Turbine blade thermal barrier coating real-time sensing and reliability integrated prediction method based on physical neural network

Abstract

The invention discloses a turbine blade thermal barrier coating real-time sensing and reliability integrated prediction method based on a physical neural network. The method comprises the steps that firstly, the temperature of the turbine blade thermal barrier coating is recorded through an infrared detector; 2, recording the peeling area of the thermal barrier coating of the turbine blade by using a CCD camera; 3, converting the peeling area into a damage degree D by using an image recognition technology, and obtaining damage time sequence data D (t); 4, constructing a thermal barrier coating microstructure geometric model by using COMSOL, and dividing grids to obtain an oxide energy release rate; 5, performing life prediction of different groups by using a physical information neural network; step 6, dynamic updating of model parameters and continuous optimization of prediction capability are realized; and step 7, performing reliability prediction of different groups by using the physical information neural network. The method has the characteristics of integration of physical mechanism driving, real-time dynamic sensing, full-process automation and high-precision reliability prediction.

Description

Technical Field

[0001] This invention belongs to the field of turbine blade thermal barrier coating technology, specifically relating to a method for real-time sensing and reliability prediction of turbine blade thermal barrier coating based on physical neural networks. Background Technology

[0002] The improvement of aero-engine performance directly depends on the increase of turbine inlet temperature, making thermal barrier coatings (TBCs) a key technology for ensuring reliable engine operation. TBCs, through their excellent thermal insulation and oxidation resistance, enable turbine blades to operate stably for extended periods in extreme environments exceeding the melting point of the base material. Currently, this coating system has evolved into a multi-layered structure consisting of a ceramic layer (typically ZrO2 partially stabilized by Y2O3) and a binder layer (MCrAlY alloy), prepared through processes such as atmospheric plasma spraying or physical vapor deposition. The integrity of the TBC directly affects the safety and lifespan of the engine; therefore, accurate monitoring and reliability assessment of the TBC are of paramount importance.

[0003] However, existing reliability evaluation models rely solely on the statistical correlation between "input parameters and output lifetime" to make predictions, without embedding the core physical laws of coating failure. This leads to two problems: first, the prediction results lack physical interpretability and cannot explain the physical logic corresponding to the lifetime results, resulting in low engineering credibility; second, under extreme operating conditions, they are prone to outputting erroneous predictions that violate common sense physics, and cannot be corrected through physical principles.

[0004] Secondly, the core data source for existing reliability evaluation models is offline finite element simulation data, which does not include real-time monitoring information during turbine blade service. In actual service, the coating may face sudden conditions such as rapid temperature rise and foreign object impact. Models based on offline data cannot perceive real-time damage changes and can only conduct static reliability evaluations. This leads to the accumulation of deviations between the predicted results and the actual state of the coating over time, making it impossible to dynamically update the remaining service life.

[0005] Finally, existing methods need to be broken down into multiple independent steps such as parameter measurement, macroscopic simulation, sample construction, microscopic calculation, model training, and reliability analysis. The connection between each step is highly dependent on manual operation (such as manually exporting macroscopic field data, organizing finite element results, and converting data formats), which is not only inefficient but also prone to inconsistent results due to operational errors. Furthermore, it has not formed an automated process from real-time perception to reliability output, making it difficult to integrate into the engine online operation and maintenance system.

[0006] Therefore, there is an urgent need for a new method that can deeply integrate real-time monitoring data with failure physical mechanisms and can efficiently and accurately predict reliability in order to overcome the above-mentioned technical defects. Summary of the Invention

