonlinear responses.

[0042] Taking the life assessment of an F-class gas turbine participating in deep peak shaving of the power grid as an example, the life migration prediction method of the hot-end components of the physical constraint network in this embodiment can be used when an operation and maintenance engineer needs to assess the damage status of the first-stage rotor blade after a certain F-class gas turbine has completed three months of deep peak shaving trial operation. The system first obtains the KKS functional location code of the unit as the unit identifier and automatically retrieves its full life operation database and the corresponding IN738LC alloy material performance parameter library accordingly. The engineer sets the peak load range to 30% to 50% of the rated power and the prediction assessment time range to the end of the last A overhaul to today. The system then selects all peak operating condition operation samples that meet the load range and are located within the specified time period from the database and extracts the cumulative number of start-stop cycles when each sample occurs as the start-stop cycle timestamp. These samples aligned with physical damage events, together with the material parameters, are input into the physical constraint network and finally generate the prediction result of the remaining life of the first-stage rotor blade considering the impact of deep peak shaving to guide the adjustment of subsequent maintenance strategies.

[0043] In one embodiment, the original monitoring information of temperature load and mechanical stress of the target gas turbine first-stage rotor blade at each start-stop cycle time point is collected, as well as the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade, including: Step S201: Obtain the design geometric parameters and material crystal orientation type of the first-stage rotor blade of the target gas turbine. The design geometric parameters include the blade root coordinates and film pore distribution density.

[0044] The design geometric parameters can be digital engineering features describing the macroscopic dimensions, spatial location, and cooling structure layout of the first-stage rotor blade. These parameters can provide a precise physical structural reference for damage area location and monitoring data spatial mapping. In this embodiment, the design geometric parameters can be extracted and structured from the original gas turbine design drawings, 3D CAD models, or manufacturing inspection reports. For example, design geometric parameters may include, but are not limited to, contour dimension parameters, cooling structure parameters, and assembly positioning parameters. The material crystal orientation type can be a classification identifier characterizing the crystallographic orientation of a single crystal or directionally solidified high-temperature alloy relative to the blade's principal axis. This can be used to reflect the anisotropic characteristics of the material's mechanical properties and thermal expansion coefficient to correct local damage assessment. Furthermore, the material crystal orientation type can be determined and recorded in the material archive using X-ray Laue backscattering or electron backscattering diffraction techniques. In an exemplary embodiment, the material crystal orientation type can be coordinated with the film gas pore distribution density to jointly determine the size and direction of the damage neighborhood calculation window. Obtaining the design geometric parameters and material crystal orientation type of the target gas turbine's first-stage rotor blade can be achieved by retrieving the geometric model file and material inspection report of the specified blade from an engineering database or digital twin platform. Furthermore, the design geometric parameters and material crystal orientation type of the first-stage rotor blade of the target gas turbine can be obtained by automatically extracting key dimensions from the STEP format 3D model and reading the crystal orientation annotations from the PDF material certificate using OCR, thereby establishing a digital prior characterization of the component's physical structure and material properties.

[0045] Step S202: Determine the damage neighborhood calculation window range based on the air film pore distribution density and material crystal orientation type.

[0046] The film cooling hole distribution density can be defined as the number of outflow holes for film cooling per unit area of ​​blade surface and their spatial arrangement density. It can be used to quantify the modulation intensity of the cooling structure on local temperature gradients and stress concentration effects. In this embodiment, the film cooling hole distribution density can be calculated by counting the number of holes in a specific area based on cooling design drawings and dividing by the corresponding surface area. For example, the film cooling hole distribution density can be coordinated with the material crystal orientation type to dynamically adjust the damage neighborhood calculation window range to adapt to different cooling efficiency regions. In an exemplary embodiment, the film cooling hole distribution density can adopt uniform array density, gradient density, or locally denser density. The damage neighborhood calculation window range can be a spatial subdomain dynamically defined around potentially high-damage-risk points for local physical field analysis and parameter extraction. It can be used to focus on local abnormal responses caused by key microstructural features to improve damage modeling accuracy. Furthermore, the damage neighborhood calculation window range can adaptively adjust its size and shape according to the film cooling hole distribution density and the material crystal orientation type. In a specific embodiment, the damage neighborhood calculation window range can be coordinated with design geometric parameters to further define the boundary range of the coupled damage calculation region. The damage neighborhood calculation window range is determined based on the film pore distribution density and the material crystal orientation type. This can be achieved by mapping the film pore density and crystal orientation type to window size and orientation parameters using preset rules or lookup tables. Furthermore, determining the damage neighborhood calculation window range based on the film pore distribution density and the material crystal orientation type can be achieved by calculating the window radius using empirical formulas, rotating the elliptical axis by the crystal orientation deflection angle, and using a machine learning model to predict the optimal window configuration. This allows for adaptive capture of local stress concentrations and thermal gradient anomalies caused by differences in cooling structure and crystal orientation.

[0047] Step S203: Based on the blade root coordinates, obtain the original monitoring information of temperature load and mechanical stress corresponding to each start-stop cycle time stamp.

[0048] The blade root coordinates can be a reference origin defining the spatial position of the primary rotor blade tenon mounting reference surface in the rotor coordinate system. This reference origin can serve as a unified anchor point for aligning all monitoring information with the geometric feature space, eliminating the influence of assembly errors. In this embodiment, the blade root coordinates can be determined based on the reference system in the blade machining drawings or the assembly measurement report. For example, the blade root coordinates can use the origin of the rotor rectangular coordinate system, the origin of the tenon groove local coordinate system, or the origin of the relative coordinate system of the center of gravity. Obtaining the original monitoring information of temperature load and mechanical stress corresponding to each start-stop cycle time stamp based on the blade root coordinates can be achieved by converting sensor readings to a unified blade body coordinate system using the blade root coordinates as a spatial reference and extracting data according to the time stamp. Furthermore, obtaining the original monitoring information of temperature load and mechanical stress corresponding to each start-stop cycle time stamp based on the blade root coordinates can be achieved by applying a rigid body transformation matrix to map the measurement point data to the blade root coordinate system and reconstructing the full-field load using finite element shape function interpolation. This ensures accurate mapping between the load data and the physical structure to support subsequent regional analysis.

[0049] Step S203: Determine the coupled damage calculation region based on the blade design geometric parameters and the damage neighborhood calculation window range, and obtain the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters within the calculation region.

