A landing gear brake disc state prediction method based on clamp-weighted PINN

By using the clamp-weighted PINN model, the problems of slow reconstruction speed and insufficient generalization ability of brake disc temperature field in traditional methods are solved, realizing real-time monitoring and digital operation and maintenance of the braking system of civil aircraft, and improving the accuracy and stability of temperature field prediction.

CN123116869APending Publication Date: 2026-11-17HARBIN INST OF TECH AT WEIHAI
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611600629.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot quickly reconstruct the brake disc temperature field under limited measurement conditions, thus failing to meet the real-time monitoring requirements of aviation braking systems. Furthermore, traditional PINN models lack generalization capabilities and cannot adapt to rapid switching in complex operating conditions.

Method used

By adopting the clamp-weighted PINN model, a physical model of the brake disc assembly is established, unsteady heat conduction equations and mixed boundary conditions are set, and temperature field simulation control equations are constructed. Using the clamp-weighted loss function and adaptive optimizer, a deep multilayer perceptron model with six-dimensional parameterized feature vectors is trained to achieve high-precision prediction of the temperature field.

Benefits of technology

It improves the model's prediction accuracy and generalization ability under complex operating conditions, realizes real-time health monitoring and digital operation and maintenance of aircraft braking systems, and enhances the accuracy and stability of temperature field reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116869A_ABST
    Figure CN123116869A_ABST
Patent Text Reader

Abstract

This invention relates to the field of monitoring the working status of multi-disc braking systems in civil aircraft. Specifically, it is a method for predicting the state of landing gear brake discs based on caliper-weighted PINN, which can quickly reconstruct the temperature field of brake discs under limited measurement conditions, enabling monitoring and digital operation and maintenance of aviation braking systems. The method includes establishing a physical model of the brake disc assembly, constructing temperature field simulation control equations and a simulation volume heat source system, constructing a braking condition matrix dataset through three-dimensional temperature field simulation, establishing a caliper-weighted PINN model, and completing model training. The accuracy of this invention is significantly improved compared to the basic model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of monitoring the working status of multi-disc brake systems in civil aircraft, specifically a method for predicting the status of landing gear brake discs based on caliper weighted PINN that can quickly reconstruct the brake disc temperature field under limited measurement conditions, enabling monitoring and digital operation and maintenance of aviation brake systems. Background Technology

[0002] With increasing safety and efficiency requirements for civil aviation aircraft, real-time and accurate acquisition of the full-range temperature field of brake discs is crucial for heat dissipation management, lifespan prediction, and fault diagnosis. However, existing methods have limitations: traditional thermocouples, constrained by installation space, harsh environments, and material limitations, can only provide discrete measurement point data and cannot reflect the three-dimensional temperature field; while finite element simulation can simulate temperature evolution, the computation time is enormous, making it difficult to meet the needs of real-time online monitoring. PINN integrates physical laws and deep learning, enabling high-precision prediction of temperature fields and inversion of thermal parameters with small samples, providing a new approach to thermal analysis. However, traditional PINN is trained for specific operating conditions, requiring retraining for parameter changes, lacking generalization ability, and unable to meet the real-time prediction needs of rapid switching of operating conditions in aviation operations. Summary of the Invention

[0003] This invention addresses the shortcomings and deficiencies of existing technologies by proposing a landing gear brake disc state prediction method based on clamp weighted PINN, which can improve the prediction accuracy and generalization ability of the model under complex operating conditions and provide a basis for real-time health monitoring and digital operation and maintenance of aviation braking systems.

[0004] This invention achieves its purpose through the following measures: A method for predicting the state of landing gear brake discs based on clamp-weighted PINN, characterized by the following steps: Step 1: Establish a physical model of the brake disc assembly, perform conformal treatment on the contact surfaces between the discs, and set the unsteady-state heat conduction equation, zoned heat transfer boundaries, and material parameters; Step 2: Construct the temperature field simulation control equations and the simulation volume heat source system. The time-varying volume heat source is implemented using a semi-implicit source model. The power time-varying curve is defined through an external dictionary file, and an over-boundary clamping function is set. The volume heat source acts on a thin layer region composed of one unit on each side of the friction surface, forming a dual volume heat source system. Step 3: Construct a braking condition matrix dataset through three-dimensional temperature field simulation. Design the condition matrix with landing mass and initial landing velocity as variables, complete the transient simulation, extract the temperature field at key moments and downsample it to form a training set, a validation set and a test set. Step 4: Establish the clamp-weighted PINN model, which specifically includes: A six-dimensional parameterized feature vector is formed, and parameter normalization is performed on each dimension. The basic framework of the PINN model is constructed using a pre-established six-dimensional parameterized feature vector. The hidden layer is set as a deep multilayer perceptron, and Tanh is used as the activation function to determine the final output temperature. T It meets the requirements of thermophysical equations in mathematics; The physical residual terms of the unsteady heat conduction equation are added to the model to restrict the function expressed by the PINN model so that it conforms to the physical laws. We set a weighted loss function with clamping mechanism to ensure that the fitting accuracy in the high-temperature region is not affected by a large number of low-temperature samples. Step 5: Complete model training. Use the Adam adaptive moment estimator and StepLR step decay scheduler to adjust the parameters, ensuring the training process is stable and the model converges to the optimal solution.