[0007] In order to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method for real-time perception and reliability prediction of turbine blade thermal barrier coating based on physical neural network. This method has the characteristics of physical mechanism driving, real-time dynamic perception, full-process automation and high-precision reliability prediction.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for real-time sensing and reliability prediction of turbine blade thermal barrier coatings based on physical neural networks includes the following steps; Step 1: Use an infrared detector to record the temperature of the thermal barrier coating on the turbine blades; Step 2: Use a CCD camera to record the area of ​​thermal barrier coating peeling off the turbine blades; Step 3: The temperature field of the turbine blade thermal barrier coating is divided into zones. Image recognition techniques (such as threshold segmentation, morphological processing, and region calculation algorithms) are used to process the synchronously acquired CCD images, and the peeling area is quantified as the degree of damage in that zone. D Acquire damage time series data D ( t ); Step 4: Use COMSOL to construct the geometric model of the thermal barrier coating microstructure and mesh it. Construct the mechanical-thermal-chemical coupling constitutive equation, oxygen diffusion equation, and oxide growth control equation of the thermal barrier coating, and input them into COMSOL as solution equations. Input the material parameters and the local temperature of the blade into COMSOL as solution material parameters. Set boundary conditions. Complete the microstructure finite element calculation to obtain the Mises stress and calculate the oxide energy release rate. Step 5: Construct a Physical Information Neural Network (PINN), which uses time... t For input neurons, based on the degree of damage D For output neurons; the core of network training lies in the multi-objective design of the loss function: on the one hand, utilizing the damage time-series data obtained from real-time monitoring. D ( t On the one hand, the data loss is calculated; on the other hand, the core physical laws of the Paris lifetime model are used as hard constraints, and physical information neural networks are used to predict the lifetime of different groups. The residual terms (i.e., the degree to which the physical equations are satisfied) are calculated using automatic differentiation techniques. By optimizing this combined loss function, a dual-objective optimization of physical constraints and data-driven approaches is achieved, forcing the network's prediction output to strictly follow the intrinsic physical correlation between "damage-life" and making the network parameters implicitly reflect the physical properties of the material. Step Six: Establish a dynamic iterative prediction and update mechanism; the system automatically collects new infrared and CCD monitoring data at preset time intervals (e.g., every second or every working cycle). When events such as foreign object impacts cause peeling area to be detected, the system will update the data. A spall When a sudden change occurs, the system quickly recalculates the degree of damage based on the above formula. D ( t The process involves combining updated temperature and physical field data with the physical information neural network for incremental training. This process ensures that the model input can respond in real time to any changes in the coating state (e.g., a sudden increase of 30% in the peeling area), thereby significantly reducing the lifetime prediction error and achieving dynamic matching between the prediction and the actual damage state. Step 7: Considering data dispersion, use a physical information neural network to predict the reliability of different groups.

[0009] In step one, the infrared detector is fixedly installed on the inner wall of the engine turbine casing. The installation axis of the infrared detector is parallel to the turbine disk axis, and its optical window is ensured to face the blade area of ​​the turbine blade to capture the complete blade surface temperature field and avoid being blocked by adjacent blades or tenon structures.

[0010] In step two, the CCD camera is fixed inside the thick-walled structure of the engine turbine casing. The optical axis of the camera at each station forms an angle of 15° to 45° with the radial direction of the turbine disk, and its field of view is aligned with at least one critical area among the leading edge, blade base, or blade back of the blade. This ensures that the CCD camera's field of view completely covers the critical areas of the turbine blade (such as the middle of the blade, the leading edge, and other areas prone to peeling), and avoids visual obstruction during blade rotation. Given the high temperature and high pressure environment inside the turbine casing, the CCD camera, its lens, and the light source must be integrated into a high-temperature, pressure-resistant protective cover with an active or passive cooling system to ensure long-term stable operation of the optical components under harsh conditions and to maintain lens cleanliness.

[0011] In step three, the surface of the blade thermal barrier coating is divided into several sub-regions with similar temperature characteristics. The similar temperature characteristics refer to the fact that, under given steady-state operating conditions, the surface temperature at each point in the same sub-region deviates from the average temperature of that sub-region by no more than ±30℃, thereby ensuring that the thermal barrier coating in that region exhibits coordinated damage evolution behavior. This partitioning strategy ensures that the thermal load conditions in each region are consistent, laying the foundation for establishing a partitioned damage evolution model. Degree of damage D =Area of ​​peeling / Area of ​​grouped regions, damage time series data D ( t This can be achieved through the following formula: In the formula,A spall ( t ) indicates time t The pixel area of ​​the coating peeling area obtained by image recognition (converted to the actual physical area after calibration). A region This represents the total area of ​​the temperature zone. This formula ensures the damage parameters... D ( t (0≤) D ≤1) Directly and quantitatively reflects the true physical peeling state of the coating surface, providing key data constraints that accurately correspond to the "actual state" for subsequent physical models and neural networks; Damage was obtained from different groups D ( t ).