[0050] The coupled damage calculation region can be an effective physical domain for performing coupled thermomechanical fatigue and creep analysis, determined by the comprehensive geometry and damage neighborhood window. It can be used to eliminate low-sensitivity redundant areas, concentrating physical parameter acquisition and network computing resources on high-risk locations. In this embodiment, the coupled damage calculation region can project the damage neighborhood calculation window range onto the solid surface or volume mesh defined by the blade design geometry parameters. For example, the coupled damage calculation region can include, but is not limited to, surface patch areas, volumetric layering areas, and interface transition areas. Determining the coupled damage calculation region based on the blade design geometry parameters and the damage neighborhood calculation window range can be achieved by cutting out a sub-region that coincides with the damage neighborhood window on the blade solid defined by the design geometry parameters. Furthermore, determining the coupled damage calculation region based on the blade design geometry parameters and the damage neighborhood calculation window range can be achieved by performing Boolean intersection operations on the 3D mesh to extract sub-domains and directly defining local UV ranges on the parametric surface, thus accurately defining the high-risk analysis domain to exclude interference from non-sensitive areas. Obtaining the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters within the calculation region can be achieved by extracting only a subset of parameters matching the spatial location of the coupled damage calculation region from the material parameter library. Furthermore, the parameters of thermomechanical fatigue coupling damage and creep physical constitutive model within the computational region can be obtained by filtering parameter entries by region label and interpolating local equivalent parameters based on the region center coordinates. This allows the physical constraint network input to focus on key parts, thereby improving modeling efficiency and prediction reliability.

[0051] Taking the life assessment of a new type of single-crystal blade peak-shaving power plant as an example, the life migration prediction method of the hot-end components of the physical constraint network in this embodiment can be as follows: In the early stage of commissioning of an F-class gas turbine peak-shaving power plant using a second-generation single-crystal alloy first-stage rotor blade, the system first obtains the design geometric parameters (including blade root coordinates and film pore distribution density) and material crystal orientation type (<001> main orientation) of the blade; Based on the combination of high film pore distribution density at the leading edge and <001> crystal orientation, a chordally elongated elliptical damage neighborhood calculation window range is dynamically determined; Then, based on the blade root coordinates, the original monitoring information of temperature load and mechanical stress in the corresponding area of ​​the window at each start-stop cycle time stamp is accurately extracted; Then, combined with the design geometric parameters, the elliptical window is projected onto the three-dimensional model of the blade to form a coupled damage calculation area; Finally, only the applicable thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters in this area are obtained and input into the physical constraint network. This process effectively avoids redundant calculations in the low-stress zone of the blade midline, enabling the model to accurately capture transient thermal shock damage around the leading-edge cooling holes even in the absence of measured degradation data, significantly improving the engineering applicability of remaining life prediction.

[0052] In one embodiment, the damage neighborhood calculation window range is determined based on the air film pore distribution density and the material crystal orientation type, including: The equivalent spacing of the air film pores is calculated based on the air film pore distribution density, and the radius of the first window is obtained from the damage window mapping table based on the equivalent spacing.

[0053] The equivalent spacing of the film cooling pores can be used to transform the non-uniformly distributed film cooling pore density into a continuous geometric equivalent distance characterizing the density of local cooling structures. This can be used to quantify the spatial modulation scale of the cooling pore array on local thermal stress concentration areas. In this embodiment, the equivalent spacing of the film cooling pores can be calculated by taking the square root of the reciprocal of the number of film cooling pores per unit area or by using the average side length of the Voronoi diagram. Exemplarily, the equivalent spacing of the film cooling pores can include, but is not limited to, the arithmetic mean spacing, the area equivalent spacing, and the nearest neighbor average spacing. The damage window mapping table can be a lookup table of the pre-set correspondence between the equivalent spacing of the film cooling pores and the radius of the first window. This can be used to quickly convert discrete geometric features into spatial scale parameters suitable for damage analysis. In an exemplary embodiment, the damage window mapping table can be generated based on a large number of thermo-mechanical coupling simulations or statistical regressions of historical damage cases and stored in a configuration file. Furthermore, the damage window mapping table can work in conjunction with the equivalent spacing of the film cooling pores to achieve a lossless mapping from structural density to computational domain size. In a specific embodiment, the damage window mapping table can employ simulation calibration mapping tables, experimental statistical mapping tables, hybrid driven mapping tables, etc.

[0054] The first window radius can be an initial spatial scale reflecting the influence range of the cooling structure, obtained by looking up the equivalent spacing of the film cooling holes in a damage window mapping table. It can be used to characterize the geometric boundaries of the local temperature gradient and thermal stress anomaly zone caused by film cooling. In this embodiment, the first window radius can be obtained by interpolation or direct matching in the mapping table using the calculated equivalent spacing of the film cooling holes as an index. For example, the first window radius can include, but is not limited to, the exact matching radius, the linear interpolation radius, and the piecewise constant radius. Calculating the equivalent spacing of the film cooling holes based on the distribution density of the film cooling holes can be achieved by converting the number of film cooling holes per unit area into a scalar value representing the average distance between holes. Furthermore, calculating the equivalent spacing of the film cooling holes based on the distribution density of the film cooling holes can be achieved by using the inverse square root formula of the area or by using Delaunay triangulation to obtain the average side length, thereby transforming the discrete hole array into a continuous geometric feature to support subsequent mapping operations. Obtaining the first window radius in the damage window mapping table based on the equivalent spacing can be achieved by querying a preset mapping table using the equivalent spacing of the film cooling holes as the key to obtain the corresponding radius value. Furthermore, the first window radius based on the equivalent spacing in the damage window mapping table can be obtained by performing precise key-value matching or by using linear or spline interpolation to obtain an intermediate value, thereby quickly converting structural density into the spatial scale of thermal stress influence.

[0055] The radius of the second window is determined based on the crystal orientation of the material, and the radius of the second window is a preset multiple of the range of influence of the material's anisotropy.

[0056] The second window radius can be a spatial scale reflecting the influence range of anisotropic mechanical response of the crystal, determined based on the material's crystal orientation type. It can be used to cover the stress redistribution region caused by directional changes in the elastic modulus and thermal expansion coefficient due to differences in grain orientation. In this embodiment, the second window radius can be obtained by multiplying the material anisotropic influence range by a preset safety factor. Exemplarily, the second window radius can include, but is not limited to, empirical conservative factors, simulation verification factors, and probability confidence factors. The material anisotropic influence range can be a physical spatial scale characterizing the significant directional dependence of the mechanical properties of single-crystal or directionally solidified materials. It can be used to provide a physical benchmark for calculating the second window radius to avoid subjective amplification or reduction of material effects. In an exemplary embodiment, the material anisotropic influence range can be obtained by measuring the distance at which the performance deflection decays to within a threshold through crystal plasticity finite element simulation or directional tensile testing. In a specific embodiment, the material anisotropic influence range can include, but is not limited to, the elastic anisotropic range, the plastic flow anisotropic range, and the thermal expansion anisotropic range. Determining the second window radius based on the material's crystal orientation can be achieved by obtaining the anisotropy influence range of the material from the crystal orientation identifier and multiplying it by a preset factor. Furthermore, determining the second window radius based on the material's crystal orientation can be achieved by obtaining the standard influence range from a table and multiplying it by a fixed factor, or by dynamically adjusting the factor based on the crystal orientation deviation angle. This allows for the quantification of the influence of crystal orientation on the spatial expansion of the mechanical response.

[0057] The first window radius is compared with the second window radius, and the larger value is taken as the base window radius.