[0005] In step one of this invention, the multi-disc brake system comprises a brake disc assembly consisting of one pressure disc, five moving discs, and four stationary discs, wherein all discs adopt a hollow cylindrical structure, and the material parameters of the brake discs are set as follows: density ρ 1700 kg / m³, specific heat capacity at constant pressure C p 1700 J / (kg·K), basic thermal conductivity k =50 W / (m·K), then the thermal diffusivity can be obtained. ; The unsteady heat conduction equation in cylindrical coordinates is shown below: (2), In the formula: T The transient temperature at any position on the brake disc; r These are radial coordinates, in meters (m). ρ This refers to the material density, expressed in kg / m³. The volumetric heat source intensity is expressed in W / m³. θ The circumferential angle is expressed in rad. x This is the axial coordinate, in meters (m). k r 、k θ 、k x These are the radial, circumferential, and axial thermal conductivity coefficients, respectively, in W / (m·K).

[0006] In step one of this invention, the heat transfer boundary of the partitioned area adopts a mixed boundary condition. The mixed boundary is controlled by the valueFraction, i.e., the mixing weight coefficient, which controls the mixing ratio of the Dirichlet boundary condition (fixed temperature) and the Neumann boundary condition (fixed gradient): valueFraction=1 corresponds to complete heat dissipation, i.e., the wall temperature is always equal to the ambient temperature of 300K, and valueFraction=0 corresponds to complete adiabatic heat transfer. The heat transfer characteristics of each region are as follows: Strong convection region at the end: The mixing weighting factor is set to 0.8, and the equivalent convective heat transfer coefficient is... ≈80 W / (m²·K); Axial transition region: The mixing weighting factor is set to 0.6, and the equivalent convective heat transfer coefficient is... ≈55 W / (m²·K); Central weak heat transfer zone: The mixing weighting coefficient is set to 0.4, and the equivalent convective heat transfer coefficient is... ≈35 W / (m²·K); Insulation boundary: The fine structure of the moving disk end face and spline groove is set as the insulation boundary; The total heat power generated during braking follows the kinetic energy consumption curve as shown in equation (5): (5), In the formula: η The heat-to-work conversion efficiency is set to 0.9, meaning that 90% of the kinetic energy is converted into heat energy in the brake disc. m For aircraft landing quality; v This refers to instantaneous velocity.

[0007] In step four of this invention, a six-dimensional parameterized feature vector is formed, and parameter normalization is performed on each dimension. Specifically, the input vector is expanded into a six-dimensional form that includes operating condition parameters: (6), In the formula: x, y, z The coordinates (m) describe the three-dimensional position within the brake disc assembly. t The time coefficient (s) represents the time progression of unsteady heat transfer. m Landing mass (kg); v 0 The initial landing velocity is (m / s), where m and v 0 together constitutes the parameterized variable; The following normalization calibration method was used: Input scaling: The six-dimensional input is linearly scaled to a specified range using Min-Max mapping; Temperature calibration: This involves setting the maximum true temperature value from the training set. T max =1808.31K is used as the upper limit for global normalization, and the output temperature is mapped to the interval [-1, 1].

[0008] In step four of this invention, the basic framework of the PINN model adopts a seven-layer architecture: one input layer, six hidden layers, and one output layer. The hidden layers use the Tanh activation function.

[0009] This invention is based on the isotropic assumption, and the thermal conductivity is taken as follows during the simulation stage. k =50 W / (m·K)) to construct the physical residual, and use the automatic differentiation technique of equation (7) to construct the unsteady-state heat conduction residual: (7), In the formula: ∇² T For the Laplace operator, describe the thermal diffusion process within the disk; The temperature prediction value (K) output by the neural network; The dynamic volumetric heat source load is determined by the operating parameters; ρ The density of the material is (kg / m³). C p Specific heat capacity at constant pressure (J / (kg·K)); t Time (s); k The thermal conductivity is denoted as W / (m·K), and the isotropic constant is taken as 50 W / (m·K) during the simulation stage.

[0010] This invention sets up a weighted MSE loss function based on the actual temperature gradient, and the weight function is defined as equation (8): (8), In the formula: clamp This is a clamping function; w min This is a bottom clamp to prevent the model from losing the entire background when fitting high temperatures; w max This is a top clamp, used to prevent excessive gradients at extreme points during aborted takeoff conditions, which could cause training oscillations. The weighted data loss expression is given by equation (9): (9), In the formula: L data This refers to the weighted data loss. N data This represents the total number of observed sample points; w i The temperature weight of the i-th sample point is calculated by equation (8); T̂ i Let be the predicted temperature (K) of the i-th sample point; T i For the first i The simulated true temperature (K) for each sample point.

[0011] This invention uses the Adam optimizer and divides the training into two stages: Phase 1 Global Search: Initial learning rate set to 1×102 -3 Set batch size to 8192; The second stage of fine-tuning: the StepLR scheduling strategy is enabled, the learning rate is decayed to 0.9 times the original value every 10,000 iterations, and the gradient norm of each parameter update is limited to a preset threshold of 1.0.

