Online calibration method and system for heavy rotor interference hot charging model parameters based on assembly residual inversion, computer equipment and storage medium

Abstract

The application discloses a heavy rotor interference hot charging model parameter online calibration method and system based on assembly residual reverse, a computer device and a storage medium, relates to the field of heavy machinery manufacturing and intelligent assembly technology, and can continuously maintain model prediction fidelity without interrupting the production line. The method comprises the following steps: keeping the main network structure of the state perception model and the basic reasoning weight frozen, and taking the physical hidden variable as a set of to-be-calibrated parameters; after completing assembly and cooling locking, obtaining a macro assembly residual and inputting an edge computing node; when it is identified that the systemic drift trigger condition is met, activating a background reverse calibration program; calling a white-box physical agent operator based on bending moment integration to obtain the correction amount of the set of to-be-calibrated parameters relative to the current reference value; performing amplitude boundary constraint on the correction amount, and updating the correction amount subjected to the amplitude boundary constraint to the state perception model, so as to complete online calibration of the model parameters of the state perception model.

Description

Technical Field

[0001] This application relates to the fields of heavy machinery manufacturing and intelligent assembly technology, specifically to the field of heavy rotor thermal assembly state perception and online adaptive correction of model parameters, and particularly to a method, system, computer equipment and storage medium for online calibration of heavy rotor interference thermal assembly model parameters based on assembly residual inverse. Background Technology

[0002] The interference-fitting process for heavy-duty rotors is a crucial step in the assembly of core components for large rotating machinery such as steam turbines and compressors. It achieves a permanent connection between the rotor and the sleeve through an interference fit. In this process, the contact thermal resistance within the mating interface directly affects the precise transfer of thermal expansion and the temperature field distribution, while the interface friction coefficient determines the resistance characteristics during the pressing process and the final axial positioning accuracy. These two factors together constitute the core hidden variables affecting assembly quality. In recent years, digital twin technology and Physical-Informed Neural Networks (PINN), which integrate prior physical laws with data-driven modeling methods, have been introduced into this process. The aim is to build a high-fidelity predictive model through offline training, accurately simulating the expansion, pressing force, and final mating state during the heat fitting process before assembly, thereby guiding the setting of process parameters and assembly operations.

[0003] In related technologies, the common implementation method in industrial settings is to build and deploy a fixed set of assembly model parameters offline before production line commissioning or when processes change. This is based on limited batch testing or historical operating data, combined with finite element simulation and empirical correction. Once the model is in use, key parameters such as contact thermal resistance coefficient and friction coefficient are locked and no longer adjusted with the actual production process. When process conditions such as ambient temperature, initial workpiece temperature, and heating method experience seasonal fluctuations or batch-specific changes, engineers typically use periodic line shutdowns to sample a small number of rotors for process verification tests. The measured pressing force curve or final displacement value is compared with the model's predicted value. If the deviation exceeds the empirical threshold, experts manually correct the empirical coefficients in the model based on their judgment, or a small-scale parameter identification experiment is conducted again to update the calibration dataset. The corrected parameters are then redeployed into the model.

[0004] In the process of implementing the relevant technologies, it is believed that the technologies have at least the following technical problems: First, once deployed, model parameters remain fixed for a long period, making it impossible to detect changes in system state caused by gradual changes in environmental temperature or batch drift due to processing tolerances. This leads to a continuous accumulation of prediction residuals, making it impossible to determine when parameter correction is needed without interrupting production. Second, parameter calibration or model reconstruction relies on expert experience for feature selection, threshold setting, and correction strategy decisions. Human intervention is not only difficult to formalize into automated algorithms, but also, when the model has high dimensionality and complex parameter coupling, human experience cannot cover the nonlinear cases of multi-parameter joint drift, resulting in subjective and unstable correction results. Third, because heavy-duty rotors are single-piece, small-batch customized products, it is impossible to obtain massive amounts of labeled samples on-site for online retraining of conventional data-driven models. Traditional offline reconstruction methods require the complete re-collection of the working condition-response dataset, which does not meet real-time requirements and cannot be embedded into the closed-loop control logic of continuous production line operations. Fourth, the deviation between model predictions and physical reality can only be passively recorded and cannot be actively compensated, and there is no adaptive correction protection mechanism based on residual statistical characteristics, which restricts the prediction fidelity and process adaptability of the model during long-term service. Summary of the Invention

[0005] In view of this, this application provides an online calibration method, system, computer equipment and storage medium for heavy rotor interference hot fitting model parameters based on assembly residual inverse. The main purpose is to solve the problem that the current rotor interference hot fitting prediction model is subject to systematic deviation caused by environmental and batch tolerance drift, and cannot be rebuilt or retrained offline in single-piece small batch customization scenarios.

[0006] According to the first aspect of this application, an online calibration method for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals is provided. The method includes: A state-aware model for predicting the assembly state of heavy rotor interference hot fitting is deployed on an edge computing node. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are kept frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as the set of parameters to be calibrated. After the heavy rotor is assembled and cooled and locked at its interference hot fitting station, the actual macroscopic assembly residual of the heavy rotor is obtained, and the macroscopic assembly residual is input into the edge computing node according to the assembly sequence. In the edge computing node, the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process are monitored. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. In response to the activation of the reverse calibration procedure, a white-box physical proxy operator based on moment integral is pre-installed in the edge computing node. The macroscopic assembly residual is used as the boundary condition. The analytical gradient of the white-box physical proxy operator is used to propagate backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current reference value. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor residual bending moment. The correction amount is subjected to amplitude boundary constraints, and the correction amount after amplitude boundary constraints is updated to the state-aware model to complete the online calibration of the model parameters of the state-aware model.

[0007] According to a second aspect of this application, an online calibration system for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals is provided. The system includes: The state monitoring and locking module is used to deploy a state-aware model on an edge computing node to predict the assembly state of heavy rotor interference hot fitting. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are kept frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as a set of parameters to be calibrated. The residual acquisition module is used to acquire the actual macroscopic assembly residual of the heavy rotor after it has been assembled and cooled and locked at its interference hot assembly station, and to input the macroscopic assembly residual into the edge computing node according to the assembly sequence. The statistical triggering module is used in the edge computing node to monitor the statistical characteristics of the macroscopic assembly residual sequence that is continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. The reverse solution module is used to respond to the activation of the reverse calibration program by calling the white-box physical proxy operator based on moment integral pre-installed in the edge computing node. Using the macroscopic assembly residual as the boundary condition, the analytical gradient of the white-box physical proxy operator is used to propagate backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current reference value. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor residual bending moment. The parameter update module is used to apply amplitude boundary constraints to the correction amount and update the correction amount after amplitude boundary constraints to the state-aware model, so as to complete the online calibration of the model parameters of the state-aware model.

[0008] According to a third aspect of this application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] By utilizing the above technical solutions, this application provides an online calibration method, system, computer equipment, and storage medium for heavy-duty rotor interference hot-fitting model parameters based on assembly residual inverse. This application freezes the main network structure and basic inference weights of the state-aware model, using only the physical latent variables within the interface as the parameter set to be calibrated. This allows online updates to avoid the need for reconstruction or retraining of the entire model, thus avoiding the dependence of offline reconstruction on the number of samples and adapting to the small-batch production characteristics of heavy-duty rotors. Simultaneously, by real-time monitoring of the statistical characteristics of the continuously accumulated macroscopic assembly residual sequence according to the assembly timeline, it can achieve calibration without manual intervention. The system automatically identifies systematic drift and triggers reverse calibration, overcoming the subjectivity and lag of manual periodic line stop judgment. Furthermore, it calls a white-box physical surrogate operator based on moment integral and uses its analytical gradient to propagate back along the physical chain to solve for the correction amount, making the parameter update have clear physical causal interpretability and avoiding the limitations of manual experience in multi-parameter coupled nonlinear scenarios. In addition, it applies amplitude boundary constraints to the correction amount and then silently updates it to the state-aware model, which can not only achieve active compensation for prediction bias, but also ensure the convergence stability of the update process, thereby continuously maintaining the model prediction fidelity without interrupting the production line.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of an online calibration method for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, provided in an embodiment of this application. Figure 2This illustration shows a schematic diagram of a residual sequence statistical drift triggering mechanism provided in an embodiment of this application; Figure 3 This illustration shows a schematic diagram of the principle of a macroscopic residual to microscopic parameter inverse mapping provided in an embodiment of this application; Figure 4 This illustration shows a schematic diagram of the error variation curve for cross-specification migration and adaptive convergence provided in an embodiment of this application; Figure 5 This paper illustrates a schematic diagram of another online calibration method for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, provided in an embodiment of this application. Figure 6 This illustration shows a schematic diagram of an online calibration system for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, provided in an embodiment of this application. Figure 7 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides an online calibration method for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, such as... Figure 1 As shown, the method includes: S10: Deploy a state-aware model on edge computing nodes to predict the assembly state of heavy rotor interference thermal fitting. During the deployment and application of the state-aware model, keep the main network structure and basic inference weights of the state-aware model frozen, and use the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor as the set of parameters to be calibrated.