[0058] The basic window radius can be the minimum necessary damage analysis spatial scale determined by considering both cooling structure and material crystal orientation. It ensures that the computational domain is sufficient to cover the superimposed effect region of all potential high-damage driving factors. In this embodiment, the basic window radius can be obtained by comparing the maximum value of the first and second window radii. Furthermore, the basic window radius can work in conjunction with the first and second window radii to ensure that the computational domain simultaneously accommodates both structural and material effects through a strategy of taking the largest value. For example, the basic window radius can include, but is not limited to, strictly taking the largest radius, a weighted fusion radius, or a conditional switching radius. Comparing the first and second window radii and taking the larger value as the basic window radius can be achieved by performing a numerical comparison of the two radius values ​​and retaining the larger one. Furthermore, taking the larger value as the basic window radius can be achieved by directly taking the maximum value or by introducing a weighted maximization factor, thereby ensuring that the computational domain simultaneously covers the damage-sensitive regions dominated by both structure and material.

[0059] By adding a fatigue damage margin to the basic window radius, the range of the damage neighborhood calculation window is obtained.

[0060] The fatigue damage margin can be an additional spatial buffer beyond the basic window radius to accommodate crack initiation and early propagation under cyclic loading. It can be used to prevent damage from exceeding the basic window and being missed due to transient overload or microscopic defects. In this embodiment, the fatigue damage margin can be calculated using a table or formula based on the material's fatigue limit strain amplitude and expected life cycle number. Furthermore, the fatigue damage margin can work in conjunction with the basic window radius to form the final damage neighborhood calculation window range. For example, the fatigue damage margin can include, but is not limited to, a fixed percentage margin, a life-related margin, and a load spectrum adaptive margin. Adding the fatigue damage margin to the basic window radius to obtain the damage neighborhood calculation window range can be achieved by superimposing the fatigue damage margin onto the basic window radius to form the final window boundary. Furthermore, adding the fatigue damage margin to the basic window radius to obtain the damage neighborhood calculation window range can be achieved by additively superimposing a fixed margin or dynamically expanding it as a percentage of the basic radius, thereby reserving physical buffer space for cyclic damage evolution to improve prediction robustness.

[0061] For example, in the scenario of adaptive window construction in the life migration prediction of single-crystal blades across units, the life migration prediction method of hot-end components in the physical constraint network of this embodiment can be as follows: When predicting the life migration of a newly commissioned F-class gas turbine first-stage rotor blade, the system first calculates the equivalent spacing of the film pores as 1.2 mm based on the high film pore distribution density in its leading edge region, and obtains the first window radius of 3.6 mm by querying the damage window mapping table; at the same time, it identifies that the blade material is a <001> oriented second-generation single-crystal alloy, finds that its material anisotropy influence range is 2.8 mm, multiplies it by a preset multiple of 1.5 to obtain the second window radius of 4.2 mm; compares the two and takes the larger value of 4.2 mm as the base window radius; then, based on the expected life cycle number corresponding to the peak shaving frequency of the unit, it determines the fatigue damage margin as 0.8 mm; finally, the damage neighborhood calculation window range is 5.0 mm. This window precisely covers the superposition range of the thermal gradient region of the leading edge cooling hole and the crystal orientation sensitive stress region, and reserves a buffer space for microcrack propagation under peak shaving cycle. This allows the physical constraint network to still lock the real high-risk area even without measured degradation data, significantly improving the engineering credibility of remaining lifetime prediction.

[0062] In one embodiment, based on the original monitoring information and damage parameters and physical constitutive model parameters corresponding to each start-stop cycle time stamp, the life migration prediction of the target cross-unit first-stage rotor blade is performed, generating the remaining life output result of the thermomechanical fatigue coupled damage of the first-stage rotor blade, including: Based on the thermomechanical fatigue coupled damage parameters, creep physical constitutive model parameters, and the original monitoring information corresponding to each start-stop cycle time stamp, the actual cumulative damage component corresponding to each start-stop cycle time stamp is calculated.

[0063] The actual cumulative damage component can be calculated based on the material's physical constitutive model and real-time monitoring data, representing the irreversible material degradation caused by the interaction of thermomechanical fatigue and creep within a single start-stop cycle. It can be used as a physical truth benchmark to quantitatively correct the prediction bias of the data-driven model. In this embodiment, the actual cumulative damage component can be obtained by numerically solving the integral equation defined by substituting the original monitoring information of temperature load and mechanical stress into the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters. For example, the actual cumulative damage component can be used in conjunction with a transfer model and as a post-processing correction factor to force the neural network output to conform to the material failure law. Furthermore, the actual cumulative damage component can include, but is not limited to, pure fatigue damage components, pure creep damage components, and fatigue-creep interactive damage components. Calculating the actual cumulative damage component corresponding to each start-stop cycle time stamp can be achieved by performing numerical integration on the damage rate function composed of the original monitoring information and physical parameters, with the start-stop cycle time stamp as the boundary of the integration interval. In one exemplary embodiment, the actual cumulative damage component corresponding to each start-stop cycle time stamp can be calculated by using the adaptive step-size Runge-Kutta method to solve the unsteady-state damage integral or by using a piecewise analytical approximation formula to replace the complex numerical integral, thereby discretizing the continuous physical degradation process into a computable time-series index synchronized with data sampling.

[0064] The initial prediction results of the migration model are corrected based on the actual cumulative damage components corresponding to each start-stop cycle time stamp, and the corrected remaining life prediction values ​​corresponding to each start-stop cycle time stamp are obtained, generating the remaining life output results of the thermomechanical fatigue coupling damage of the first-stage blade.

[0065] The transfer model can be a deep learning life prediction network pre-trained on the source unit's full lifespan and adapted to the target cross-unit. It can provide a basic remaining lifespan trend estimate when fault data is lacking in the early stages of the target unit's service. In one specific embodiment, the transfer model can be obtained by fine-tuning the distribution differences using a small number of samples in the target domain or by adjusting the distribution differences after training the feature extractor and regression head on the source domain data. Furthermore, the transfer model can be coordinated with the actual damage accumulation component, and its initial prediction result can be used as the object being corrected, receiving anchoring corrections based on physical quantities. For example, the transfer model can employ instance-based transfer models, feature-based transfer models, shared representation-based transfer models, etc. The initial prediction result can be the original remaining lifespan estimate output by the transfer model based solely on data-driven inference before applying physical correction. It can be used to reflect the degree of data distribution matching but may deviate from the actual physical degradation trajectory as an intermediate computational state. In this embodiment, the initial prediction result can include, but is not limited to, single-point instantaneous prediction values, sliding window average prediction values, and sequence trend prediction values.