[0012] This invention takes the multi-disc braking system of civil aircraft as the research object, and uses OpenFOAM to conduct transient heat conduction numerical simulations under 18 typical braking conditions, forming a highly reliable temperature field dataset. Then, a clamp-weighted PINN model is constructed, which integrates spatial coordinates ( x,y,z The six-dimensional parameterized input layer, consisting of time and braking parameters including aircraft mass and initial landing velocity, enables the model to predict temperature fields under various conditions. Addressing the highly uneven temperature distribution in the simulation data, a temperature-weighted loss function is designed to improve the model's learning ability and prediction stability in high-temperature regions. After training, the model can be predicted using the initial landing velocity. v 0 and landing mass m This allows us to obtain the predicted three-dimensional temperature field under the corresponding working conditions, with a significant improvement in accuracy compared to the basic model. Attached Figure Description

[0013] Appendix Figure 1 This is a flowchart of the present invention.

[0014] Appendix Figure 2 This is the power attenuation curve under the reference operating condition in this invention.

[0015] Appendix Figure 3 This is a three-dimensional diagram of the dual-volume heat source in this invention.

[0016] Appendix Figure 4 This is a diagram showing the kinetic energy scaling factor and peak temperature distribution for 18 working conditions in this invention.

[0017] Appendix Figure 5 This is a comparison chart of data distribution before and after downsampling, in which... Figure 5 (a) is a comparison chart of the data distribution before and after downsampling. Figure 5 (b) is a comparison chart of the data distribution after downsampling.

[0018] Appendix Figure 6 This is a schematic diagram of the clamp-weighted PINN principle framework in this invention.

[0019] Appendix Figure 7 This is the temperature distribution histogram in this invention.

[0020] Appendix Figure 8 This is the baseline operating condition temperature cloud map in this invention. Figure 8 (a) shows the temperature contour map of the baseline operating condition at 5 seconds. Figure 8 (b) shows the temperature contour map of the reference operating condition at 10s. Figure 8 (c) shows the temperature contour map of the baseline operating condition at 15s. Figure 8 The middle (d) image shows the temperature contour map of the reference operating condition at 20 seconds. Figure 8 (e) is the temperature contour map of the reference operating condition at 25s. Figure 8 The middle (f) is the temperature cloud map of the reference operating condition at 30s.

[0021] Appendix Figure 9 This is a comparison chart of the scatter plot density of the PINN model prediction and the true value in an embodiment of the present invention, wherein... Figure 9 (a) is the basic model. Figure 9 (b) is the weighted model.

[0022] Appendix Figure 10 This is a comparison of histograms showing the prediction error distribution of the PINN model in embodiments of the present invention. Figure 10 (a) corresponds to the basic model. Figure 10 (b) corresponds to the weighted model.

[0023] Appendix Figure 11 This is a MAE bar chart for different temperature ranges in an embodiment of the present invention.

[0024] Appendix Figure 12 This is a volume rendering of the three-dimensional temperature field of the brake pads in an embodiment of the present invention.

[0025] Appendix Figure 13 This is a temperature cloud map of a VTK file slice in an embodiment of the present invention, wherein... Figure 13 (a) is the axial view. Figure 13 (b) is the radial diagram.

[0026] Appendix Figure 14 This is a spatial distribution cloud map of prediction error in an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] This invention addresses the problem of reconstructing the three-dimensional transient temperature field during the braking process of multi-disc carbon / carbon composite brake discs in civil aircraft. It employs clamp-weighted PINN optimization, and the specific process is as follows.

[0029] like Figure 1The method of this invention is supported by physical modeling and multi-condition simulation data, with clamped weighted PINN as the core. Through the synergistic effect of data fitting, physical residuals and temperature weighted loss, it improves the reconstruction accuracy and engineering inference efficiency of complex transient temperature fields.

[0030] Based on the actual structural characteristics of a multi-disc brake system, the following details the construction process, key geometric parameters, and coordinate system settings for a complete assembly model of a brake disc assembly consisting of one pressure disc, five moving discs, and four stationary discs. Part specifications: All discs are hollow cylindrical structures with an outer diameter of 0.25m, an inner diameter of 0.22m, and a radial wall thickness of 0.015m. Axial arrangement is used: the pressure disc thickness is 0.0235m (located at the bottom, with its lower surface contacting the hydraulic piston), the stationary discs have a uniform thickness of 0.0235m, and the moving discs have a thickness of 0.020m. The total assembly height is 0.2175m. Adjacent disc friction surfaces achieve zero-gap mechanical contact through overlapping fit, and concentric fit ensures alignment of the central axes of all discs.

[0031] The material parameters for the brake disc are set as follows: density ( ρ : 1700 kg / m³, corresponding to a medium-dense carbon brake disc after 3-4 CVI densification cycles, reflecting the lightweight and high-strength characteristics of the carbon fiber reinforced matrix. Specific heat capacity at constant pressure ( C p ): 1700 J / (kg·K), which is the effective average value in the temperature range of 300~1411K. At the extreme temperature rise of 1800K, the high specific heat capacity effectively prevents the material from instantaneously melting. Thermal conductivity: basic thermal conductivity. k =50 W / (m·K), simplified to an isotropic constant in the simulation stage. The thermal diffusivity can be calculated using these parameters.