[0015] In this embodiment, a state-aware model for predicting the assembly state of heavy-duty rotor interference heat fitting needs to be deployed on an edge computing node. The edge computing node refers to an intelligent terminal device with computing and data storage capabilities deployed near the heavy-duty rotor heat fitting station, responsible for performing model inference and calibration calculations. The state-aware model is a thermo-mechanical neural operator pre-trained offline, used to simulate the temperature field evolution and mechanical response during rotor heat fitting.

[0016] In this embodiment, during the deployment and application of the state-aware model, its main network structure and basic inference weights need to be completely frozen. This means that any online adjustments to the connection weights between layers within the deep learning model are prohibited to ensure that the model's real-time inference response is not interfered with by online update operations. Simultaneously, the physical latent variables that cannot be directly measured within the heavy-duty rotor interface are selected as the parameter set to be calibrated. The interface refers to the annular contact surface between the outer cylindrical surface of the rotor shaft and the inner surface of the sleeve or impeller on which it is heated and fitted. This interface is enclosed within the metal structure after assembly, making it impossible to directly obtain its contact state information using conventional sensors. The physical latent variables specifically include the interface contact thermal resistance and the interface friction coefficient. The contact thermal resistance characterizes the degree of resistance to heat transfer at the interface, while the friction coefficient characterizes the frictional characteristics of the interface during the press-fitting process. Together, these constitute the core physical parameters affecting assembly quality.

[0017] In this way, by freezing the main network structure and locking only the physical latent variables of a limited dimension as the calibration object, the model parameter update does not require reconstruction or retraining of the entire network. This avoids the dependence of traditional offline reconstruction on massive sample data and adapts to the inherent constraint of sample scarcity in the scenario of customized production of single heavy rotors in small batches.

[0018] For example, in the thermal assembly wiring of a multi-stage rotor of a heavy-duty centrifugal compressor, a TMNO (Thermal-Mechanical Neural Operator) model that has been trained offline is preloaded into the edge computing node, all network layer weights are frozen, and the two physical latent variables, namely the interface contact thermal resistance and the interface friction coefficient, are set as calibration objects that can be adjusted online. In the subsequent operation of the model, only these two parameters are tracked and corrected.

[0019] S20: After the heavy rotor is assembled and cooled and locked at its interference hot assembly station, obtain the actual macroscopic assembly residual of the heavy rotor, and input the macroscopic assembly residual into the edge calculation node according to the assembly sequence.

[0020] In this embodiment, after the heavy rotor is assembled and cooled and locked at its interference hot fitting station, it is necessary to obtain the actual macroscopic assembly residual of the heavy rotor. The interference hot fitting station is a dedicated production station for performing hot fitting processes such as rotor induction heating, sleeve axial pressing, and furnace-cooled locking. After cooling and locking, the heavy rotor is removed from the assembly fixture and transferred to the metrology and inspection station. At this point, the mating interface has completely contracted and tightened, and the macroscopic form and position deviations caused by the asymmetric distribution of the microscopic contact pressure field are finally frozen. The macroscopic assembly residual refers to the total runout of the rotor after assembly, measured using a non-contact laser tracker or digital displacement sensor arranged around the metrology and inspection station. This total runout reflects the total radial offset of the actual rotation axis of the heavy rotor relative to the ideal rotation axis during one circumferential rotation, and is the final measured index characterizing the assembly accuracy at the macroscopic level. In this embodiment of the application, after obtaining the macro assembly residual, the macro assembly residual needs to be continuously input into the calibration interface of the edge computing node according to the assembly time sequence as the feedback input data for subsequent reverse calibration. Subsequently, by continuously obtaining the total runout after each assembly is completed, the actual fluctuation trend of the current assembly accuracy can be grasped.

[0021] For example, after each stage of rotor and sleeve is heat-fitted and cooled to room temperature, the outer circle of the rotor is circumferentially rotated to measure the total runout value, and this value is automatically sent to the edge calculation node according to the assembly sequence as the measured macroscopic assembly residual of the rotor.

[0022] S30: In the edge computing node, monitor the statistical characteristics of the macroscopic assembly residual sequence that is continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, activate the background reverse calibration program.

[0023] In this embodiment, the edge computing node needs to monitor the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. Specifically, to achieve this monitoring function, in this embodiment, the edge computing node maintains a residual sliding window with a fixed historical capacity in the background. As the production process progresses, the latest macroscopic assembly residuals are pushed into the window one by one, and the earliest data is removed, thereby forming a dynamic residual sequence that always reflects the recent assembly quality trend. Then, the moving mean and moving variance of the residual sequence within the sliding window are extracted as statistical features. The moving mean reflects the average deviation level of recent assembly accuracy, and the moving variance reflects the degree of consistency fluctuation of recent assembly results. At the same time, the mean drift tolerance and variance fluctuation limit are set as judgment thresholds. If the absolute value of the moving mean within the window does not exceed the tolerance and the moving variance does not exceed the fluctuation limit, the current assembly process is judged to be in a controlled steady state and calibration is not triggered. If either statistic exceeds the corresponding threshold, it is judged that the state of the mating interface has undergone systematic drift, that is, it is determined that the external environment or incoming material conditions have undergone substantial changes and have a detectable systematic impact on the assembly quality, and the background reverse calibration program is automatically activated.

[0024] In this way, through this statistical automatic triggering mechanism, the process changes can be continuously perceived and the calibration timing can be determined autonomously without human intervention. This can overcome the lag and subjectivity caused by relying on periodic line stop sampling and human experience judgment in traditional methods.

[0025] For example, the edge computing node maintenance includes a sliding window of the measured total runout values ​​of the most recent 50 rotors, which is updated one by one daily according to the production rhythm. When the mean of the runout sequence in the window deviates from the initial benchmark value by more than the set tolerance or the variance increases significantly, the background calibration program is automatically started without manual judgment and intervention.

[0026] S40: In response to the activation of the reverse calibration procedure, the white-box physical surrogate operator based on moment integral, which is pre-placed in the edge computing node, is invoked. The macroscopic assembly residual is used as the boundary condition. The analytical gradient of the white-box physical surrogate operator is used to propagate backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current benchmark value.

[0027] In this embodiment, in response to the activation of the reverse calibration procedure, a white-box physical surrogate operator based on moment integral, pre-placed in the edge computing node, is invoked. The white-box physical surrogate operator refers to a differentiable physical mathematical model with an explicit analytical expression. Its internal operational rules are constructed based on mechanical principles rather than fitting by a black-box neural network, thus possessing interpretability and differentiability. Furthermore, the white-box physical surrogate operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor's residual bending moment. Specifically, if the distribution of the contact pressure field along the circumferential and axial directions at the mating interface is asymmetrical after the heat-shrinking and locking process, a residual bending moment will be generated around the neutral axis of the rotor cross-section. This residual bending moment directly determines the magnitude of the rotor's rotational displacement tolerance. After calibration activation, this embodiment uses the measured macroscopic assembly residual, i.e., the total runout, as the boundary condition. The runout deviation observed at the macroscopic level is substituted into the above-mentioned differentiable analytical mapping. The analytical gradient of the white-box operator is used to propagate in a chain-like backward manner along the physical causal chain of total runout, residual bending moment, microscopic contact pressure field, and physical latent variables to be calibrated. The correction amount of the parameter set to be calibrated relative to the current reference value is obtained by reverse calculation from the macroscopic residual.

[0028] In the above process, since each link of the backpropagation path is an explicit physical expression, the gradient calculation is accurate and controllable, and there is no problem of unreliable gradients in black-box models. Therefore, the correction result has clear physical causal interpretability and can achieve accurate solution of the single correction amount without a large number of samples. This avoids the limitation of traditional methods that rely on human experience to repeatedly try to solve multi-parameter coupled nonlinear problems.

[0029] For example, when residual drift is detected, the white-box moment integral operator is activated, and the current measured total runout is input into the operator. The operator automatically calculates the correction values ​​of contact thermal resistance and friction coefficient in reverse along the physical gradient chain. Assuming that the contact thermal resistance needs to be increased by 5% and the friction coefficient needs to be decreased by 3%, the entire solution process is completed in a very short time without the need for manual analysis or experience judgment.

[0030] S50: Apply amplitude boundary constraints to the correction amount and update the correction amount after amplitude boundary constraints to the state-aware model to complete the online calibration of the model parameters of the state-aware model.

[0031] In this embodiment, the obtained correction amount needs to be subject to amplitude boundary constraints, and the correction amount after amplitude boundary constraints is updated to the state-aware model to complete the online calibration of the model parameters. Amplitude boundary constraints refer to setting an upper limit on the norm of a single correction amount, i.e., constraining it to not exceed a set percentage threshold of the current parameter baseline norm. If the norm of the solved correction amount falls within the constraint range, it is retained as a valid correction amount; if it exceeds the constraint range, the correction amount is truncated or discarded. This constraint mechanism prevents system oscillations or control divergence caused by excessively large correction amounts due to occasional abnormal data or numerical fluctuations. The correction amount after amplitude constraint filtering is updated to the corresponding adjustable parameters in the state-aware model by the parameter calibration interface in a silent overwrite manner. This process is executed automatically in the background, without affecting the normal operation rhythm of the production line or interrupting equipment operation. After the update is completed, the corrected parameters are used in the production inference of the next rotor.