[0066] The corrected remaining lifetime prediction value can be the final lifetime assessment result that satisfies both data statistical laws and material physical constraints after being corrected by the actual damage accumulation component. It can be used as a direct constituent unit of the remaining lifetime output result of the thermomechanical fatigue coupling damage of the first-stage rotor blade for operation and maintenance decision-making. Furthermore, the corrected remaining lifetime prediction value can include, but is not limited to, a fully physical coverage correction value, a weighted fusion correction value, and a residual compensation correction value. Correcting the initial prediction result of the transfer model based on the actual damage accumulation component can be achieved by using the actual damage accumulation component as a feedback signal to adjust the amplitude or trend slope of the initial prediction result to approximate the physical true value. In an exemplary embodiment, correcting the initial prediction result of the transfer model based on the actual damage accumulation component can be achieved by constructing a physical residual network to learn the mapping deviation between data prediction and physical calculation, or by using a Kalman filter framework to update the data prediction state with physical quantities as observations. This can eliminate the systematic prediction error of the data-driven model caused by the target unit's operational boundary offset. Obtaining the corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp can be achieved by reorganizing the corrected values ​​in chronological order into a lifetime time series sequence strictly aligned with the start-stop cycle time stamps. For example, the corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp can be obtained by serially outputting the correction value time stamp by time stamp and caching it or by batch generating a complete correction sequence and then uniformly packaging and outputting it, thereby forming a structured lifetime prediction product that can be directly used for engineering evaluation and has physical consistency.

[0067] Taking the online monitoring and early warning of the life of the first-stage rotor blade of a peak-shaving gas turbine as an example, the life migration prediction method of the hot-end component of the physical constraint network in this embodiment can be as follows: During the operation of an F-class gas turbine undertaking the task of deep peak shaving of the power grid, the system automatically captures the original monitoring information of temperature load and mechanical stress corresponding to the time stamp of the cycle after each start-stop cycle. Combined with the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the blade material, the actual cumulative damage component of the current cycle is calculated by an adaptive step-size integral algorithm. At the same time, the migration model outputs an initial prediction result based on the current operating condition characteristics. The system then starts the correction processing module to use the actual cumulative damage component as a physical anchor point and corrects the overestimation deviation of life caused by the difference in cooling efficiency of the new unit in the initial prediction result through the residual compensation mechanism. Finally, the corrected remaining life prediction value is obtained and updated to the monitoring screen. When the correction value of multiple consecutive cycles shows an accelerated downward trend, an early maintenance warning is triggered to avoid the risk of missed detection that may be caused by relying solely on the uncorrected data model.

[0068] In one embodiment, the initial prediction results of the migration model are corrected based on the actual cumulative damage component corresponding to each start-stop cycle time stamp to obtain the corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp, including: Iterate through multiple peak-shaving operation samples throughout the entire lifecycle of the runtime sequence to obtain the first set of peak-shaving operation samples.

[0069] The first set of peak-shaving operation samples can be specific peak-shaving condition data units selected sequentially from the entire lifecycle runtime sequence for performing single damage mechanism separation and correction calculations. These units can serve as the operational carrier for damage region division and component extraction within the current iteration cycle. In this embodiment, the first set of peak-shaving operation samples may include, but is not limited to, rapid load increase samples, deep load decrease samples, and load maintenance samples. Obtaining the first set of peak-shaving operation samples by traversing multiple peak-shaving operation samples in the entire lifecycle runtime sequence can be achieved by sequentially accessing peak-shaving condition segments in the entire lifecycle runtime sequence according to time order or index number and loading the current batch of data. Furthermore, obtaining the first set of peak-shaving operation samples by traversing multiple peak-shaving operation samples in the entire lifecycle runtime sequence can be achieved by using a sliding window to advance sample by sample or by using a priority queue to select samples sorted by damage severity, thereby ensuring that all peak-shaving conditions are included in the damage mechanism separation processing flow without omission.

[0070] Extract the first start-stop cycle timestamp of the first set of peak-shaving operation samples, and obtain the first actual damage accumulation component based on the first start-stop cycle timestamp.

[0071] The first start-stop cycle time stamp can be a unique time index identifying the start and end boundaries of the physical process corresponding to the first set of peak-shaving operation samples. It can be used to limit the calculation interval of the first actual damage accumulation component and the temporal range of subsequent damage region division. The first actual damage accumulation component can be the total damage accumulation quantification value calculated for the first set of peak-shaving operation samples within the first start-stop cycle time stamp interval. It can be used as a physical basis for determining the relative dominance of creep and thermal fatigue under the current operating condition. For example, the first actual damage accumulation component can work in conjunction with the creep-dominated damage region and the thermal fatigue-dominated damage region, and its numerical distribution characteristics directly determine the boundary position of the two types of regions. Extracting the first start-stop cycle time stamp of the first set of peak-shaving operation samples and obtaining the first actual damage accumulation component can be achieved by parsing the current sample metadata to obtain the time boundary identifier and calling the pre-calculated damage integral result or performing the integration operation in real time. In an exemplary embodiment, extracting the first start-stop cycle time stamp of the first set of peak-shaving operation samples and obtaining the first actual damage accumulation component can be achieved by directly reading the key value from the cache dictionary or triggering the online calculation of a lightweight integrator, thereby establishing an accurate mapping relationship between the current sample and its physical damage quantity.

[0072] Based on the first actual damage accumulation component, the first group of peak-shaving operation samples are divided into creep-dominated damage region and thermal fatigue-dominated damage region.

[0073] The creep-dominated damage region can be a time segment in the first set of peak-shaving operation samples where the temperature is high, the stress duration is long, and the material degradation is mainly driven by time-dependent plastic deformation. This region can be used to accurately locate the operating condition intervals where steady-state high-temperature damage features need to be extracted to avoid mechanism confusion. In this embodiment, the creep-dominated damage region can be automatically identified and generated based on continuous time periods where the proportion of creep damage in the first actual damage accumulation component exceeds a preset threshold. Furthermore, the creep-dominated damage region can complement the thermal fatigue-dominated damage region to jointly cover the physical process of the complete start-stop cycle. The creep-dominated damage region can include, but is not limited to, the peak temperature creep region, the plateau temperature creep region, and the residual creep region during the cooling phase. The thermal fatigue-dominated damage region can be a time segment in the first set of peak-shaving operation samples where the temperature gradient changes drastically and the material degradation is mainly driven by alternating thermal stress. This region can be used to isolate intervals with significant pure thermal shock effects to support independent extraction of cyclic plastic damage features. In a specific embodiment, the thermal fatigue-dominated damage region can be automatically identified and generated based on transient time periods where the proportion of fatigue damage in the first actual damage accumulation component is significantly higher than that of creep damage. The thermal fatigue-dominated damage region can be temporally adjacent to or intersect with the creep-dominated damage region, and their boundary is dynamically determined by the competitive relationship of the damage mechanisms. The thermal fatigue-dominated damage region may include, but is not limited to, the ignition heating shock zone, the load shedding rapid cooling zone, and the grid-connected transition thermal shock zone.

[0074] Dividing the creep-dominated damage region and the thermal fatigue-dominated damage region based on the first actual cumulative damage component can be achieved by analyzing the compositional changes of the first actual cumulative damage component along the time axis and segmenting the time series intervals with the mechanism dominance switching point as the boundary. Furthermore, the division of the creep-dominated damage region and the thermal fatigue-dominated damage region based on the first actual cumulative damage component can be achieved through hard segmentation by setting a fixed damage ratio threshold or through soft boundary transition segmentation using a fuzzy membership function, thereby enabling adaptive decoupling of different physical damage mechanisms under complex variable working conditions.