[0032] The assembly imported from SolidWorks was meshed using the Salome platform. The resulting mesh was mainly composed of linear tetrahedral elements, with a total of 412,728 elements and 71,022 nodes.

[0033] Achieving unity between physical constraints and numerical solutions through coordinate transformation matrices: (1), In the formula: ρ The density of the material is (kg / m³). C p Specific heat capacity at constant pressure (J / (kg·K)); T Temperature (K); t Time (s); kq is the thermal conductivity (W / (m·K)), and the isotropic constant is taken as 50 W / (m·K) in the simulation stage; q''' is the volumetric heat source intensity (W / m³); ∇ T For temperature gradient (K / m); For thermal conductivity diffusion (W / m³); K Let be the thermal conductivity tensor.

[0034] The tensor-based expression described above can flexibly handle the performance differences of carbon / carbon materials in the radial and axial directions. For the axisymmetric cylindrical configuration of civil aircraft brake assemblies, this invention employs an unsteady heat conduction equation in cylindrical coordinates in the construction of the physical loss function: (2), In the formula: T The transient temperature at any position on the brake disc; r Radial coordinates (m); ρ The density of the material is (kg / m³). q ''' represents the volumetric heat source intensity (W / m³); θ Circumferential angle (rad); x Here is the axial coordinate (m); k r 、k θ 、k x The values ​​are the radial, circumferential, and axial thermal conductivity (W / (m·K)).

[0035] This equation relates to the temperature field. T Defined as a spatiotemporal variable The function. Considering the strict geometric symmetry of the brake disc in the circumferential direction and the uniform distribution of frictional heat load along the circumferential direction, the equation can be simplified to an axisymmetric form, effectively reducing the computational load of subsequent PINN automatic differentiation. The left side of the equation describes the accumulation of energy over time, while the diffusion and source terms on the right side represent the heat transfer mechanism in a complex anisotropic medium. Regarding surface convection heat transfer, considering that the brake assembly is located within the wheel hub space of the aircraft landing gear, there are differences in the heat dissipation environment between different parts. To achieve the universality of convection heat transfer under multiple operating conditions, this invention adopts a mixed boundary condition, the physical essence of which is based on Newton's law of cooling: (3), where: q conv For convective heat dissipation, the heat flux density is (W / m²). h The convective heat transfer coefficient is (W / (m²·K)). T surface The surface temperature of the disk (K); T ambient The ambient temperature is set to 300K.

[0036] A weighting coefficient is applied spatially to the equivalent heat transfer coefficient. The mixing boundary controls the mixing ratio of the Dirichlet boundary condition (fixed temperature) and the Neumann boundary condition (fixed gradient) through valueFraction (i.e., the mixing weighting coefficient, ranging from 0 to 1): valueFraction = 1 corresponds to complete heat dissipation (wall temperature is constant equal to the ambient temperature of 300K), and valueFraction = 0 corresponds to complete adiabatic conditions. The heat transfer characteristics of each region are as follows: The end high-convection zone (near the pressure plate and moving plate 5): Due to direct impact from the hub airflow, it has the strongest heat transfer capacity. The mixing weight coefficient is set to 0.8, and the equivalent convective heat transfer coefficient is... ≈80 W / (m²·K).

[0037] Axial transition zone (moving disks 2-4 and adjacent stationary disks): Convection enters through the gaps between the disks, but is partially blocked. The mixing weight coefficient is set to 0.6, and the equivalent convective heat transfer coefficient is... ≈55 W / (m²·K).

[0038] Central weak heat transfer zone (around stationary discs 2-3): Located in the core area of ​​multiple disc stacks, airflow is severely obstructed. The mixing weighting factor is set to 0.4, and the equivalent convective heat transfer coefficient is... ≈35 W / (m²·K).

[0039] Adiabatic boundary: The end face of the moving disk and the spline groove and other fine structures are set as adiabatic boundaries. These areas have extremely small surface areas and extremely poor airflow conditions, so their heat transfer contribution is negligible.

[0040] High-temperature radiative heat transfer is described by the Stefan-Boltzmann law (4): (4), where: q rad The heat flux density for radiative heat dissipation (W / m²). ε For surface emissivity, the C / C composite material is taken as 0.7~0.9 (dimensionless); σ is the Stefan-Boltzmann constant, taken as 5.67×10⁻ 8 W / (m²·K 4 ).

[0041] Its effect is directly implied in the valueFraction parameter of the mixing boundary and is not calculated separately. Under the condition of a peak temperature of about 1400K, the proportion of radiative heat dissipation in the total heat dissipation is limited (estimated to be no more than 15%), and its impact on the temperature field distribution trend is within the range of engineering accuracy.