[0012] In step four, the geometric model of the thermal barrier coating microstructure is constructed during the modeling process. The microstructure includes the multi-layer structure of the ceramic layer, oxide layer, and adhesive layer of the thermal barrier coating, as well as the interface roughness and pore microstructure. The failure driving force of the coating is quantified by solving the physical field equations of the mechanical-thermal coupling. The specific equation is as follows: (1) (2) (3) (4) (5) (6) (7) In the formula This represents the symmetric Cauchy stress tensor. ε It is the total strain. u It is displacement. ε e It is elastic strain. ε ox It is growth strain. ε th It is thermal strain. It is the force per unit volume. c Indicates the concentration of oxygen. n To express the volume fraction of oxides, G It is the energy release rate. D It refers to the degree of damage to the thermal barrier coating. D=1 indicates thermal barrier coating failure; the others are material parameters. Equation (4) couples thermal mismatch and high-temperature oxidation, while equations (5) and (6) consider interfacial oxidation. High-temperature oxidation failure often occurs at the TGO interface cracking, therefore in equation (7) The average Von Mises stress is represented at the TC / TGO interface and the TGO / BC interface. Optionally, failure mechanisms such as corrosion and sintering can also be considered.

[0013] The parameters of the mechanothermal coupling constitutive equation for the mechanothermal barrier coating considered in step four include one or more of the following: high-temperature interface oxidation, creep, sintering, and high-temperature corrosion.

[0014] The physical information neural network in step five is built using the PyTorch framework. PyTorch's advantages lie in its automatic differentiation function, flexibility, and powerful GPU acceleration support, enabling it to effectively handle complex physical problems. It includes an input layer, an output layer, and multiple hidden layers. The number of hidden layers is adjusted according to the degree of nonlinearity. Adjacent layers in the physical information neural network are connected by weights, and a hyperbolic tangent activation function is used to introduce nonlinear characteristics. The hyperbolic tangent function maps the input to an interval (…). 1,1), with smooth gradients, helps alleviate the problems of gradient vanishing or exploding, and can usually effectively improve training stability; The key to physical information neural networks is embedding information from physical equations during network training. The residuals of the physical equations are added as an additional loss term to the total loss function. This is typically achieved using automatic differentiation techniques. In PyTorch, calculating the residuals of the physical equations involves calculating the derivative of the network output with respect to the input. The time and spatial second derivatives of the neural network output are calculated using PyTorch's `torch.autograd` module; specifically, the derivatives are calculated using the `torch.autograd.grad` function, and these derivatives are substituted into the physical equations to calculate the residuals. The loss function is designed to consider not only data fitting errors but also the physical residuals. By minimizing the combined loss of the physical equation residuals and data errors, the network learns solutions that conform to physical laws.

[0015] To optimize network training, the Adam optimizer was used with a learning rate of 0.001 and weight decay of 0.001. An early stopping mechanism was also implemented, and the optimal model was dynamically saved based on the performance on the validation set. A random seed was also set to ensure the reproducibility of the experiment. To avoid gradient explosion, the input and output training data were Z-score normalized.

[0016] In step six, based on the constructed physical information neural network, the damage time series data obtained in step three are fused together for different temperature groups. D (t The energy release rate data obtained in step four G ( T , t Each network group is trained independently. After training, for each group, the desired future time parameters are input, and the corresponding damage level is output through the physical information neural network. D The predicted values ​​are used to achieve accurate prediction of the lifespan of each coating group.

[0017] In step seven, the data dispersion is described using a Weibull distribution, specifically: (8) parameter m For fixed constants, parameters n Follows a Weibull distribution and has a fixed constant. m and parameters n The mean and variance are determined by the data obtained from CCD image recognition.

[0018] In step seven, the failure probability is obtained using the Monte Carlo method and mapped onto the computational grid of the turbine blade thermal barrier coating. A reliability cloud map is then obtained through visualization software.