[0075] For the first group of peak-shaving operation samples, the first damage component is extracted in the creep-dominated damage region, and the second damage component is extracted in the thermal fatigue-dominated damage region.

[0076] The first damage component can be an independent damage metric characterizing time-dependent material degradation extracted within the creep-dominated damage region. It can provide pure creep damage input unaffected by thermal fatigue to support subsequent nonlinear coupling synthesis. For example, the first damage component can be obtained by applying a creep constitutive integral to the temperature and stress time series within the region and removing the fatigue contribution. The second damage component can be an independent damage metric characterizing cyclic thermal stress-induced damage extracted within the thermal fatigue-dominated damage region. It can provide pure thermal fatigue damage input to ensure accurate weighting of each mechanism during coupling synthesis. In this embodiment, the second damage component can be obtained by applying a thermomechanical fatigue model to the strain amplitude and temperature phase difference within the region and removing the creep effect. Extracting the first damage component in the creep-dominated damage region and the second damage component in the thermal fatigue-dominated damage region can be achieved by applying corresponding single-mechanism damage models for local integration or feature extraction within the two defined regions. Furthermore, extracting the first damage component in the creep-dominated damage region and the second damage component in the thermal fatigue-dominated damage region can be achieved by using a mechanism-specific sub-model for parallel computation or by using a masking mechanism to shield non-target region data before unified computation. This allows for obtaining independent damage characterization quantities with clear physical meaning and no mutual contamination.

[0077] The first damage component and the second damage component are coupled and synthesized to generate the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp.

[0078] The corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp is obtained by using the same method as obtaining the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp.

[0079] The coupling synthesis of the first and second damage components can be achieved by combining two independent damage components into a unified lifetime index through a nonlinear interaction model based on the physical laws of material failure. In an exemplary embodiment, the coupling synthesis of the first and second damage components can be achieved by using the Lemaitre-Chaboche nonlinear accumulation rule or by using a coupling function based on energy dissipation, thereby realistically reflecting the accelerated damage effect under the interaction of thermomechanical fatigue and creep rather than a simple linear superposition.

[0080] Taking the refined assessment of the lifespan of deep peak-shaving gas turbine blades as an example, the life migration prediction method for hot-end components of the physical constraint network in this embodiment can be as follows: when a gas turbine performs a deep peak-shaving task with a load of 30% to 100%, the system traverses its entire lifespan operating sequence to obtain the current deep load reduction sample as the first set of peak-shaving operation samples; after extracting the start-stop cycle time stamps corresponding to the sample, the first actual cumulative damage component calculated in advance is retrieved; analysis shows that the creep proportion of this component reaches 75% during the load holding stage, while the thermal fatigue proportion exceeds 80% in the early stage of rapid load reduction. Based on this, it is automatically divided into a creep-dominated damage region and a thermal fatigue-dominated damage region; then, in the creep-dominated region, only the Norton-Bailey model is used to extract the first damage component, and in the thermal fatigue-dominated region, only the Manson-Coffin equation is used to extract the second damage component; finally, the two are synthesized into the corrected remaining life prediction value through the Chaboche nonlinear coupling model, avoiding the problem of underestimating interactive damage in the traditional linear Miner's rule, and making the prediction results highly consistent with the coexistence of grain boundary voids and transgranular cracks observed by metallographic examination.

[0081] In one embodiment, for the first group of peak-shaving operation samples, a first damage component is extracted in the creep-dominated damage region, and a second damage component is extracted in the thermal fatigue-dominated damage region, including: Obtain the coupling damage separation threshold.

[0082] The coupling damage separation threshold can be a physical criterion parameter used to define the respective contribution weights of creep and thermomechanical fatigue in the mixed damage process. It can serve as a quantitative switch for damage component separation to ensure that the mechanism decoupling process has clear physical boundaries rather than subjective divisions. In this embodiment, the coupling damage separation threshold can be obtained based on fitting material interaction damage test data or a dimensionless critical ratio set according to engineering experience. Furthermore, the coupling damage separation threshold can be coordinated with the cumulative damage introduced by Larson-Miller creep and the cumulative damage introduced by thermomechanical fatigue to determine when to activate which separation algorithm. For example, the coupling damage separation threshold can include, but is not limited to, a time fraction threshold, a strain energy density ratio threshold, and a damage rate ratio threshold. The coupling damage separation threshold can be obtained by reading from a preset material parameter library or by calculating the currently applicable separation criterion value through a real-time adaptive algorithm. In an exemplary embodiment, the coupling damage separation threshold can be obtained by querying a static JSON configuration file to load a fixed threshold or by dynamically interpolating a threshold based on the current temperature and stress level, thereby providing a physical benchmark adapted to the current working conditions for damage mechanism decoupling.

[0083] In the creep-dominated damage region, the first group of peak-shaving operation samples are subjected to damage component separation processing based on the coupling damage separation threshold, and the cumulative damage introduced by Larson-Miller creep is separated and extracted to obtain the first damage component.

[0084] The cumulative damage introduced by Larson-Miller creep can be an independent damage quantity characterizing the time-dependent degradation of materials under high-temperature long-term stress, calculated using the Larson-Miller parametric method. This can be used to provide a pure steady-state high-temperature damage metric consistent with classical creep theory to avoid interference from transient thermal shock. In a specific embodiment, the cumulative damage introduced by Larson-Miller creep can be obtained by substituting the temperature and stress time series within the creep-dominant region into the Larson-Miller equation, integrating, and normalizing to the fracture life. Damage component separation processing of the first group of peak-shaving operation samples based on the coupled damage separation threshold can be achieved by allocating the sample data stream to the corresponding single-mechanism damage extraction channel based on the coupled damage separation threshold. For example, damage component separation processing of the first group of peak-shaving operation samples based on the coupled damage separation threshold can be achieved by using a hard threshold switch for binary classification routing or by using a soft weighting function to achieve smooth transition separation, enabling precise diversion of mixed damage signals to independent physical channels. The first damage component, derived from the cumulative damage introduced by Larson-Miller creep, can be obtained by integrating the shunt data using the Larson-Miller constitutive model within the creep-dominant region and outputting a normalized damage value. Furthermore, this extraction can be achieved by using a lookup table to accelerate LM parameter calculation or by directly solving the continuous damage equation using a numerical integrator, thus obtaining a highly physically deterministic quantification result of pure creep damage.