[0042] The contact surface between adjacent discs is the main path for heat conduction and dissipation. Under hydraulic pressure of several megapascals, the friction surfaces are pressed together, and the macroscopic contact thermal resistance is negligible. Salome Partition conformal processing is used to achieve node sharing to simulate ideal contact. To address the intense frictional heat generation between the moving and stationary discs, load modeling is performed based on the extreme conditions of a typical emergency landing of a civil aircraft. The total heat power generated during braking is... Follow the kinetic energy consumption curve (5): (5), where: η The heat-to-work conversion efficiency is set at 0.9, meaning 90% of the kinetic energy is converted into heat energy in the brake disc. m For aircraft landing quality; v This refers to instantaneous velocity.

[0043] Baseline operating condition ( m =22000 kg, v At 0 = 72 m / s, the peak heat flux density of the friction surface can reach 5 × 10⁻⁶ m / s. 6 W / m². To adapt to PINN's multi-condition inference requirements, this example constructs 18 groups of different masses. m (Range 18000~28000 kg) and initial velocity v An orthogonal parameter space consisting of 0 (range 67~77 m / s). The kinetic energy distribution of each working condition is in the range of 0.6~1.4 times that of the baseline working condition, covering a variety of operating scenarios from light load and low speed to heavy load and high speed.

[0044] Construction of the volumetric heat source scheme for temperature field simulation: During braking, frictional heat is generated only in a thin layer of approximately 2 mm thickness in the near field of 10 friction surfaces, and the thermal power continuously decreases with braking time: from an initial value of approximately 2.5 × 10⁻⁶. 9 W / m³ decays to approximately 1.1 × 10⁻⁶ at the end of 30 s. 9 W / m³ Figure 2 The heat source scheme must simultaneously meet three technical conditions: precise spatial limitation (acting only on near-field elements of the friction surface), dynamic and time-varying power (following the law of kinetic energy decay), and stable compatibility with the solver (not introducing additional numerical instabilities).

[0045] Therefore, a volumetric heat source combining a semiimplicitsource model was chosen: the heat source was transformed from surface excitation on the boundary into volumetric excitation within the domain, and the spatial range and power curve were flexibly controlled through an external dictionary file, completely avoiding compilation dependencies and permission restrictions. For spatial layout, the topoSet tool was used to extract one layer of elements from each side of the friction surface to construct the cell zone. The first nine groups of friction surfaces generated the main heat source, containing approximately 100,000 elements. The tenth group, located at the top of the moving disk 5, generated an independent heat source of approximately 500 elements. These two parts combined constitute a dual-volume heat source system (e.g., ...). Figure 3 ).

[0046] In terms of time implementation, baseline operating conditions are defined through a powerTable file. m =22000 kg, v The temperature rise rate at each moment was obtained by using 0=72 m / s. The values ​​were determined through theoretical estimation and multiple simulation calibrations: the temperature rose to 1363K at the end of 5 seconds under the baseline condition and reached a peak of 1411K at about 10 seconds. The temperature evolution was highly consistent with the physical expectations. Two independent semi-implicit source models were configured in fvModels, with volumeMode set to specific (specific power mode), which were applied to the dual heat source regions respectively. The powerTable was referenced and an out-of-bounds clamping function was set. If the time exceeded the range of 0 to 30 seconds, it was clamped to the boundary value.

[0047] The advantages of this scheme are as follows: First, it precisely limits the spatial range, ensuring that the heat source only acts on a thin layer of approximately 100,000 units near the friction surface, without affecting the entire region; second, it allows for flexible setting of power levels, as the time-varying curve can be adjusted using an external dictionary file without modifying the solver's own code; third, it is perfectly compatible with existing solvers, as the semi-implicit source model is a standard source term model included with OpenFOAM and can directly interface with the Laplace solver; and fourth, the operation process is repeatable, with all configurations written in an ASCII format dictionary file, facilitating version control and parameter adjustment.

[0048] To enable the PINN model to perceive different flight conditions, this invention constructs 18 sets of parametric flight condition matrices using aircraft mass m (18000~28000kg) and initial landing velocity (67~77 m / s) as variables. The frictional heat power of each set of conditions is bound to the baseline condition through a kinetic energy scaling factor. The principle of this method is: under the condition that the braking system design, friction area, and materials are the same, the frictional heat power is proportional to the kinetic energy to be consumed by the aircraft. The difference in temperature rise rate between different conditions is determined only by the initial kinetic energy. This method ensures that the physical evolution characteristics of the 18 sets of conditions are the same: the temperature change of each set of conditions over time follows the same decay curve shape, with only the amplitude scaled proportionally to the kinetic energy. The kinetic energy scaling factor R for each condition is approximately 0.70~1.45, and the peak temperature ranges from 1100~1800K. Figure 4 ).

[0049] After the simulation is completed, the cell center coordinates and the corresponding temperature values ​​at the time need to be extracted from the ASCII format field file output by OpenFOAM to construct a structured dataset for PINN training.

[0050] The temperature field file generated by OpenFOAM contains coordinates and temperature vectors for millions of nodes. In the WSL2Ubuntu environment, firstly, cell center coordinate component files are generated in each case directory. Then, spatial field extraction is performed on six key time sections (5, 10, 15, 20, 25, and 30 s) within the entire braking cycle (0–30 s). The data volume extracted for a single operating condition is approximately 4.46 million cells × 6 time steps ≈ 26.76 million rows, with a CSV file size of approximately 400–600 MB. The total data for all 18 operating conditions is approximately 8–10 GB.