[0032] In this way, through the amplitude constraint and silent update mechanism, the prediction bias can be actively compensated and the convergence stability of the update process can be guaranteed. This enables the model to maintain prediction fidelity without interrupting the production line, and ultimately forms a fully autonomous closed loop of residual monitoring, implicit parameter inversion, and online adaptive correction.

[0033] For example, assuming that the magnitude of a single correction is set not to exceed 10% of the current baseline value, if the calculated value of the contact thermal resistance correction is increased by 12%, it will be automatically truncated to an increase of 10% before the update is performed. The update process is completed silently in the background of the edge node, without the operator's awareness. When the next rotor is put into production, the corrected parameters will be used to predict the assembly status.

[0034] In this embodiment, optionally, a state-aware model for predicting the assembly state of heavy rotor interference thermal fitting is deployed on an edge computing node. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as a set of parameters to be calibrated. This includes: constructing a state-aware model with a two-layer architecture and deploying it on the edge computing node, wherein the two-layer architecture includes a deep learning model as a simulation baseline layer and a reality residual correction layer superimposed on the prediction output of the deep learning model. The deep learning model uses a pre-trained thermodynamic neural operator as the main network structure, and the inter-layer connection weights of the main network structure are frozen during deployment and application; utilizing the contact thermal resistance at the mating interface... coefficient of friction with interface Construct the set of parameters to be calibrated as shown in the following expression. : , in, Characterizes the contact thermal conductivity at the mating interface. Characterize the frictional properties at the mating interface.

[0035] In this embodiment, a state-aware model for predicting the assembly state of heavy-duty rotor interference heat fitting needs to be deployed on edge computing nodes. Edge computing nodes are intelligent terminal devices with computing and data storage capabilities deployed near the heavy-duty rotor heat fitting station, responsible for performing model inference and calibration calculations. In this embodiment, the state-aware model adopts a two-layer architecture, including a deep learning model as the simulation baseline layer and a reality residual correction layer superimposed on the predicted output of the deep learning model. The deep learning model uses a pre-trained offline thermodynamic neural operator as its main network structure. This thermodynamic neural operator is a deep learning model that integrates heat conduction and solid mechanics physical constraints to simulate the temperature field evolution and mechanical response during rotor heat fitting. The reality residual correction layer is a lightweight, adjustable parameter network layer superimposed on the predicted output of the main network structure, used to perform minor compensation correction on the baseline prediction results.

[0036] During the deployment and application of the state-aware model, it is necessary to keep the inter-layer connection weights of its main network structure completely frozen, that is, prohibit any online adjustment of the connection weights between the layers within the deep learning model, to ensure that the real-time response capability of the model inference is not interfered with by online update operations. Simultaneously, the contact thermal resistance and interface friction coefficient within the mating interface of the heavy-duty rotor, two physical latent variables that cannot be directly measured, are selected as the parameter set to be calibrated. The mating interface refers to the annular contact surface between the outer cylindrical surface of the rotor spindle and the inner surface of the sleeve or impeller on which it is heated and fitted. This interface is enclosed within the metal structure after assembly, and its contact state information cannot be directly obtained through conventional sensors. Contact thermal resistance characterizes the degree of resistance to heat transfer at the mating interface, and the friction coefficient characterizes the frictional characteristics of the mating interface during the press-fitting process. Together, they constitute the core physical parameters affecting assembly quality. In this embodiment, the contact thermal resistance and interface friction coefficient at the mating interface are locked as the physical latent variables to be calibrated, constructing the parameter set to be calibrated as shown in the expression in Formula 1 below: Formula 1: , in, Contact thermal resistance characterizes the contact thermal conductivity at the mating interface. is the interfacial friction coefficient, which characterizes the frictional properties at the mating interface. In this embodiment, the inverse solution and parameter update operations only apply to the adjustable parameters in the actual residual correction layer corresponding to the parameter set to be calibrated, while the main network structure remains unchanged.

[0037] In this way, by freezing the main network structure and using only the physical latent variables of a limited dimension as the calibration objects, online parameter updates do not require reconstruction or retraining of the entire model. This avoids the dependence on massive sample data for traditional offline reconstruction and adapts to the inherent constraints of sample scarcity in the scenario of customized production of single heavy rotors in small batches. At the same time, since the weights of the main network are frozen, the computational load of the model inference process does not increase due to online updates, which can ensure the real-time inference response capability of edge computing nodes.

[0038] For example, during the deployment phase of the thermal assembly wiring of a multi-stage rotor of a heavy-duty centrifugal compressor, a TMNO model that has been trained offline is preloaded into the edge computing node, all its network layer weights are frozen, and the interface contact thermal resistance and interface friction coefficient are set as calibration objects that can be adjusted online. At the same time, the state-aware model is constructed as a two-layer structure with a frozen baseline layer and an adjustable correction layer. In subsequent operation, only the adjustable parameters corresponding to these two physical latent variables in the correction layer are tracked and corrected, while the main network remains frozen.

[0039] Optionally, in this embodiment, after the heavy rotor is assembled and cooled and locked at its interference hot fitting station, the actual macroscopic assembly residual of the heavy rotor is obtained, and the macroscopic assembly residual is input into the edge computing node according to the assembly sequence. This includes: using a measuring device arranged at the metrology and inspection station to measure the form and position tolerances of the assembled heavy rotor to collect the total runout of the heavy rotor after assembly. Total bounce As a macroscopic assembly residual, the heavy rotor undergoes induction heating, axial pressing, and cooling locking stages sequentially at the interference hot assembly station, causing it to detach from the assembly fixture and transfer to the metrology and inspection station; the total runout is... Characterizes the total radial offset of the actual axis of rotation of a heavy-duty rotor relative to the ideal axis of rotation during one circumferential rotation; and includes macroscopic assembly residuals. The calibration interface of the edge computing node is input according to the assembly sequence as feedback input for subsequent reverse calibration.

[0040] In this embodiment, after the heavy rotor is assembled and cooled and locked at its interference hot fitting station, the actual macroscopic assembly residual of the heavy rotor is obtained. The interference hot fitting station is a dedicated production station for performing hot fitting processes such as rotor induction heating, sleeve axial pressing, and furnace-cooled locking. At this station, the heavy rotor sequentially undergoes three stages: induction heating to expand the inner hole of the sleeve, axial pressing to bring the sleeve and rotor spindle to a predetermined axial position, and complete cooling and locking to allow the sleeve to contract and clamp the rotor spindle. After cooling and locking, the macroscopic form and position deviations caused by the asymmetric distribution of the microscopic contact pressure field at the mating interface are ultimately frozen inside the workpiece. Subsequently, the heavy rotor is removed from the assembly fixture and transferred to the metrology and inspection station. Using a high-precision non-contact laser tracker or digital displacement sensor located at the metrology and inspection station, the form and position tolerances of the assembled heavy rotor are measured, and the final macroscopic assembly residual of the assembled workpiece is collected. The macroscopic assembly residual is the total runout after assembly. Among them, the total bounce The total radial deviation of the actual rotation axis of a heavy-duty rotor relative to the ideal rotation axis during one circumferential rotation is a macroscopically measured indicator that comprehensively reflects assembly accuracy. This application embodiment will use the collected total runout... As macroscopic assembly residuals, they are input into edge computing nodes through the calibration interface according to the assembly sequence, serving as feedback input data for subsequent reverse calibration.

[0041] In this way, by continuously acquiring the total runout of each assembled part and inputting it into the edge computing node in a time sequence during the above process, the actual fluctuation trend of the current assembly accuracy can be grasped, providing a reliable measured data basis for subsequent statistical drift monitoring and reverse calibration.

[0042] For example, in the thermal assembly line of a multi-stage rotor of a heavy-duty centrifugal compressor, after each stage rotor and sleeve are induction heated, axially pressed in, and naturally cooled to room temperature in the air, multiple laser trackers arranged around the metering and testing station are used to perform circumferential rotation scanning measurement on the outer surface of the rotor, read the total runout value, and automatically send the value to the edge computing node according to the order of assembly completion as the actual measured macroscopic assembly residual corresponding to the rotor.