[0085] In the thermal fatigue-dominated damage region, after creep coupling correction of the first group of peak-shaving operation samples, damage component separation is performed in combination with the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

[0086] The cumulative damage introduced by thermo-mechanical fatigue can be an independent damage quantity characterizing the cyclic plastic degradation of the material under the combined action of alternating thermal stress and mechanical load after removing residual creep effects. This can be used to obtain pure thermo-mechanical fatigue damage input to ensure accurate calculation of nonlinear interaction terms during subsequent coupling synthesis. In this embodiment, the cumulative damage introduced by thermo-mechanical fatigue can be obtained by first subtracting the creep coupling correction amount within the thermal fatigue-dominant region and then applying a fatigue damage model to the remaining strain amplitude. For example, the cumulative damage introduced by thermo-mechanical fatigue can include, but is not limited to, Manson-Coffin fatigue damage, Coffin-Manson-Taira fatigue damage, and strain energy density fatigue damage. Creep coupling correction for the first set of peak-shaving operation samples can be performed by estimating and subtracting the pseudo-fatigue damage contributed by the creep mechanism within the thermal fatigue-dominant region. In an exemplary embodiment, creep coupling correction for the first set of peak-shaving operation samples can be achieved by extrapolating and subtracting based on the creep state at the previous moment or by using a coupled constitutive model to decouple the creep component in reverse, thereby eliminating the contamination of fatigue damage assessment by residual creep during transient processes. The damage component separation process, which combines the coupling damage separation threshold to separate and extract the cumulative damage introduced by thermo-cyclic mechanical fatigue, yields a second damage component. This can be achieved by applying the separation threshold again to the data after creep coupling correction to confirm purity, followed by fatigue damage model calculation. Furthermore, the damage component separation process, which combines the coupling damage separation threshold to separate and extract the cumulative damage introduced by thermo-cyclic mechanical fatigue, can be achieved by sequentially executing a correction-separation-calculation process or by using end-to-end differentiable modules to jointly optimize the correction and extraction parameters, resulting in a pure thermo-mechanical fatigue damage metric that is unaffected by creep.

[0087] For example, in the scenario of damage decoupling in the cross-unit migration life assessment of F-class gas turbine blades, the hot-end component life migration prediction method of the physical constraint network in this embodiment can be as follows: When predicting the life migration of a newly commissioned F-class gas turbine first-stage rotor blade, the system first obtains the coupling damage separation threshold of the alloy material in the current peak-shaving temperature range. For a set of deep peak-shaving samples, if the damage rate ratio is detected to exceed the threshold during the load holding phase, it is determined to be creep-dominated. Then, the first damage component for this period is extracted based on the Larson-Miller equation integral. In the subsequent rapid load reduction phase, although thermal fatigue is dominant, the system first performs creep coupling correction, deducts the residual creep strain caused by the high temperature holding phase, and then confirms the thermal fatigue dominance by combining the same separation threshold. Finally, the Manson-Coffin model is applied to extract the pure second damage component. This step-by-step decoupling strategy can accurately distinguish the contribution of the two damage mechanisms even if the target unit has a short operating history and few fault samples, avoiding the misjudgment of creep as fatigue-induced overestimation of life, and significantly improving the consistency between the migration prediction results and the measured degradation trend of the source unit.

[0088] In one embodiment, in the thermal fatigue-dominated damage region, after creep coupling correction is performed on the first group of peak-shaving operation samples, damage component separation is performed based on the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component, including: In the first set of peak-shaving operation samples, the original monitoring values ​​and creep coupling components of each sampling point in the thermal fatigue-dominated damage area were obtained.

[0089] The original monitoring values ​​can be direct readings from temperature and mechanical stress sensors at each sampling point within the thermal fatigue-dominated damage region, without physical correction. These can serve as the input reference for the creep coupling corrector and reflect the original physical state, including mixed damage effects. The creep coupling component can characterize the contribution of transient creep deformation caused by unsteady temperature and stress history within the thermal fatigue-dominated damage region. It can be used to quantify the pseudo-fatigue damage portion that needs to be removed from the original monitoring values ​​to support pure feature extraction. In this embodiment, the creep coupling component can be obtained by extrapolation from the constitutive model or real-time integration based on the creep state at the previous moment and the current operating parameters. Furthermore, the creep coupling component can work in conjunction with the creep coupling corrector to provide a precise basis for deduction. Obtaining the original monitoring values ​​and creep coupling components at each sampling point within the thermal fatigue-dominated damage region can be achieved by simultaneously reading the sensor data stream within the time window of that region and the pre-calculated transient creep contribution. In one exemplary embodiment, the original monitoring values ​​and creep coupling components of each sampling point within the thermal fatigue-dominated damage region can be obtained by batch pulling aligned data from a ring buffer or by acquiring and calculating creep components on demand via an event-triggered mechanism, thereby providing a complete and time-consistent input dataset for the correction process.

[0090] The creep coupling corrector is used to perform creep coupling correction on each sampling point in the thermal fatigue-dominated damage area to obtain the corrected damage value, and the original monitoring value is replaced with the corrected damage value.

[0091] The creep coupling corrector can be an algorithm module that performs creep effect stripping operations and converts the original monitoring values ​​into a pure fatigue feature space. It can be used to eliminate the nonlinear interference of residual creep during transient processes on thermomechanical fatigue damage assessment. For example, the creep coupling corrector can incorporate a creep-fatigue interactive constitutive relation and apply inverse compensation or residual decomposition processing to the input data. Furthermore, the creep coupling corrector can work in conjunction with the corrected damage value, and its output can directly constitute the pure input for subsequent separation processing. In a specific embodiment, the creep coupling corrector can include, but is not limited to, a constitutive inversion type corrector, a state observation type corrector, and an incremental subtraction type corrector. The corrected damage value can be a corrected physical quantity that retains only the thermocyclic mechanical fatigue-related characteristics after processing by the creep coupling corrector. It can be used to replace the original monitoring value as the input for damage component separation processing and ensure that the extraction results are not contaminated by creep. In this embodiment, the corrected damage value can include, but is not limited to, the corrected strain amplitude, the corrected stress range, and the corrected hysteresis loop area.

[0092] The corrected damage value is obtained by performing creep coupling correction on each sampling point using a creep coupling corrector. This can be achieved by inputting the original monitoring value and the creep coupling component into the corrector and performing a physical stripping operation to output the correction result. Furthermore, the corrected damage value obtained by performing creep coupling correction on each sampling point can be achieved by using point-by-point serial correction to ensure causality or by using vectorized parallel computation to improve processing efficiency, thereby generating a clean fatigue feature sequence free from creep variations. The original monitoring value is replaced with the corrected damage value, which can be done by overwriting the original value with the corrected value in memory or data stream and updating the input source of subsequent processing modules. In an exemplary embodiment, replacing the original monitoring value with the corrected damage value can be achieved by modifying the data structure in-situ to save memory or by creating a new data copy to retain the original record for auditing, thereby ensuring that downstream separation processing is performed only based on the physically corrected clean data.

[0093] Based on the corrected damage value and the coupling damage separation threshold, the damage component separation process is performed on the first group of peak-shaving operation samples, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

[0094] The damage component separation process based on the corrected damage value and the coupled damage separation threshold can be achieved by using the corrected damage value as input and the coupled damage separation threshold as the criterion to execute a thermal fatigue damage extraction algorithm. Furthermore, the damage component separation process based on the corrected damage value and the coupled damage separation threshold can be achieved by applying threshold gating to select valid data segments and then integrating, or by using weighted filtering to smooth the transition region and then calculating the cumulative amount. This allows the output of independent thermal cycle mechanical fatigue cumulative damage that only reflects the alternating thermal stress.