[0051] The total dataset exceeds the memory capacity of conventional deep learning frameworks on ordinary GPUs. This invention employs a spatially uniform downsampling method: following the cell index, one sample point is extracted from every 10 cells, while filtering out non-physical regions (such as invalid cells outside material boundaries). The downsampled training dataset contains approximately 5 million rows of data and can be directly loaded and trained in a standard runtime environment. Details lost during uniform downsampling can be partially recovered using PINN's physical loss across the entire region. Thus, even with only 5 million rows of training data, the model can maintain the spatial continuity and physical accuracy of temperature field predictions. Figure 5 The downsampled data was randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0052] This invention proposes a PINN model combining a clamping weighted loss function: on the one hand, it utilizes simulation data to learn the continuous mapping law of the temperature field with operating conditions; on the other hand, it uses the clamping weighted loss function to address the uneven heating and cooling of the data, improving the reconstruction quality of the friction surface and adjacent areas. The model embeds the transient heat conduction equation as a soft constraint into the loss function, achieving dual-driven development by data and physics. This allows the network to follow physical laws while fitting the data, thereby supporting reasonable extrapolation and ensuring that the temperature field prediction under operating conditions without simulation data remains within a reasonable range.

[0053] This invention sequentially forms the various functional parts of the clamp-weighted PINN model, and the specific steps are as follows: First, a six-dimensional parameterized feature vector is formed, namely spatial coordinates. x, y, z ,time t Aircraft quality m Initial landing velocity v The combination of zeros yields a six-dimensional feature vector, enabling the model to perceive different operating conditions. Next, parameter normalization is performed on each dimension, unifying inputs with different dimensions to the same scale. This step also lays the dynamic foundation for the subsequent use of the clamped weighted loss function method.

[0054] Secondly, the basic framework of the PINN model is constructed. Using the established six-dimensional parameterized feature vector, the hidden layers are set as deep multilayer perceptrons, and Tanh is used as the activation function. The Tanh function ensures the propagation of the second derivative within the model, providing an effective gradient signal to the physical constraints, thus ensuring the final output temperature... T It meets the requirements of the thermophysical equations in mathematics.

[0055] Furthermore, the physical residual term of the unsteady heat conduction equation is added to the model. The heat conduction equation serves as a soft constraint embedded loss function, restricting the function expressed by the PINN model to conform to physical laws. The physical residual term also ensures that the model has reasonable extrapolation capabilities.

[0056] Then, a weighted loss function with a clamping mechanism was designed to address the uneven distribution of hot and cold samples in the brake disc simulation dataset, ensuring that the fitting accuracy in the high-temperature region is not affected by the large number of low-temperature samples. In addition, an adaptive moment estimation (Adam) optimizer and a step decay scheduler (StepLR) were used to adjust the parameters, and gradient pruning was employed to ensure the stability of the training process, ultimately leading the model to converge to the optimal solution. After training, the model performance was tested, primarily evaluating prediction accuracy and computational power, thereby demonstrating the effectiveness of neural networks in temperature field simulation.

[0057] Construction of a six-dimensional parameterized input space and parameter normalization: This invention constructs a six-dimensional parameterized input space and normalizes parameters. Expanded to a six-dimensional form including operating condition parameters, it is as follows: (6), In the formula: x, y, z The coordinates (m) describe the three-dimensional position within the brake disc assembly. t The time coefficient (s) represents the time progression of unsteady heat transfer. m Landing mass (kg); v 0 The initial landing velocity is (m / s), where m and v 0 together constitutes the parameterized variable.

[0058] Will and The physical basis for embedding into the input space lies in the fact that the thermal response of the brake disc is essentially affected by the initial kinetic energy. Dominance. Under the premise of identical braking system design, friction area, and material properties, the temperature field under different operating conditions exhibits remarkably similar spatial distribution: the high-temperature zone is consistently concentrated in a ring-shaped region near the friction surface, and the temperature decrease pattern is largely consistent along both the axial and radial directions. The main difference lies in the fact that the temperature scales proportionally with the initial kinetic energy. The neural network, through training, learns the aircraft's mass... m and initial landing velocity The mapping relationship between the model and the temperature field enables a single model to cover all 18 typical operating conditions, giving it the ability to generalize predictions across parameters.

[0059] Considering the significant differences in magnitude between mass, velocity, and spatial coordinates, the following normalization and scaling methods were adopted: Input scaling: The six-dimensional input was linearly scaled to a specified interval using a Min-Max mapping. The normalization range for each aspect was determined during the previous data extraction stage. Temperature scaling: The maximum true temperature of the training set was used as the calibration value. T max =1808.31K was used as the upper bound for global normalization, and the output temperature was mapped to the interval [-1, 1]. This upper bound includes the highest temperature that will occur in all 18 operating conditions, ensuring that the network output is always within the linear sensitivity range of the Tanh activation function.