[0043] Optionally, in this embodiment of the application, the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process are monitored in the edge computing node. When the statistical characteristics are found to meet the preset systematic drift triggering condition, the background reverse calibration program is activated, including: maintaining a fixed historical capacity based on the edge computing node. A residual sliding window is used to push the currently acquired macroscopic assembly residuals into the sliding window piece by piece as production progresses, and to remove the earliest data in the window; the moving average of the macroscopic assembly residual sequence within the sliding window is calculated. With moving variance Statistical characteristics of the macroscopic assembly residual sequence; determining the mean drift tolerance indicated by the preset systematic drift triggering conditions. and variance fluctuation limit , The maximum allowable offset of the moving average. Maximum allowable fluctuation in moving variance; continuous monitoring of moving average. Is the absolute value greater than the mean drift tolerance? Moving variance Is it greater than the variance fluctuation limit? When the moving average is detected The absolute value is greater than the mean drift tolerance. or moving variance Greater than the variance fluctuation limit At that time, the statistical characteristics are determined to meet the systematic drift triggering conditions, and the background inverse calibration program is activated; correspondingly, the method also includes: when a moving mean is detected... The absolute value is less than or equal to the mean drift tolerance. and moving variance Less than or equal to the variance fluctuation limit When the system is in a controlled steady state, calibration is not triggered, and the input of the next macroscopic assembly residual is monitored.

[0044] In this embodiment, the edge computing node needs to monitor the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. The edge computing node maintains a database with a fixed historical capacity in the background. The residual sliding window is essentially a first-in, first-out (FIFO) data buffer. As production progresses, the currently acquired macroscopic assembly residuals are pushed to the end of the window piece by piece, while the oldest data at the beginning of the window is removed, ensuring that the window always contains the most recent data. The measured residual values ​​of the rotor are then determined. Next, the moving average of the macroscopic assembly residual sequence within this sliding window is extracted. With moving variance As statistical characteristics, the moving mean reflects the average deviation level of recent assembly accuracy, while the moving variance reflects the consistency and fluctuation degree among recent rotor assembly results. Furthermore, this embodiment also presets a systematic drift trigger condition, which indicates a mean drift tolerance. and variance fluctuation limit , The maximum allowable offset of the moving average. This represents the maximum allowable fluctuation in moving variance.

[0045] Subsequently, drift characteristic checks are continuously performed in the background. If the absolute value of the moving mean is detected to be greater than the mean drift tolerance, the drift characteristic will be checked. Or the moving variance is greater than the variance fluctuation limit. If a systematic drift occurs at the interface, it is determined that changes in external environmental temperature or batch fluctuations in incoming material tolerances have had a detectable systematic impact on assembly quality, and the background reverse calibration program is automatically activated. Correspondingly, if the absolute value of the moving average is detected to be less than or equal to the mean drift tolerance... And the moving variance is less than or equal to the variance fluctuation limit. If the current assembly process is in a controlled steady state, calibration is not triggered, and the edge computing node continues to monitor the input of the next macroscopic assembly residual.

[0046] In this way, through the statistical automatic monitoring and triggering mechanism, the changing trend of the production process can be continuously perceived and the calibration timing can be determined autonomously without human intervention. This can overcome the lag and subjectivity caused by the reliance on periodic line stop sampling and human experience judgment in traditional methods. At the same time, it can avoid high-frequency oscillation updates of the model caused by occasional random measurement noise of a single workpiece, ensuring the reliability and stability of calibration triggering.

[0047] For example, in the thermal assembly line of a multi-stage rotor of a heavy-duty centrifugal compressor, the edge computing node maintains a sliding window containing the measured values ​​of the total runout of the most recent 50 rotors. The window is updated one by one according to the production rhythm every day. Assuming that the mean drift tolerance is set to 0.02mm and the variance fluctuation limit is 0.0004mm², when the mean of the runout sequence in the window deviates more than 0.02mm or the variance exceeds 0.0004mm² from the initial calibration reference value, the background calibration program is automatically started without manual judgment and intervention.

[0048] Optionally, in this embodiment, the moving average of the macroscopic assembly residual sequence within the sliding window is calculated. With moving variance Statistical characteristics of the macroscopic assembly residual sequence include: calculating the moving mean of the macroscopic assembly residual sequence within the sliding window using the following formula. : , in, Represents the first in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value Given the historical capacity of the sliding window; the moving variance of the macroscopic assembly residual sequence within the sliding window is calculated using the following formula. : , in, Represents the first in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value For the historical capacity of the sliding window, Represents the moving average; [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] With moving variance As a statistical characteristic.

[0049] In this embodiment, a database with a fixed historical capacity is maintained in the background of the edge computing node. The residual sliding window is a first-in-first-out (FIFO) data buffer used to store the measured macroscopic assembly residuals of several rotors assembled recently. As production progresses, the currently acquired macroscopic assembly residuals are pushed to the end of the window one by one, and the earliest data at the beginning of the window is removed, ensuring that the window always contains the most recent data. The measured residual values ​​of the rotor were then calculated. Subsequently, the moving average of the macroscopic assembly residual sequence within the sliding window was calculated. With moving variance As a statistical characteristic, the moving mean is one of them. The calculation formula is shown in Formula 2 below: Formula 2: , in, This represents the first position in the macroscopic assembly residual sequence within the sliding window, arranged according to assembly time. The first macroscopic assembly residual value, i.e., the first... The measured total runout of the rotor. This represents the historical capacity of the sliding window. The moving average is calculated using Formula 2. This reflects the average deviation level of recent assembly accuracy.

[0050] Furthermore, moving variance The calculation formula is shown in Formula 3 below: Formula 3: , In Formula 3, the meanings of the parameters are the same as those in Formula 2. This is the moving mean. The moving variance is calculated using Formula 3. This reflects the consistency and degree of fluctuation among the recent rotor assembly results.

[0051] Calculate the moving average With moving variance Subsequently, the embodiments of this application will use the moving average. With moving variance As a statistical feature, combined with preset systematic drift triggering conditions, it is determined whether the background reverse calibration program needs to be activated.

[0052] Thus, in the above process, the statistical calculation based on the sliding window can effectively filter out the interference of occasional random measurement noise of a single workpiece, avoid the model from oscillating and updating due to noise, and ensure the reliability and stability of calibration triggering.

[0053] In summary, the residual sequence statistical drift triggering mechanism in the embodiments of this application is as follows: Figure 2 As shown, the total runout will be used as the macroscopic assembly residual. Record data and set up historical capacity. A residual sliding window of 50 (Window Length N=50) is used, and the window is slid one item at a time. Statistical features are first extracted, and the moving mean is calculated using formula 2 above. Simultaneously, the moving variance is calculated using Formula 3 above. Furthermore, the core decision-making logic of this mechanism is to determine whether a systematic drift has occurred in the current assembly process, that is... Figure 2 The "Indicating Systemic Drift?" problem shown is determined by examining whether the statistical characteristics of the residual sequence within the sliding window indicate a statistical measure that exceeds a set tolerance band. Specifically, this triggering mechanism sets explicit threshold conditions, including mean drift tolerance. With variance fluctuation limit ,in The maximum allowable offset of the moving average. This represents the maximum permissible fluctuation in moving variance. During ongoing background drift characteristic checks, once the absolute value of the moving mean of the residual sequence within the window is detected to exceed the mean drift tolerance, [the drift is considered acceptable]. Or the moving variance is greater than the variance fluctuation limit. This means that if the current statistical value exceeds the set tolerance zone, it indicates that the incoming material tolerance batch or the on-site temperature environment has undergone substantial changes and has had a detectable systematic impact on assembly quality. Based on this, a systematic drift is determined to have occurred. And based on the deployment-state automatic triggering mechanism, the background reverse calibration program is automatically activated, locking the current window data and sending it to the white-box physical proxy operator to calculate the correction amount. Correspondingly, only when the absolute value of the moving average is less than or equal to the mean drift tolerance and the moving variance is less than or equal to the variance fluctuation limit, i.e. and Only when the state is in a controlled steady state is the calibration not triggered and monitoring continues.

[0054] In this embodiment of the application, optionally, a white-box physical surrogate operator based on moment integral, pre-placed in the edge computing node, is invoked. Using the macroscopic assembly residual as the boundary condition, the analytical gradient of the white-box physical surrogate operator is propagated backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current benchmark value. This includes: determining the white-box physical surrogate operator based on moment integral, wherein the differentiable analytical mapping established by the white-box physical surrogate operator is as shown in the following expression: , in, Indicates the heavy rotor at time The residual bending moment, This represents the integral operation over the axial mating interface of a heavy-duty rotor. Indicates time The lower-corrected interfacial micro-contact pressure field in cylindrical coordinates Spatial distribution values ​​at that location Represents circumferential coordinates. Indicates the axial coordinate. This indicates the area of ​​the infinitesimal element on the interface. The vertical lever arm represents the spatial coordinates of the infinitesimal element area relative to the neutral axis of the heavy-duty rotor cross-section; where the vertical lever arm is [value missing] when the infinitesimal element area is located above the neutral axis. Taking a positive value, when the area of ​​the infinitesimal element is below the neutral axis, the vertical lever arm... Take the negative value; treat the macroscopic assembly residual as... The boundary conditions are input to a differentiable analytical mapping to construct a loss function aimed at reducing prediction jitter error. According to the following expression, the analytical gradient of the differentiable analytical mapping is backpropagated along the physical chain to inversely deduce the microscopic interface pressure field deviation that causes macroscopic assembly residual anomalies. Based on the deviation of the microscopic interface pressure field The physical correlation between the parameter set to be calibrated and the actual parameter set is used to solve for the correction amount. : , in, Indicates the set of parameters to be calibrated and , This indicates the contact thermal resistance at the interface. This represents the coefficient of interfacial friction at the mating interface. Represents the loss function Regarding the parameter set to be calibrated The gradient; Indicates macroscopic assembly residuals; Represents the residual bending moment of a heavy-duty rotor; This represents the corrected interfacial microscopic contact pressure field; where, , This indicates the correction amount for contact thermal resistance. This represents the correction amount for the interfacial friction coefficient.