[0095] For example, in the scenario of accurately quantifying thermal fatigue damage of peak-shaving gas turbine blades, the hot-end component life migration prediction method of the physical constraint network in this embodiment can identify the thermal fatigue-dominated damage region during the rapid load reduction process of a certain F-class gas turbine and obtain the original monitoring values ​​of temperature and strain at sampling points per second in this region. At the same time, based on the creep state of the previous steady-state segment, the creep coupling component in the current transient process is extrapolated. The creep coupling corrector then performs constitutive inversion operation on each sampling point and deducts about 15% of the pseudo-fatigue strain contribution to generate a corrected damage value and replaces the original data in situ. Subsequently, the corrected value is combined with a preset strain energy density ratio threshold for separation processing, and the Manson-Coffin model is applied to the effective cycle segment that exceeds the threshold for integration. Finally, the cumulative damage introduced by pure thermal cycle mechanical fatigue is obtained as the second damage component. This avoids the residual creep in the early stage of load reduction being mistakenly included in the fatigue damage and makes the cross-unit migration prediction results highly consistent with the pure fatigue stripe spacing observed by metallographic inspection.

[0096] In one embodiment, the construction steps of the creep coupling corrector include: Collect the sample creep coupling record set and the sample original damage set, and annotate them to obtain the sample corrected damage set.

[0097] The sample creep coupling record set can be a structured dataset containing the transient creep contribution and corresponding operating parameters of each sampling point in historical operation or experiment. It can be used as one of the input features for training the creep coupling corrector and provide a quantitative characterization of creep disturbances. In this embodiment, the sample creep coupling record set can be generated and organized into a time-aligned format through bench test measurements, high-fidelity simulation calculations, or inversion of source unit operating data. Exemplarily, the sample creep coupling record set can include, but is not limited to, bench test record sets, digital twin simulation record sets, and field inversion record sets. The original sample damage set can be a set of uncorrected sensor monitoring values ​​or preliminary damage assessment values ​​time-aligned with the sample creep coupling record set. It can be used as another input channel for corrector training and reflects the original physical state containing mixed damage effects. In an exemplary embodiment, the original sample damage set and the sample creep coupling record set can together constitute a dual-channel input for the corrector to learn the mixed-to-pure mapping. The sample-corrected damage set can be a pure damage label dataset containing only thermomechanical fatigue components, verified by expert annotation or a high-precision physical model. It can be used as a supervisory signal during training to guide the corrector in learning physically consistent decoupling mappings. Furthermore, the sample-corrected damage set can be obtained by manually reviewing the original sample damage set and the sample creep coupling record set, or by constitutive inversion, to generate supervisory ground truth. In a specific embodiment, the sample-corrected damage set can employ expert-annotated sets, constitutive inversion annotation sets, multi-source cross-validation annotation sets, etc.

[0098] Collecting sample creep coupling record sets and sample original damage sets can be done by batch extracting and verifying the temporal consistency of the two sets of data from experimental databases, simulation platforms, or historical operating systems. Furthermore, the collection of sample creep coupling record sets and sample original damage sets can be achieved automatically by ETL tools and cleaned, or manually exported to CSV and then timestamped using a script, thus constructing a high-quality dual-channel training input foundation. Annotating the sample creep coupling record sets and sample original damage sets to obtain sample correction damage sets can be done by combining domain knowledge or high-precision physical models to generate corresponding clean fatigue damage labels for each set of input data. Furthermore, the annotation of sample creep coupling record sets and sample original damage sets to obtain sample correction damage sets can be achieved by organizing an expert team to blindly review the annotations or by calling a unified constitutive solver to automatically generate labels and conduct sampling verification, thereby establishing a physically realistic supervised benchmark to guide the learner's learning.

[0099] Using the sample creep coupling record set and the sample original damage set as inputs, and the sample corrected damage set as the supervision signal, the creep coupling corrector is trained by combining the pre-trained weights of the physical constraint network for transfer fine-tuning.

[0100] The pre-trained weights of the physical constraint network can be the initial state of neural network parameters that have been trained on the source unit's full-lifetime data and incorporate material constitutive knowledge. This can be used to efficiently inject general material physics priors into the corrector training process and reduce the risk of overfitting on small samples. In this embodiment, the pre-trained weights of the physical constraint network can be obtained by extracting encoder or shared layer weights from a validated physical constraint network model as the transfer starting point. For example, the pre-trained weights of the physical constraint network can work in conjunction with a sample correction damage set, where the former provides physical structural constraints and the latter provides target domain adaptation signals. In an exemplary embodiment, the pre-trained weights of the physical constraint network can include, but are not limited to, full network weights, feature extraction layer weights, and physical embedding layer weights. Using the sample creep coupling record set and the original sample damage set as inputs and the sample correction damage set as the supervision signal, the dual-channel data can be fed into the corrector for forward propagation, and the loss between the output and the sample correction damage set can be calculated for reverse updating. Furthermore, using the sample creep coupling record set and the original sample damage set as input, and the sample corrected damage set as the supervision signal, numerical bias can be measured by employing mean squared error loss or by introducing a physical residual term to enhance the consistency constraint of the mechanism. This can drive the model to learn an accurate mapping from mixed observations to pure fatigue features. The creep coupling corrector can be generated by transfer fine-tuning training with pre-trained weights of the physical constraint network. This can be achieved by loading the pre-trained weights to initialize the corrector parameters and performing a limited number of fine-tuning rounds on the sample set with a small learning rate. Furthermore, the creep coupling corrector can be generated by transfer fine-tuning training with pre-trained weights of the physical constraint network. This can be achieved by freezing the bottom physical embedding layer and only fine-tuning the top adaptation module, or by gradually releasing the parameters using a progressive unfreezing strategy. This allows the corrector to adapt to the target domain characteristics while preserving physical priors and generating a corrector with both generalization and accuracy.

[0101] For example, in the scenario of rapidly constructing a creep coupling corrector for newly commissioned gas turbine blades, the hot-end component life migration prediction method of the physical constraint network in this embodiment can be used when a new type of F-class gas turbine is put into operation and its own fault data is scarce. The system retrieves three years of peak-shaving operation data from the same type of source unit as a sample creep coupling record set and a sample original damage set. The materials expert team combines a high-precision constitutive model to annotate two hundred typical cycles to generate a sample corrected damage set. Then, the pre-trained weights of the physical constraint network that have been trained on the full life data of the source unit are loaded and its physical embedding layer is frozen. Only the top-level adaptation module is fine-tuned for fifty rounds. The creep coupling corrector generated in the end can accurately remove the residual creep component in the load reduction stage on the first month's operation data of the new unit, reducing the thermal fatigue damage extraction error by more than 60% compared with the de novo training model, effectively supporting the rapid deployment of life migration prediction in the early stage of service.

[0102] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a hot-end component lifetime migration prediction program for a physical constraint network, wherein when the hot-end component lifetime migration prediction program for a physical constraint network is executed by a processor, the program implements the steps of the hot-end component lifetime migration prediction method for a physical constraint network as described above.