[0060] Constructing the PINN model: Addressing the highly nonlinear characteristics of the brake disc temperature field: The near-field temperature gradient of the friction surface reaches as high as 10. 5 The temperature variation is on the order of K / m, and the temperature change is gradual deep within the disk. Therefore, this invention constructs a deep multilayer perceptron as a surrogate model. The network adopts a 7-layer architecture: one input layer (6 neurons), six hidden layers (256 neurons per layer), and one output layer (1 neuron), with a total of approximately 460,000 parameters. The considerations for this design are as follows: Network depth: 6 hidden layers. This deep network can simultaneously capture two distinct temperature change characteristics: the extremely rapid spatial temperature rise gradient near the friction surface and the relatively gentle temperature distribution inside the disk. The 6-layer depth represents a balance between expressive power and training feasibility.

[0061] Hidden layer width: 256 neurons, with a six-dimensional input space complexity that is much higher than the three-dimensional or four-dimensional input commonly used in traditional PINNs. It requires a sufficiently wide hidden layer to represent the complex cross-coupling relationship between spatial coordinates, time, and operating parameters.

[0062] The hidden layers use a uniform Tanh activation function. In the PINN mode, the choice of activation function affects the expressive power of forward propagation and also directly changes the feasibility of calculating the physical loss in backpropagation.

[0063] Operational constraint: Linear rectifier units cannot be used as activation functions in physical information neural networks.

[0064] Physical basis: The formation of the physical loss function involves the calculation of the second-order spatial derivative. If ReLU is used in the hidden layer, the terms in the PDE residuals will be... The computation will be zero across all input spaces; the thermal diffusion term disappears, and the physical constraints become completely ineffective. The Tanh function is continuously differentiable everywhere, and its first derivative equals... The second derivative equals Both of these are non-zero and smooth functions, which can provide meaningful gradient signals to physical constraints, enabling the optimizer to update parameters along the direction that satisfies the heat conduction equation.

[0065] This invention is based on the isotropic assumption (the thermal conductivity is taken during the simulation stage). k =50 W / (m·K)) to construct the physical residual. The unsteady-state heat conduction residual is constructed using the automatic differentiation technique of equation (7). : (7), In the formula: ∇² T For the Laplace operator, describe the thermal diffusion process within the disk; T̂ The temperature prediction value (K) output by the neural network; q''' The dynamic volumetric heat source load is determined by the operating parameters; ρ The density of the material is (kg / m³). C p Specific heat capacity at constant pressure (J / (kg·K)); T Temperature (K); t Time (s); kThe thermal conductivity is denoted as W / (m·K), and the isotropic constant is taken as 50 W / (m·K) during the simulation stage.

[0066] Automatic differentiation accurately calculates partial derivatives of each order along the chain rule of the computation graph, completely avoiding the truncation error of numerical difference. This technical advantage ensures that the model still has physically consistent extrapolation capabilities inside a sensorless disk. Even without internal observation data, the global constraints of the PDE residuals can confine the temperature field within a function space that satisfies energy conservation.

[0067] The temperature distribution in the brake disc simulation data is severely uneven: samples in the low-temperature region (300~500K) account for about 70%, while those in the high-temperature region (>1000K) account for less than 5%. Under uniform MSE loss, the cumulative error of the massive low-temperature samples dominates the optimization direction, resulting in a systematic underestimation of the high-temperature region prediction. The MAE of the basic model in the high-temperature region is about 65K, and the MAE of the entire region is about 34K.

[0068] To avoid excessively high fitting errors in the temperature field of high-temperature regions due to insufficient sample size, this invention designs a weighted MSE loss function based on the actual temperature gradient. The weighting function is defined as equation (8). (8), In the formula: clamp This is a clamping function; w min This is a bottom clamp to prevent the model from losing the entire background when fitting high temperatures; w max This is a top clamp, used to prevent excessive gradients at extreme points during aborted takeoff conditions, which could cause training oscillations.

[0069] The weighted data loss expression is given by equation (9): (9), In the formula: L data This refers to the weighted data loss. N data This represents the total number of observed sample points; w i The temperature weight of the i-th sample point is calculated by equation (8); T̂ i Let be the predicted temperature (K) of the i-th sample point; T i For the first i The simulated true temperature (K) for each sample point.

[0070] Under the weighted MSE mechanism, the pulling force of a 1500K high-temperature sample point on the optimization direction is roughly equivalent to the combined effect of five 300K low-temperature sample points. This forces the optimizer to devote more attention to the high-temperature region during training, thereby improving the reconstruction quality of the friction surface and its adjacent areas. The optimization effect of the weighted model is significant, with the global mean absolute error (MAE) decreasing from approximately 34K to 13.12K, and the MAE in the high-temperature region decreasing from approximately 65K to 28K (see Table 1).

[0071] Table 1 Performance Comparison of Basic Model and Weighted Model

[0072] Training convergence optimization strategy: The Adam optimizer is selected, whose adaptive learning mechanism can effectively address the imbalance problem that occurs when the sensitivity of PDE residuals spans several orders of magnitude in different spatial regions. Based on this, training is divided into two stages: 1. Global search: The initial learning rate is set to 1×10⁻⁶. -3 The batch size was set to 8192. A high initial learning rate allowed the optimizer to quickly traverse the flat regions of the loss surface and lock into the globally optimal region. Second, fine-tuning: the StepLR scheduling strategy was enabled. Every 10,000 iterations, the learning rate decreased to 0.9 times its original value. The learning rate was gradually reduced in the later stages of training, allowing the optimizer to perform fine-grained local searches near the minimum of the loss surface, avoiding continuous oscillations near the optimal solution and preventing convergence. The total number of iterations was 50,000, with each training session taking approximately 2.8 hours.