[0055] In this embodiment, in response to the activation of the reverse calibration procedure, a white-box physical proxy operator based on moment integral, pre-placed in the edge computing node, needs to be invoked. The white-box physical proxy operator refers to a differentiable physical mathematical model with an explicit analytical expression, constructed entirely based on mechanical principles. Its internal operational rules are transparent and interpretable, and it supports automatic differentiation. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor's residual bending moment. Due to the asymmetry in the circumferential and axial distribution of the contact pressure field at the mating interface during the hot-fitting process of a heavy-duty rotor, a residual bending moment will be generated around the neutral axis of the rotor cross-section after cooling and locking. This residual bending moment directly determines the magnitude of the rotor's rotational displacement tolerance. Specifically, the differentiable analytical mapping established by the white-box physical proxy operator is shown in Equation 4 below: Formula 4: , in, Indicates the heavy rotor at time The residual bending moment, i.e. the resultant bending moment generated by the asymmetric contact pressure field at the mating interface on the rotor cross section; This represents the area integral operation over the axial mating interface of a heavy-duty rotor; Indicates time Corrected interfacial micro-contact pressure field in cylindrical coordinates Spatial distribution values ​​at, where Circumferential coordinates For axial coordinates; This represents the area of ​​a micro-element on the interface. This represents the vertical component of the spatial coordinates of the micro-element area relative to the neutral axis (horizontal centerline) of the heavy-duty rotor cross-section, also known as the vertical lever arm. This component is positive when the micro-element area is above the neutral axis and negative when it is below. In practical applications, the principle of the inverse mapping from macroscopic residuals to microscopic parameters in this embodiment reveals the physical causal relationship between the final macroscopic runout after the heavy-duty rotor is thermally assembled and the spatial distribution of the microscopic contact pressure field at the mating interface. Specifically, as follows... Figure 3 As shown, the principle focuses on the analysis of the micro-element level on the mating interface, specifically taking the area of ​​a micro-element on the mating interface. Local contact pressure acts on the area of ​​this micro-element. ,in This represents the circumferential coordinates of the infinitesimal element. Indicates its axial coordinate. This represents the local contact pressure value at that spatial location. This local contact pressure generates a torque effect on the neutral axis (horizontal centerline) of the rotor cross-section, and the magnitude of this torque is equal to the local contact pressure. Multiply by the area of ​​the infinitesimal element Then multiply by the vertical force arm of the infinitesimal element to the neutral axis. ,in This represents the vertical component of the spatial coordinates of the infinitesimal element area relative to the neutral axis. The lever arm is positive when the infinitesimal element is above the neutral axis and negative when it is below. Integrating and summing the moments across all infinitesimal areas on the entire mating interface yields the residual bending moment acting on the rotor. This residual bending moment directly determines the final macroscopic runout of the rotor after cooling and locking. Through this physical mapping relationship, the actual measured macroscopic assembly residual... When deviations exceeding the tolerance occur, a white-box physical surrogate operator based on bending moment integrals can be used to reverse-engineer the microscopic interface pressure field deviations that cause macroscopic assembly residual anomalies using the aforementioned formula 4. Then, the correction amount of the physical parameter to be calibrated that causes the deviation of the pressure field can be solved, that is, the correction amount of the interfacial friction coefficient. and the correction amount for contact thermal resistance .

[0056] After the calibration procedure is activated, the embodiments of this application will use the measured macroscopic assembly residuals. The above differentiable analytical mapping is used as a boundary condition input, and a loss function is constructed with the goal of reducing the prediction sway error. Subsequently, the analytical gradient of this white-box physical surrogate operator is backpropagated along the complete physical causal chain, which is the macroscopic assembly residual. Residual bending moment Microscopic contact pressure field Parameter set to be calibrated The calculation expression for backpropagation is shown in Formula 5 below: Formula 5: , in, Indicates the set of parameters to be calibrated and , This indicates the contact thermal resistance at the interface. This represents the coefficient of interfacial friction at the mating interface. Represents the loss function Regarding the parameter set to be calibrated The gradient of the loss function is the direction of optimization that makes the loss function decrease the fastest. This represents the measured macroscopic assembly residual, i.e., the total runout. Represents the residual bending moment of a heavy-duty rotor; This represents the corrected interfacial micro-contact pressure field.

[0057] Thus, through the back propagation of this analytical derivative chain, the deviation of the microscopic interfacial pressure field that causes the overall bending moment anomaly can be precisely derived. Furthermore, based on the deviation of the microscopic interface pressure field... With the parameter set to be calibrated The physical relationship between them allows for the unique solution of the correction amount for the set of parameters to be calibrated. ,in This indicates the correction amount for contact thermal resistance. This represents the correction amount for the interfacial friction coefficient.

[0058] In the above process, since each link in the backpropagation path is an explicit physical expression, the gradient calculation is accurate and controllable. There is no problem of unreliable gradients or gradient vanishing in black-box models. Therefore, the correction result has clear physical causal interpretability and can achieve accurate solution of single correction amount without a large number of samples. This avoids the limitation of traditional methods that rely on human experience to repeatedly try to solve multi-parameter coupled nonlinear problems.

[0059] For example, in the thermal assembly wiring of a multi-stage rotor of a heavy-duty centrifugal compressor, when residual drift is detected, the white-box bending moment integral operator is activated. The current measured total runout is input into the differentiable analytical mapping shown in Formula 4 above. The operator automatically calculates the correction values ​​of contact thermal resistance and friction coefficient in reverse along the physical gradient chain. For example, it is calculated that the contact thermal resistance needs to be increased by 5% and the friction coefficient needs to be decreased by 3%, without the need for manual analysis or experience judgment.

[0060] Optionally, in this embodiment, the correction amount is subjected to amplitude boundary constraints, and the correction amount after amplitude boundary constraints is updated to the state-aware model to complete the online calibration of the model parameters of the state-aware model. This includes: setting the amplitude boundary constraints shown in the following expression, and using the amplitude boundary constraints to perform amplitude verification on the correction amount, wherein the amplitude boundary constraints are used to limit the norm of a single correction amount from not exceeding a preset percentage threshold of the norm of the parameter reference values ​​stored in the current state-aware model. , in, This represents the correction amount for the set of parameters to be calibrated. , This indicates the correction amount for contact thermal resistance. This indicates the correction amount for the interfacial friction coefficient; Indicates the baseline value of the parameter. , This indicates the current reference value for contact thermal resistance. This represents the current reference value for the interfacial friction coefficient; Norm operators; This represents a preset percentage threshold. If the norm of the correction quantity satisfies the amplitude boundary constraint, the correction quantity is taken as the effective correction quantity, and the effective correction quantity is silently overwritten and updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated, thus completing the online calibration of the model parameters of the state-aware model. If the norm of the correction quantity does not satisfy the amplitude boundary constraint, the correction quantity is truncated, so that the norm of the truncated correction quantity is equal to... The truncated correction amount is then silently overwritten and updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated, thus completing the online calibration of the model parameters of the state-aware model.

[0061] In this embodiment, after obtaining the correction amount for the parameter set to be calibrated, it is necessary to impose amplitude boundary constraints on the correction amount to prevent system oscillation or control divergence caused by excessively large single correction amplitude due to occasional abnormal data or numerical fluctuations. Specifically, the amplitude boundary constraints are set by limiting the single parameter correction amount. The norm does not exceed the parameter baseline value stored in the current state-aware model. Preset percentage threshold of the norm The expression is shown in Formula 6 below: Formula 6: ; in, This represents the correction amount for the set of parameters to be calibrated. , This indicates the correction amount for contact thermal resistance. This indicates the correction amount for the interfacial friction coefficient; This represents the baseline values ​​of the parameters stored in the current state-aware model. , This indicates the currently stored reference value for contact thermal resistance. This represents the current stored reference value for the interface friction coefficient; The norm operator is used to measure the overall magnitude of the correction or baseline value; The preset percentage threshold ranges from 0 to 1. The specific value can be preset according to the actual needs of the project. For example, setting it to 0.10 means that the single correction amount does not exceed 10% of the baseline value amplitude.