[0019] The beneficial effects of this invention are: This invention, through the structured design of a physical information neural network, transforms the core physical mechanism of thermal barrier coating failure into a hard constraint of the model, completely solving the fundamental defect of existing technologies that are "data-driven but lack physical logic." Its specific advantages are reflected in two aspects: This invention incorporates a physical equation residual term into the loss function of a physical information neural network, embedding the core physical laws of the Paris lifetime model into the training process. This allows the intermediate layer parameters of the network to indirectly correspond to the real physical properties of materials, such as elastic modulus and fracture toughness. The Paris lifetime model can directly reveal the power-law relationship between damage and lifetime.

[0020] The physical information neural network of this invention forces the prediction results to satisfy the inherent physical correlation between "damage-life" through dual-objective optimization of physical constraints and data-driven approaches.

[0021] This invention employs a feature-level fusion design of multi-source real-time data, including the spatiotemporal temperature field distribution on the blade surface acquired by an infrared detector. T ( x , y, z , t The image sequence of the peeling area on the coating surface captured by the CCD camera, and the microstructure stress field and energy release rate obtained by COMSOL simulation calculation. G ( T , tThe spatial registration of temperature and image data, coupled modeling of physical fields and damage data, and dynamic iterative prediction mechanism break through the static limitations of existing technologies that "rely on offline simulation data," and construct a prediction system that dynamically iterates according to service status. Its advantages are as follows: Real-time perception of damage status reduces prediction bias This invention introduces a collaborative acquisition module combining infrared temperature monitoring and CCD image recognition. It processes temperature field data in real time and uses an image segmentation algorithm to convert the peeling area into a quantified degree of damage. D ( t The data is then used as a constraint input into the physical information neural network. For example, when a foreign object impact causes a sudden increase of 30% in the local peeling area, the system can update the damage parameters within 10 seconds, reducing the remaining life prediction error from 22% in traditional methods to 4.5%, accurately matching the actual state of the coating.

[0022] Dynamically update the prediction model to support operation and maintenance decisions. This invention designs an online adaptive training mechanism: the system automatically injects newly collected "real-time monitoring data - damage degree" samples and energy release rate data into the incremental training process of the physical information neural network at preset time intervals.

[0023] This invention integrates the entire process of monitoring, modeling, and analysis through an end-to-end modular design, solving the engineering pain points of existing technologies such as "fragmented processes and heavy reliance on manual labor." Its specific advantages are as follows: Automating the connection between various processes improves efficiency and consistency. This invention develops a data interface module for COMSOL-Physical Information Neural Network (PINN), which automatically reads macroscopic field simulation data. Simultaneously, it achieves seamless integration of monitoring equipment, physical modeling, and neural network training through a standardized data bus. In practical applications, the entire process time is reduced from 72 hours to 4.5 hours, and by eliminating manual operation, the consistency of results is improved to 99.2%, solving the industry problem of "significant differences in results among different operators."

[0024] The end-to-end architecture is adapted to online operation and maintenance systems, enhancing engineering practicality; the integration of this invention makes it the first thermal barrier coating reliability prediction solution that can be directly deployed in online operation and maintenance scenarios. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the evaluation method of the present invention.

[0026] Figure 2 This is a specific application scenario in one embodiment of the present invention.

[0027] Figure 3This is a local microscale finite element geometric model and mesh generation in one embodiment of the present invention.

[0028] Figure 4 This is a diagram of a physical information neural network structure for this invention.

[0029] Figure 5 This is the loss function for the training process of the physical information neural network in one embodiment of the present invention.

[0030] Figure 6 This invention provides a lifetime prediction for thermal barrier coatings at different temperature groups in one embodiment of the present invention.

[0031] Figure 7 This represents the failure probability of the thermal barrier coating at 1100°C in one embodiment of the present invention.