[0103] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting the lifetime migration of hot-end components in a physical constraint network, characterized in that, The method includes: The full life cycle operation sequence of the target gas turbine power unit is obtained. The full life cycle operation sequence includes multiple peak-shaving operating condition samples, and each operating sample has a corresponding start-stop cycle time stamp. The original monitoring information of temperature load and mechanical stress of the first-stage rotor blade of the target gas turbine at the time stamp of each start-stop cycle is collected, as well as the thermomechanical fatigue coupling damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade. Based on the original monitoring information and the damage parameters and physical constitutive model parameters corresponding to the start-stop cycle timestamps of each cycle, the life migration prediction of the target cross-unit first-stage rotor blade is performed, and the remaining life output result of the thermomechanical fatigue coupling damage of the first-stage rotor blade is generated.

2. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 1, characterized in that, The acquisition of the full lifecycle operating sequence of the target gas turbine power unit includes multiple peak-shaving operating condition samples, each sample having a corresponding start-stop cycle timestamp, including: Obtain the unit identifier to which the target first-stage rotor blade belongs, and retrieve the corresponding full-life operation database and material performance parameter database based on the unit identifier; Receive the peak load range and prediction assessment time range set by the user, wherein the prediction assessment time range includes the start time of operation and the end time of prediction; Based on the peak load range, the start time, and the termination prediction time, multiple peak load operation samples are selected from the full life cycle operation database, and the start and stop cycle counts of each operation sample are extracted simultaneously as start and stop cycle timestamps for each operation sample.

3. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 1, characterized in that, The acquisition of raw monitoring information on temperature load and mechanical stress corresponding to the start-stop cycle timestamps of the target gas turbine first-stage rotor blade, as well as the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters of the material corresponding to the first-stage rotor blade, includes: The design geometric parameters and material crystal orientation type of the first-stage rotor blade of the target gas turbine are obtained. The design geometric parameters include the blade root coordinates and the distribution density of film gas holes. The damage neighborhood calculation window range is determined based on the air film pore distribution density and the material crystal orientation type. Based on the blade root coordinates, the original monitoring information of temperature load and mechanical stress corresponding to each start-stop cycle time stamp is obtained; The coupled damage calculation region is determined based on the blade design geometric parameters and the damage neighborhood calculation window range, and the thermomechanical fatigue coupled damage parameters and creep physical constitutive model parameters within the calculation region are obtained.

4. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 3, characterized in that, The step of determining the damage neighborhood calculation window range based on the air film pore distribution density and the material crystal orientation type includes: The equivalent spacing of the air film pores is calculated based on the air film pore distribution density, and the first window radius is obtained from the damage window mapping table based on the equivalent spacing. The second window radius is determined based on the crystal orientation type of the material, and the second window radius is a preset multiple of the influence range of material anisotropy; Compare the first window radius with the second window radius, and take the larger value as the base window radius; The damage neighborhood calculation window range is obtained by adding a fatigue damage margin to the basic window radius.

5. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 1, characterized in that, The method involves predicting the life migration of the target cross-unit first-stage rotor blade based on the original monitoring information corresponding to each start-stop cycle time stamp, the damage parameters, and the physical constitutive model parameters, generating the remaining life output result of the thermomechanical fatigue coupled damage of the first-stage rotor blade, including: Based on the thermomechanical fatigue coupled damage parameters, creep physical constitutive model parameters and the original monitoring information corresponding to each start-stop cycle time stamp, the actual cumulative damage component corresponding to each start-stop cycle time stamp is calculated. The initial prediction results of the migration model are corrected based on the actual cumulative damage components corresponding to each start-stop cycle time stamp, and the corrected remaining life prediction values ​​corresponding to each start-stop cycle time stamp are obtained, generating the remaining life output results of the thermomechanical fatigue coupling damage of the first-stage blade.

6. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 5, characterized in that, The step of correcting the initial prediction results of the migration model based on the actual cumulative damage component corresponding to each start-stop cycle time stamp to obtain the corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp includes: Traverse multiple peak-shaving operation samples in the entire lifecycle operation sequence to obtain the first set of peak-shaving operation samples; Extract the first start-stop cycle timestamp of the first group of peak-shaving operation samples, and obtain the first actual damage accumulation component based on the first start-stop cycle timestamp; Based on the first actual damage accumulation component, the first group of peak-shaving operation samples are divided into creep-dominated damage region and thermal fatigue-dominated damage region. For the first group of peak-shaving operation samples, the first damage component is extracted in the creep-dominated damage region, and the second damage component is extracted in the thermal fatigue-dominated damage region. The first damage component and the second damage component are coupled and synthesized to generate the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp. The corrected remaining lifetime prediction value corresponding to each start-stop cycle time stamp is obtained by using the same method as obtaining the corrected remaining lifetime prediction value corresponding to the first start-stop cycle time stamp.

7. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 6, characterized in that, The step of extracting a first damage component in the creep-dominated damage region and a second damage component in the thermal fatigue-dominated damage region for the first group of peak-shaving operation samples includes: Obtain the coupling damage separation threshold; In the creep-dominated damage region, the first group of peak-shaving operation samples are subjected to damage component separation processing based on the coupled damage separation threshold, and the cumulative damage introduced by Larson-Miller creep is separated and extracted to obtain the first damage component. In the thermal fatigue-dominated damage region, after creep coupling correction is performed on the first group of peak-shaving operation samples, damage component separation processing is performed in combination with the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

8. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 7, characterized in that, In the thermal fatigue-dominated damage region, after creep coupling correction is performed on the first group of peak-shaving operation samples, damage component separation processing is performed in conjunction with the coupling damage separation threshold, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component, including: In the first group of peak-shaving operation samples, the original monitoring values ​​and creep coupling components of each sampling point in the thermal fatigue-dominated damage area are obtained; The original monitoring value is replaced by the creep coupling correction value after creep coupling correction is obtained by performing creep coupling correction on each sampling point in the thermal fatigue-dominated damage area. Based on the corrected damage value and the coupled damage separation threshold, the first group of peak-shaving operation samples are subjected to damage component separation processing, and the cumulative damage introduced by thermal cycling mechanical fatigue is separated and extracted to obtain the second damage component.

9. The method for predicting the lifetime migration of hot-end components in a physical constraint network as described in claim 8, characterized in that, The construction steps of the creep coupling corrector include: Collect a sample creep coupling record set and a sample original damage set, and annotate based on the sample creep coupling record set and the sample original damage set to obtain a sample corrected damage set; Using the sample creep coupling record set and the sample original damage set as inputs, and the sample corrected damage set as a supervision signal, the creep coupling corrector is trained and generated by combining the pre-trained weights of the physical constraint network for transfer fine-tuning.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a hot-end component lifetime migration prediction program for a physical constraint network, which, when executed by a processor, implements the steps of the hot-end component lifetime migration prediction method for a physical constraint network as described in any one of claims 1 to 9.

Images