[0062] Furthermore, the magnitude of the obtained correction is checked using the aforementioned magnitude boundary constraints. If the norm of the correction satisfies the aforementioned inequality, the correction is retained as a valid correction. If the norm of the correction does not satisfy the aforementioned inequality, the magnitude of the correction is truncated so that the norm of the truncated correction equals the magnitude of the correction. This means that the values ​​are compressed proportionally to the constraint boundaries. Subsequently, the correction amount after amplitude boundary constraints is updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated in a silent overlay manner, thus completing the online calibration of the model parameters of the state-aware model.

[0063] It should be noted that the above update process is executed automatically in the background without interrupting the normal inference service of the edge computing nodes. The operators are unaware of it. The updated parameters are directly used in the subsequent rotor assembly status prediction. Thus, by introducing amplitude boundary constraints, it is possible to actively compensate for prediction deviations and effectively avoid system oscillations or control divergence that may be caused by excessive single correction amplitude, thus ensuring the safety and convergence stability of the online adaptive calibration process.

[0064] For example, in the thermal assembly line of a multi-stage rotor of a heavy-duty centrifugal compressor, the amplitude of a single correction is set to not exceed 10% of the current reference value amplitude. If the norm of the calculated value of the contact thermal resistance correction is 12% of the reference value amplitude, it will be automatically truncated to 10% before the update is performed. The update process is completed silently in the background of the edge node. When the next rotor is put into production, the corrected parameters will be used to predict the assembly status.

[0065] Furthermore, in the scenario of single-piece, small-batch, multi-specification customized production of heavy-duty rotor interference heat fitting, the production line frequently faces the switch between different rotor specifications. When the production line switches from a historical specification A rotor to a completely new specification B rotor, due to the significant abrupt change in geometric dimensions and heat fitting interference tolerances, there are significant differences between the model parameters obtained from the original calibration for specification A and the physical characteristics of specification B. If the original parameters are directly used, the sensing system will face a severe shock at the specification switch point, and the prediction error will increase sharply. To address this engineering challenge of cross-specification switch, such as... Figure 4 As shown, this embodiment of the application does not adopt a cold start method with zero foundation at the moment of switching, that is, it does not reset the parameters to be calibrated to the initial default value, but automatically extracts the calibrated parameters of the historical specification A that is closest to the new specification B in terms of structural dimensions or materials. As the initial warm-start values ​​for the new model, that is Figure 4The parameter hotstart shown inherits parameters from neighboring specifications, avoiding large errors during cold starts. Hotstart refers to using converged parameters from existing similar specifications as initial values ​​for the new specification model, ensuring a reasonable parameter baseline from the outset. This keeps prediction errors during the initial specification switch within a reasonable range, avoiding the initial large error shock caused by cold starts. As the new specification B enters the formal assembly line and begins continuous production, the sliding window maintained in the edge computing node continuously receives the measured macroscopic assembly residuals of each rotor of the new specification B. When the statistical characteristics of the residual sequence within the window exceed the preset systematic drift trigger condition, the background reverse calibration program is automatically activated. Subsequently, based on the feedback from the measured residuals, a white-box physical proxy operator based on bending moment integral is invoked. The analytical gradient is used to backpropagate along the physical chain to perform inverse inversion solutions for interface contact thermal resistance and friction coefficient, achieving rapid optimization of interface hidden variables. Silent adaptive updates are then performed in the background. The entire process is executed automatically in the background without interrupting normal production line operations, and is imperceptible to operators, thus completing the cross-specification parameter migration. Thus, by using a very small number of samples, such as less than 30 samples, for cross-specification parameter adaptive migration, the model's prediction accuracy for the assembly runout of the new specification B rotor can be improved by approximately 32%. Therefore, through the aforementioned parameter hot-start and background silent adaptive update mechanism, this embodiment of the application can achieve rapid parameter migration and accuracy convergence in cross-specification scenarios with small samples without interrupting the production line or accumulating massive amounts of samples, significantly improving the model's adaptability and assembly quality assurance level in multi-specification customized production scenarios.

[0066] The technical solution in this application embodiment can be applied to a calibration system. The calibration system includes an initialization layer, an online monitoring and triggering layer, and a background reverse solving and updating layer. Based on this calibration system, the logical process of the technical solution in this application embodiment is summarized as follows: like Figure 5As shown, the process first executes step S1, "State Monitoring and Parameter Locking," deploying a physical information neural network as a state-aware model on the edge computing node, freezing its main network structure, and selecting the contact thermal resistance and interface friction coefficient, which cannot be directly measured at the interface, as the parameter set to be calibrated. Next, step S2, "Macroscopic Residual Acquisition," is executed. After the heavy rotor completes heat fitting and cooling locking, its actual total runout is acquired through offline measurement equipment as the macroscopic assembly residual. Then, step S3, "Update Residual Sliding Window Statistics," is executed. The currently measured residual is pushed into the residual sliding window maintained by the edge computing node, and the moving mean and moving variance of the recent production batch residual sequence are calculated. The statistics within the window are compared with the preset tolerance to determine whether the statistical drift characteristics exceed the set tolerance. If not, calibration is not triggered and the monitoring of the next workpiece continues; if it exceeds the tolerance, the background reverse calibration program is automatically activated. After activating the background reverse calibration program, step S4, "Macroscopic to Microscopic Reverse Solving," is executed. The white-box physical proxy formula based on bending moment integral, namely Formula 4 above, is introduced. The measured macroscopic assembly residual is used as the boundary condition. By backpropagating the analytical gradient along the complete physical chain, the deviation of the microscopic interface pressure field that causes the bending moment anomaly is deduced. Then, the correction amount of the physical latent variable to be calibrated is obtained, and it is determined whether the correction amount meets the amplitude boundary constraint, such as not exceeding 10% of the current benchmark amplitude. If it meets the constraint, the correction amount is retained; if it does not meet the constraint, amplitude truncation or abandonment of this correction is performed. Finally, the S5 parameter direct update and correction step is executed. The correction amount after amplitude boundary constraint is updated to the residual correction network of the state-aware model in a silent overlay manner, completing the online adaptive closed loop of the model parameters. The updated parameters are directly put into subsequent production inference. The whole process covers the complete closed loop link from state monitoring, residual acquisition, statistical triggering, physical inversion to parameter update, realizing continuous online adaptive correction without reconstruction after model deployment.

[0067] The method provided in this application freezes the main network structure and basic inference weights of the state-aware model, using only the physical latent variables within the interface as the parameter set to be calibrated. This eliminates the need for online updates to reconstruct or retrain the entire model, thus avoiding the dependence of offline reconstruction on the number of samples and adapting to the production characteristics of single-piece, small-batch production of heavy-duty rotors. Simultaneously, by real-time monitoring of the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence, it can automatically identify systematic drift and trigger reverse calibration without manual intervention, overcoming the subjectivity and lag of manual periodic line stop judgment. Furthermore, by calling a white-box physical surrogate operator based on bending moment integral and using its analytical gradient to propagate back along the physical chain to solve for the correction amount, the parameter update has clear physical causal interpretability, avoiding the limitations of manual experience in multi-parameter coupled nonlinear scenarios. In addition, by applying amplitude boundary constraints to the correction amount before silently updating it to the state-aware model, it can achieve active compensation for prediction bias and ensure the convergence stability of the update process, thereby continuously maintaining the model prediction fidelity without interrupting the production line.

[0068] Furthermore, as Figure 1 In a specific implementation of the method, this application provides an online calibration system for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, such as... Figure 6 As shown, the system includes: a status monitoring and locking module 601, a residual acquisition module 602, a statistical triggering module 603, a reverse solution module 604, and a parameter update module 605.

[0069] The state monitoring and locking module 601 is used to deploy a state-aware model on an edge computing node for predicting the assembly state of heavy rotor interference hot fitting. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are kept frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as a set of parameters to be calibrated. The residual acquisition module 602 is used to acquire the actual macroscopic assembly residual of the heavy rotor after the heavy rotor is assembled and cooled and locked at its interference hot assembly station, and input the macroscopic assembly residual into the edge computing node according to the assembly sequence. The statistical triggering module 603 is used in the edge computing node to monitor the statistical characteristics of the macroscopic assembly residual sequence that is continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. The reverse solution module 604 is used to respond to the activation of the reverse calibration program by calling the white-box physical proxy operator based on moment integral pre-installed in the edge computing node, using the macroscopic assembly residual as the boundary condition, and using the analytical gradient of the white-box physical proxy operator to propagate backward along the physical chain of differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current reference value. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor residual bending moment. The parameter update module 605 is used to apply amplitude boundary constraints to the correction amount and update the correction amount after amplitude boundary constraints to the state-aware model, so as to complete the online calibration of the model parameters of the state-aware model.

[0070] In a specific application scenario, the state monitoring and locking module 601 is used to construct the state-aware model with a two-layer architecture and deploy it on the edge computing node. The two-layer architecture includes a deep learning model as a simulation baseline layer and a real-world residual correction layer superimposed on the predicted output of the deep learning model. The deep learning model uses a pre-trained offline thermodynamic neural operator as its main network structure, and the inter-layer connection weights of the main network structure are frozen during application deployment. The contact thermal resistance at the mating interface is utilized. coefficient of friction with interface Construct the set of parameters to be calibrated as shown in the following expression. : , in, Characterizes the contact thermal conductivity at the mating interface. Characterizes the frictional properties at the mating interface.