[0032] Figure 8 This is a reliability failure probability cloud map of the turbine blade thermal barrier coating after 400 hours and 800 hours in one embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0034] like Figure 1 As shown, the integrated real-time sensing and reliability prediction method for turbine blade thermal barrier coating based on physical neural networks of the present invention includes the following steps: (1) To obtain real-time temperature field data of the blade surface, such as Figure 2 As shown, an infrared detector is used to record the temperature of the thermal barrier coating on the turbine blades to achieve non-contact real-time monitoring of the temperature field, providing key input for subsequent thermal load zoning and damage analysis; (2) To collect image data of the blade surface, such as Figure 2 As shown, the area of ​​the thermal barrier coating peeled off the turbine blade was recorded using a CCD to establish a visual record of the coating damage, providing an image basis for the quantitative conversion of the damage degree. (3) In order to divide the blade surface into zones according to the temperature distribution and convert the peeling area into the degree of damage. D The turbine blade thermal barrier coating temperatures are rationally grouped. In this embodiment, the blade coating is divided into three groups based on the temperatures recorded by the infrared detector: 900-1350K, 1350-1400K, and 1400-1460K. This is based on the damage... D=Area of ​​peeling away / Area of ​​grouped regions, using image recognition technology to convert the area of ​​peeling away into the degree of damage. D Damage was obtained in different groups. D ( t This enables the quantification of damage extent and spatial partitioning management, providing damage time-series data for subsequent physical information neural networks; (4) In order to simulate the response behavior of the coating under thermo-mechanical coupling, a thermal barrier coating microstructure geometric model considering local microscale structure was constructed in COMSOL software. In this example, the average interface roughness was taken. R =10μm, and for a more comprehensive approach, non-periodic interface roughness and internal pores and other microstructures can be considered during modeling, such as Figure 3 When setting boundary conditions, Figure 3 The bottom of the model is fixed, and the left boundary is... x The directional displacement is zero, and the left and right boundaries are set to be symmetrical. This example only considers two failure mechanisms: high-temperature oxidation and thermal mismatch. The specific equations are as follows: (1) (2) (3) (4) (5) (6) (7) In the formula This represents the symmetric Cauchy stress tensor. ε It is the total strain. u It is displacement. ε e It is elastic strain. ε ox It is growth strain. ε th It is thermal strain. It is the force per unit volume. c Indicates the concentration of oxygen. n To express the volume fraction of oxides, G It is the energy release rate. In the above equations (5) and (6), interfacial oxidation is considered, and equation (4) couples thermal mismatch and high-temperature oxidation. Optionally, failure mechanisms such as corrosion and sintering can also be considered. Through multiphysics coupling analysis, the Mises stress and energy release rate are accurately calculated. G This study reveals the microscopic physical mechanism of coating failure.

[0035] (5) To achieve accurate dynamic prediction of coating lifetime, a deep neural network was further constructed. Damage at different temperature groups was obtained based on infrared detectors, CCD image recognition, and data conversion technology. D Over time t A changing dataset, where time... t As the input neuron of PINN, damage D As the output neuron of PINN, the Paris lifetime model serves as a physical constraint: (8) In the formula D It is thermal barrier coating damage, when D When the value is 1, the thermal barrier coating fails. a , b If the material parameters can be iteratively optimized by the neural network based on online identification data, then the network's physical loss function is: (9) Damage monitored in real time D ( t As a data constraint, the network's data loss function is: (10) The total loss function of the network is: (11) in w 1 represents the data loss weight. w 2 represents the weights of the physical function. Based on the coordination between the physical loss function and the data loss function, the weights are... w Set 1 to 1. w 2 is set to 100. The network parameters are updated by minimizing the total loss function. For example... Figure 4 As shown, the network has two hidden layers, each with 64 neurons. Adjacent layers are connected by weights, and a hyperbolic tangent activation function is used. To optimize network training, the Adam optimizer is employed with a learning rate of 0.001, weight decay of 0.001, and an early stopping mechanism. The optimal model is dynamically saved based on validation set performance, and a random seed of index 42 is used to ensure experimental repeatability. To avoid gradient explosion, the input and output training data are Z-score standardized. By fusing physical constraints and data-driven approaches, the accuracy and generalization ability of lifetime prediction are significantly improved. The loss function changes during training as shown below. Figure 5 It can be observed that the loss function converged after 20,000 training iterations, indicating that training is complete. Based on the trained physical information neural network model, lifetime predictions are performed for different groups, such as... Figure 6 It can be seen that the network predictions and experimental values ​​at different temperatures are in good agreement.