[0071] In a specific application scenario, the residual acquisition module 602 is used to perform form and position tolerance measurement on the assembled heavy rotor using measuring equipment arranged at the metrology and testing station, so as to collect the total runout of the heavy rotor after assembly. The total bounce As the macroscopic assembly residual, the heavy rotor, after undergoing induction heating, axial pressing, and cooling locking stages sequentially at the interference hot fitting station, is detached from the assembly fixture and transferred to the measurement and testing station; wherein, the total runout... The total radial offset of the actual rotation axis of the heavy-duty rotor relative to the ideal rotation axis during one circumferential rotation; the macroscopic assembly residual The calibration interface of the edge computing node is input according to the assembly sequence as feedback input for subsequent reverse calibration.

[0072] In specific application scenarios, the statistics triggering module 603 is used to maintain a fixed historical capacity based on the edge computing node. A residual sliding window is used to push the currently acquired macroscopic assembly residuals into the sliding window piece by piece as production progresses, and the earliest data in the window is removed; the moving average of the macroscopic assembly residual sequence within the sliding window is calculated. With moving variance As a statistical characteristic of the macroscopic assembly residual sequence; determine the mean drift tolerance indicated by the preset systematic drift triggering condition. and variance fluctuation limit , The maximum allowable offset of the moving average. The maximum allowable fluctuation in moving variance; continuously monitor the moving mean. Is the absolute value greater than the mean drift tolerance? The moving variance Is it greater than the variance fluctuation limit? When the moving average is detected The absolute value is greater than the mean drift tolerance. or the moving variance Greater than the variance fluctuation limit When the statistical feature satisfies the systematic drift triggering condition, the background reverse calibration program is activated; correspondingly, the statistical triggering module 603 is also used to detect the moving mean. The absolute value is less than or equal to the mean drift tolerance. And the moving variance Less than or equal to the variance fluctuation limit When the state is in a controlled steady state, calibration is not triggered and the input of the next macroscopic assembly residual is monitored.

[0073] In a specific application scenario, the statistical triggering module 603 is used to calculate the moving average of the macroscopic assembly residual sequence within the sliding window using the following formula. : , in, This represents the first [number] in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value The historical capacity of the sliding window is given; the moving variance of the macroscopic assembly residual sequence within the sliding window is calculated using the following formula. : , in, This represents the first [number] in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value The historical capacity of the sliding window, Represents the moving average; the moving average With the aforementioned moving variance As the statistical characteristic mentioned above.

[0074] In a specific application scenario, the inverse solving module 604 is used to determine the white-box physical proxy operator based on the bending moment integral, wherein the differentiable analytic mapping established by the white-box physical proxy operator is as shown in the following expression: , in, Indicates the heavy rotor at time The residual bending moment, This represents the integral operation over the axial mating interface of the heavy-duty rotor. Indicates time The lower-corrected interfacial micro-contact pressure field in cylindrical coordinates Spatial distribution values ​​at that location Represents circumferential coordinates. Indicates the axial coordinate. This represents the area of ​​the infinitesimal element on the mating interface. The vertical lever arm represents the spatial coordinates of the infinitesimal element area relative to the neutral axis of the heavy-duty rotor cross-section; wherein, when the infinitesimal element area is located above the neutral axis, the vertical lever arm... Taking a positive value, when the area of ​​the infinitesimal element is located below the neutral axis, the vertical lever arm... Take a negative value; the macroscopic assembly residual The boundary conditions are input to the differentiable analytical mapping to construct a loss function aimed at reducing the prediction jitter error. According to the following expression, the analytical gradient of the differentiable analytical map is backpropagated along the physical chain to inversely deduce the microscopic interface pressure field deviation that causes the macroscopic assembly residual anomaly. Based on the deviation of the microscopic interface pressure field The correction amount is obtained by solving the physical correlation between the correction and the set of parameters to be calibrated. : , in, Represents the set of parameters to be calibrated and , This indicates the contact thermal resistance at the mating interface. This represents the interfacial friction coefficient at the mating interface; Represents the loss function Regarding the set of parameters to be calibrated The gradient; This represents the macroscopic assembly residual; This represents the residual bending moment of the heavy-duty rotor; This represents the corrected interfacial microscopic contact pressure field; where, , This represents the correction amount for the contact thermal resistance. This represents the correction amount for the interface friction coefficient.

[0075] In specific application scenarios, the parameter update module 605 is used to set the amplitude boundary constraint shown in the following expression, and to use the amplitude boundary constraint to perform amplitude verification on the correction amount, wherein the amplitude boundary constraint is used to limit the norm of the correction amount in a single instance to not exceed a preset percentage threshold of the norm of the parameter reference value stored in the current state-aware model: , in, This represents the correction amount for the set of parameters to be calibrated. , This represents the correction amount for the contact thermal resistance. This represents the correction amount for the interface friction coefficient; This represents the baseline value of the parameter. , This indicates the current reference value for contact thermal resistance. This represents the current reference value for the interfacial friction coefficient; Norm operators; The preset percentage threshold is represented; if the norm of the correction amount satisfies the amplitude boundary constraint, the correction amount is taken as the effective correction amount, and the effective correction amount is updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated in a silent overlay manner, thus completing the online calibration of the model parameters of the state-aware model; if the norm of the correction amount does not satisfy the amplitude boundary constraint, the correction amount is truncated, so that the norm of the truncated correction amount is equal to The truncated correction amount is then silently overwritten and updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated, thus completing the online calibration of the model parameters of the state-aware model.

[0076] The system provided in this application freezes the main network structure and basic inference weights of the state-aware model, using only the physical latent variables within the interface as the parameter set to be calibrated. This allows online updates to avoid the dependence of offline reconstruction on the number of samples, thus adapting to the production characteristics of single-piece, small-batch production of heavy-duty rotors. Simultaneously, by real-time monitoring of the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence, it can automatically identify systematic drift and trigger reverse calibration without manual intervention, overcoming the subjectivity and lag of manual periodic line stop judgment. Furthermore, by calling a white-box physical surrogate operator based on bending moment integral and using its analytical gradient to propagate back along the physical chain to solve for the correction amount, the parameter update has clear physical causal interpretability, avoiding the limitations of manual experience in multi-parameter coupled nonlinear scenarios. In addition, by applying amplitude boundary constraints to the correction amount before silently updating it to the state-aware model, it can achieve active compensation for prediction bias and ensure the convergence stability of the update process, thereby continuously maintaining the model prediction fidelity without interrupting the production line.

[0077] It should be noted that other corresponding descriptions of the functional units involved in the online calibration system for heavy-duty rotor interference thermal assembly model parameters based on assembly residual inverse provided in this application embodiment can be found in the following references. Figures 1 to 5 The corresponding descriptions in [the document] will not be repeated here.

[0078] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0079] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0080] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0081] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0082] In an exemplary embodiment, see Figure 7 An electronic device is also provided, comprising a bus 701, a processor 702, a memory 703, and a communication interface 704. It may also include an input / output interface 705 and a display device 706. The various functional units can communicate with each other via the bus. The memory 703 stores a computer program, which, from bottom to top, includes a kernel 7034 that directly manages the hardware, middleware 7033 that provides general services for applications, an application programming interface 7032 that defines calling specifications, and an application 7031 that implements specific business logic. The processor 702 is used to read from the memory 703 via the bus 701 and execute the program stored in the memory, performing the online calibration method for heavy-duty rotor interference hot-fitting model parameters based on assembly residual inverse described in the above embodiments.

[0083] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the online calibration method for parameters of a heavy-duty rotor interference thermal assembly model based on assembly residual inverse.

[0084] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0085] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0086] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0087] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0088] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for online calibration of parameters of a heavy-duty rotor interference thermal assembly model based on inverse assembly residuals, characterized in that, include: A state-aware model for predicting the assembly state of heavy rotor interference hot fitting is deployed on an edge computing node. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are kept frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as the set of parameters to be calibrated. Using measuring equipment located at the metrology and testing station, the dimensional and positional tolerances of the assembled heavy-duty rotor are measured to obtain the total runout of the assembled heavy-duty rotor. The total bounce As the macroscopic assembly residual, the heavy rotor, after undergoing induction heating, axial pressing, and cooling locking stages sequentially at the interference hot fitting station, is detached from the assembly fixture and transferred to the measurement and testing station; wherein, the total runout... The total radial offset of the actual rotation axis of the heavy-duty rotor relative to the ideal rotation axis during one circumferential rotation; the macroscopic assembly residual The calibration interface of the edge computing node is input according to the assembly sequence as feedback input for subsequent reverse calibration; In the edge computing node, the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process are monitored. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. In response to the activation of the reverse calibration procedure, a white-box physical proxy operator based on moment integral is pre-installed in the edge computing node. The macroscopic assembly residual is used as the boundary condition. The analytical gradient of the white-box physical proxy operator is used to propagate backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current reference value. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor residual bending moment. The correction amount is subjected to amplitude boundary constraints, and the correction amount after amplitude boundary constraints is updated to the state-aware model to complete the online calibration of the model parameters of the state-aware model.

2. The method according to claim 1, characterized in that, The state-aware model deployed on edge computing nodes for predicting the assembly state of heavy-duty rotor interference thermal fitting, during the deployment and application of the state-aware model, maintains the main network structure and basic inference weights of the state-aware model frozen, and uses the physical latent variables within the mating interface of the heavy-duty rotor that cannot be directly measured as a set of parameters to be calibrated, including: The state-aware model with a two-layer architecture is constructed and deployed on the edge computing node. The two-layer architecture includes a deep learning model as a simulation baseline layer and a reality residual correction layer superimposed on the predicted output of the deep learning model. The deep learning model uses a thermodynamic neural operator that has been trained offline in advance as the main network structure. The inter-layer connection weights of the main network structure are frozen during the deployment and application. Utilizing the contact thermal resistance at the mating interface coefficient of friction with interface Construct the set of parameters to be calibrated as shown in the following expression. : ， in, Characterizes the contact thermal conductivity at the mating interface. Characterizes the frictional properties at the mating interface.

3. The method according to claim 1, characterized in that, In the edge computing node, the statistical characteristics of the macroscopic assembly residual sequence continuously accumulated according to the assembly time sequence in the current production process are monitored. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated, including: Based on the aforementioned edge computing nodes, maintain a fixed historical capacity. The residual sliding window pushes the currently acquired macroscopic assembly residual into the sliding window piece by piece as the production process progresses, and removes the earliest data in the window; Calculate the moving mean of the macroscopic assembly residual sequence within the sliding window. With moving variance As a statistical characteristic of the aforementioned macroscopic assembly residual sequence; Determine the mean drift tolerance indicated by the preset systematic drift triggering condition. and variance fluctuation limit , The maximum allowable offset of the moving average. The maximum allowable fluctuation in moving variance; Continuously detect the moving average Is the absolute value greater than the mean drift tolerance? The moving variance Is it greater than the variance fluctuation limit? ; When the moving average is detected The absolute value is greater than the mean drift tolerance. or the moving variance Greater than the variance fluctuation limit When the statistical feature satisfies the systematic drift triggering condition, the background reverse calibration program is activated. Accordingly, the method further includes: When the moving average is detected The absolute value is less than or equal to the mean drift tolerance. And the moving variance Less than or equal to the variance fluctuation limit When the state is in a controlled steady state, calibration is not triggered and the input of the next macroscopic assembly residual is monitored.

4. The method according to claim 3, characterized in that, The moving average of the macroscopic assembly residual sequence within the sliding window is calculated. With moving variance The statistical characteristics of the macroscopic assembly residual sequence include: The moving mean of the macroscopic assembly residual sequence within the sliding window is calculated using the following formula. : ， in, This represents the first [number] in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value The historical capacity of the sliding window; The moving variance of the macroscopic assembly residual sequence within the sliding window is calculated using the following formula. : ， in, This represents the first [number] in the macroscopic assembly residual sequence within the sliding window. A macroscopic assembly residual value The historical capacity of the sliding window, This represents the moving average; The moving average With the aforementioned moving variance As the statistical characteristic mentioned above.

5. The method according to claim 1, characterized in that, The invocation of the white-box physical proxy operator based on bending moment integral, pre-installed in the edge computing node, uses the macroscopic assembly residual as the boundary condition. The analytical gradient of the white-box physical proxy operator is propagated backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current benchmark value, including: The white-box physical surrogate operator based on bending moment integral is determined, wherein the differentiable analytic mapping established by the white-box physical surrogate operator is as shown in the following expression: ， in, Indicates the heavy rotor at time The residual bending moment, This represents the integral operation over the axial mating interface of the heavy-duty rotor. Indicates time The lower-corrected interfacial micro-contact pressure field in cylindrical coordinates Spatial distribution values ​​at that location Represents circumferential coordinates. Indicates the axial coordinate. This represents the area of ​​the infinitesimal element on the mating interface. The vertical lever arm represents the spatial coordinates of the infinitesimal element area relative to the neutral axis of the heavy-duty rotor cross-section; wherein, when the infinitesimal element area is located above the neutral axis, the vertical lever arm... Taking a positive value, when the area of ​​the infinitesimal element is located below the neutral axis, the vertical lever arm... Take the negative value; The macro assembly residual The boundary conditions are input to the differentiable analytical mapping to construct a loss function aimed at reducing the prediction jitter error. ; According to the following expression, the analytical gradient of the differentiable analytical map is backpropagated along the physical chain to inversely deduce the microscopic interface pressure field deviation that causes the macroscopic assembly residual anomaly. Based on the deviation of the microscopic interface pressure field The correction amount is obtained by solving the physical correlation between the correction and the set of parameters to be calibrated. : ， in, Represents the set of parameters to be calibrated and , This indicates the contact thermal resistance at the mating interface. This represents the interfacial friction coefficient at the mating interface; Represents the loss function Regarding the set of parameters to be calibrated The gradient; This represents the macroscopic assembly residual; This represents the residual bending moment of the heavy-duty rotor; This represents the corrected interfacial micro-contact pressure field; in, , This represents the correction amount for the contact thermal resistance. This represents the correction amount for the interface friction coefficient.

6. The method according to claim 2, characterized in that, The step of applying amplitude boundary constraints to the correction amount and updating the correction amount after amplitude boundary constraints to the state-aware model to complete the online calibration of the model parameters of the state-aware model includes: The amplitude boundary constraint is set as shown in the following expression, and the amplitude boundary constraint is used to perform an amplitude test on the correction amount, wherein the amplitude boundary constraint is used to limit the norm of the correction amount in a single instance from not exceeding a preset percentage threshold of the norm of the parameter reference value stored in the current state-aware model: ， in, This represents the correction amount for the set of parameters to be calibrated. , This represents the correction amount for the contact thermal resistance. This represents the correction amount for the interface friction coefficient; This represents the baseline value of the parameter. , This indicates the current reference value for contact thermal resistance. This represents the current reference value for the interfacial friction coefficient; Norm operators; This represents the preset percentage threshold; If the norm of the correction quantity satisfies the amplitude boundary constraint condition, then the correction quantity is taken as the effective correction quantity, and the effective correction quantity is updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated in a silent overlay manner, thereby completing the online calibration of the model parameters of the state-aware model. If the norm of the correction quantity does not satisfy the amplitude boundary constraint condition, then the correction quantity is truncated, such that the norm of the truncated correction quantity is equal to... The truncated correction amount is then silently overwritten and updated to the adjustable parameters in the state-aware model corresponding to the parameter set to be calibrated, thus completing the online calibration of the model parameters of the state-aware model.

7. An online calibration system for the parameters of a heavy-duty rotor interference thermal assembly model based on the inverse of assembly residuals, characterized in that, include: The state monitoring and locking module is used to deploy a state-aware model on an edge computing node to predict the assembly state of heavy rotor interference hot fitting. During the deployment and application of the state-aware model, the main network structure and basic inference weights of the state-aware model are kept frozen, and the physical latent variables that cannot be directly measured within the mating interface of the heavy rotor are used as a set of parameters to be calibrated. The residual acquisition module is used to measure the form and position tolerances of the assembled heavy-duty rotor using measuring equipment arranged at the metrology and inspection station, so as to collect the total runout of the assembled heavy-duty rotor. The total bounce As the macroscopic assembly residual, the heavy rotor, after undergoing induction heating, axial pressing, and cooling locking stages sequentially at the interference hot fitting station, is detached from the assembly fixture and transferred to the measurement and testing station; wherein, the total runout... The total radial offset of the actual rotation axis of the heavy-duty rotor relative to the ideal rotation axis during one circumferential rotation; the macroscopic assembly residual The calibration interface of the edge computing node is input according to the assembly sequence as feedback input for subsequent reverse calibration; The statistical triggering module is used in the edge computing node to monitor the statistical characteristics of the macroscopic assembly residual sequence that is continuously accumulated according to the assembly time sequence in the current production process. When the statistical characteristics are found to meet the preset systematic drift triggering conditions, the background reverse calibration program is activated. The reverse solution module is used to respond to the activation of the reverse calibration program by calling the white-box physical proxy operator based on moment integral pre-installed in the edge computing node. Using the macroscopic assembly residual as the boundary condition, the analytical gradient of the white-box physical proxy operator is used to propagate backward along the physical chain of the differentiable analytical mapping to obtain the correction amount of the parameter set to be calibrated relative to the current reference value. The white-box physical proxy operator establishes a differentiable analytical mapping between the microscopic contact pressure field at the mating interface and the rotor residual bending moment. The parameter update module is used to apply amplitude boundary constraints to the correction amount and update the correction amount after amplitude boundary constraints to the state-aware model, so as to complete the online calibration of the model parameters of the state-aware model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images