[0036] (6) In order to establish an online adaptive mechanism for the coating life prediction model, so that the prediction system can continuously optimize the model parameters based on real-time monitoring data and adapt to the state changes of the coating during service, the system automatically collects real-time monitoring data at preset time intervals and updates the newly acquired "time" data. The "damage level" sample and energy release rate data are injected into the physical information neural network for incremental training, that is, steps one to five are repeated to achieve dynamic updating of model parameters and continuous optimization of prediction ability. (7) To account for data dispersion and achieve quantitative assessment and spatial visualization of coating failure probability, data dispersion is described using a Weibull distribution, specifically: (12) parameter n Following a Weibull distribution, in this example, within the 900-1350K group... m It is 4.43567e-6. n The mean is 1.38, and the shape parameter is 23; within the 1350-1400K group. m It is 4.43567e-6. n The mean is 1.5, and the shape parameter is 23; within the 1400-1460K group. m It is 4.43567e-6. n The mean is 1.8, and the shape parameter is 23; Select D The network was trained using real-time monitoring data with a value within 0.2, and predictions were made based on the trained network. The prediction results for the 1350-1400K groups are as follows: Figure 7 As shown in (a), and further using the Monte Carlo method to obtain its failure probability, as follows: Figure 7 As shown in (b), the data is mapped onto the computational grid of the turbine blade thermal barrier coating, and a reliability cloud map is obtained through visualization software, as shown in [example]. Figure 8 As shown, the failure probability of the blade leading edge is 0.2 at 400h and reaches 1 at 800h, indicating that the thermal barrier coating in this area will basically peel off at this time.

[0037] This invention uses the core physical laws governing thermal barrier coating failure (represented by the Paris lifetime model) as physical constraints, embedding them into the loss function of a physical information neural network to construct a dual-driven prediction model of "physical equation constraints + real-time monitoring data constraints." By using a specific physical model (such as the Paris model) as a hard constraint for the neural network, the model's prediction results are ensured to conform to both statistical data patterns and strict adherence to physical mechanisms.

[0038] This invention combines infrared temperature monitoring with CCD image recognition to obtain real-time damage levels. D (t Using real-time monitoring data (especially damage levels quantified through image recognition) as data constraints for the physical information neural network, a method for dynamically updating the life prediction model was achieved.

[0039] This invention designs a process from real-time data acquisition (infrared / CCD) to image recognition and feature extraction (damage level). D This solution provides a fully automated, integrated end-to-end solution for reliability assessment of thermal barrier coatings, encompassing physical modeling (COMSOL microscale finite element analysis), physical information neural network training and prediction, and reliability analysis and visualization. It integrates data interfaces and processing flows between specific hardware (infrared, CCD) and specific software (COMSOL, physical information neural network framework).

[0040] This invention, building upon deterministic lifetime prediction using physical information neural networks, further considers the dispersion of monitoring data. It employs the Weibull distribution to describe lifetime distribution and utilizes Monte Carlo sampling for statistical sampling, ultimately achieving probabilistic reliability assessment rather than merely point prediction. This invention addresses data dispersion by combining physical information neural networks with probabilistic statistical models (such as the Weibull distribution) for quantitative reliability assessment and visualization.

Claims

1. A physical neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method, characterized in that, Includes the following steps; Step 1: Use an infrared detector to record the temperature of the thermal barrier coating on the turbine blades; Step 2: Use a CCD camera to record the area of ​​thermal barrier coating peeling off the turbine blades; Step three, group the temperature of turbine blade thermal barrier coating, use image recognition technology to convert the spalling area into damage degree D , obtain damage timing data D ( t ) Step 4: Use COMSOL to construct the geometric model of the thermal barrier coating microstructure and mesh it. Construct the mechanical-thermal-chemical coupling constitutive equation, oxygen diffusion equation, and oxide growth control equation of the thermal barrier coating, and input them into COMSOL as the solution equations. Input the material parameters and the local temperature of the blade into COMSOL as the solution material parameters. Set the boundary conditions, complete the finite element calculation of the microstructure to obtain the Mises stress, and calculate the oxide energy release rate. Step five, build a physical information neural network, which takes time t as input neurons and damage degree D as output neurons; calculate data loss using real-time monitoring damage time series data D ( t ) ; use the core physical law of Paris life model as a hard constraint to perform life prediction for different groups using a physical information neural network; Step six, dynamic iteration prediction and update mechanism is established; the system automatically collects new infrared and CCD monitoring data at preset time intervals, and when events such as foreign object impact are monitored, the peeling area A spall When mutation occurs, the damage degree is recalculated D ( t ), and the physical information neural network is injected for incremental training combined with updated temperature and physical field data; Achieve dynamic matching between predictions and actual damage states; Step 7: Use a physical information neural network to predict the reliability of different groups.

2. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 1, characterized in that, In step one, the infrared detector is fixedly installed on the inner wall of the engine turbine casing. The installation axis of the infrared detector is parallel to the turbine disk axis, and its optical window is ensured to face the blade area of ​​the turbine blade to capture the complete blade surface temperature field.

3. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 2, characterized in that, In step two, the CCD camera is fixed inside the wall structure of the engine turbine casing. The optical axis of the camera at each station forms an angle of 15° to 45° with the radial direction of the turbine disk, and its field of view center is aligned with at least one key part among the leading edge of the blade, the blade base, or the blade back, to ensure that the field of view of the CCD camera can completely cover the key area of ​​the turbine blade and avoid visual obstruction during the rotation of the blade.

4. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 3, characterized in that, In step three, the surface of the blade thermal barrier coating is divided into several sub-regions with similar temperature characteristics. This zoning strategy ensures consistent thermal load conditions within each region, laying the foundation for establishing a zoning damage evolution model; Degree of injury D = area of peeling / area of grouping region, injury timing data D t ) is achieved by the following equation:​ wherein A spall ( t ) indicates the time t The pixel area of the coating peeling area obtained by image recognition (converted to the actual physical area by calibration), A region The total area of the temperature partition, and the damage timing data of different groups are obtained D ( t ).

5. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method of claim 4, wherein, In step four, the geometric model of the thermal barrier coating microstructure is constructed during the modeling process. The microstructure includes the multi-layer structure of the ceramic layer, oxide layer, and adhesive layer of the thermal barrier coating, as well as the interface roughness and pore microstructure. The failure driving force of the coating is quantified by solving the physical field equations of the mechanical-thermal coupling. The specific equation is as follows: （1） （2） （3） （4） （5） （6） （7） where is the symmetric Cauchy stress tensor, ε is the total strain, u is the displacement, ε e is the elastic strain, ε ox is the growth strain, ε th is the thermal strain, is the force per unit volume, c denotes the concentration of oxygen, n denotes the volume fraction of oxide, G is the energy release rate, D is the thermal barrier coating damage, when D = 1 the thermal barrier coating fails, otherwise it is a material parameter.

6. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 5, characterized in that, The parameters of the mechanothermal coupling constitutive equation for the mechanothermal barrier coating considered in step four include one or more of the following: high-temperature interface oxidation, creep, sintering, and high-temperature corrosion.

7. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 6, characterized in that, The physical information neural network in the step five is built by using a PyTorch framework, and contains an input layer, an output layer and multiple hidden layers, the number of layers of the hidden layers is adjusted according to the degree of nonlinearity, the adjacent layers in the physical information neural network are connected through weights, and a hyperbolic tangent activation function is used to introduce a nonlinear characteristic, the hyperbolic tangent function maps the input to the interval (0, 1). 1,1). In PyTorch, calculating the residuals of physical equations involves calculating the derivative of the network output with respect to the input. The time derivative and the second spatial derivative of the neural network output are calculated using PyTorch's torch.autograd module; the derivatives are then calculated using the torch.autograd.grad function and substituted into the physical equations to calculate the residuals.

8. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 7, characterized in that, In step six, based on the constructed physical information neural network, damage time-series data are fused for different temperature groups. D ( t ) and energy release rate data G ( T , t Each network group is trained independently. After training, for each group, the input of the desired future time parameters is used to output the corresponding degree of damage through the physical information neural network. D The predicted values ​​enable accurate prediction of the lifespan of each coating group.

9. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method according to claim 8, characterized in that, In step seven, the data dispersion is described using a Weibull distribution, specifically: Parameters m are fixed constants, parameters n are fixed constants m and parameters n The mean and variance of the Weibull distribution are determined from the data identified by the CCD image.

10. The physics-based neural network-based turbine blade thermal barrier coating real-time perception and reliability integrated prediction method of claim 9, wherein, In step seven, the failure probability is obtained using the Monte Carlo method and mapped onto the computational grid of the turbine blade thermal barrier coating. A reliability cloud map is then obtained through visualization software.

Images