[0073] Under the weighted mechanism, the loss contribution of samples in the high-temperature region is amplified, thereby increasing the gradient norm of the loss function with respect to the network parameters. If not restricted, individual extremely high-temperature samples (such as extremely thin layer units with a thickness of less than 1 mm near the peak temperature of 1411 K) may lead to excessive parameter updates in a single step, causing training oscillations or even divergence.

[0074] The system implements gradient pruning, limiting the gradient norm to within 1.0. This operation limits the single-step update magnitude while ensuring the correct direction of loss descent, preventing interference from extreme samples and guaranteeing stability over 50,000 long-cycle training sessions.

[0075] The performance of this invention will be verified and analyzed below: Baseline operating condition (aircraft mass) m =22000 kg, initial landing velocity v The speed of 0 = 72 m / s corresponds to the extreme braking scenario when a passenger plane aborts takeoff. At this time, the heat flux density on the friction surface reaches a maximum of 5 × 10⁻⁶ m / s. 6 W / m², the temperature field completely records the process of rapid rise from room temperature, stagnation at the peak, and then natural cooling. (See W / m²) Figure 8Simulation results show that the highest temperature of the disc increased from the ambient temperature of 300K to 1363K within 5 seconds, reaching a peak of 1411K at around 10 seconds. Then, the braking power gradually decreased, and the system began cooling, dropping to 976K by the end of 30 seconds. The main temperature parameters at each time point are shown in Table 3.

[0076] In the initial braking phase, frictional power is high, with a temperature rise rate reaching 1982 K / s. Heat generation far exceeds convective heat dissipation, and heat rapidly accumulates on the friction surface without sufficient time to dissipate. Around the 10th second, the temperature reaches its peak, at which point heat generation and dissipation reach a dynamic equilibrium: the temperature rise rate drops to about 80% of the peak value, approximately 1586 K / s, while the heat dissipation capacity continues to increase with the average temperature of the disc. Afterward, frictional power decreases, heat dissipation becomes dominant, and the temperature slowly declines. The entire process is consistent with the braking power decay pattern, verifying the rationality of the heat source scheme and boundary conditions.

[0077] Table 2 Key Temperature Indicators at Different Times Under Baseline Operating Conditions

[0078] Spatial thermal gradient distribution characteristics: At the temperature peak time (t=10s), the analysis was conducted on the middle moving disk (axial position x=0.1105~0.1305m). Along the thickness direction, the temperature decreases symmetrically with the central layer as the boundary. The temperature of both friction surfaces is approximately 1390K (1117℃), while the central layer is approximately 1050K (777℃), with a surface-to-core temperature difference of approximately 340℃. There are two reasons for this: first, the thermal conductivity of the carbon / carbon composite material is lower in the thickness direction than in the planar direction (anisotropy ratio 1.5~3), limiting the inward transfer of heat; second, the braking time is only 10s, and the heat is concentrated on the surface layer and does not penetrate the 20mm disk thickness. The temperature difference between the two friction surfaces is only 5℃, which is negligible, indicating that the heating is uniform from top to bottom. Along the radial direction, the temperature decreases from the outside to the inside, with approximately 1410 K (1137℃) on the outer diameter side (r=0.125m) and approximately 1180 K (907℃) on the inner diameter side (r=0.110m), a difference of about 230℃. This is because the outer ring has a higher linear velocity, resulting in stronger frictional heat generation per unit area, and the heat generation rate dominates the temperature distribution during the rapid heating phase. This pattern is consistent with the actual wear and heating characteristics of multi-disc brakes, verifying the rationality of the volumetric heat source application method of this invention.

[0079] Multi-disc coupled heat transfer and statistical characteristics of 18 operating conditions: The brake disc assembly employs a ten-disc alternating stacked structure, and the thermal coupling between the discs significantly impacts the overall temperature distribution. From a thermal coupling perspective, the stationary and moving discs in the middle of the assembly are heated on both sides, resulting in limited heat dissipation paths and a tendency for heat to accumulate, leading to the most significant effect. In contrast, the discs at the ends are heated on only one side, with the other side facing the cavity inside the wheel hub, resulting in relatively better heat exchange conditions. During the natural cooling phase, the temperature difference between the discs gradually decreases. By the end of 30 seconds, the temperature difference across the entire disc assembly has decreased from approximately 100K at its peak to approximately 40K. This indicates that solid-state heat conduction between the discs plays a dominant role in the later stages of cooling, causing the temperature to become more uniform. This trend towards temperature homogenization also indirectly verifies the rationality of the conformal mesh processing method: if there is significant numerical thermal resistance at the contact surface, heat transfer between the discs will be artificially blocked, and the process of temperature uniformity will deviate significantly from physical reality. All 18 sets of parameterized driving conditions converged smoothly to 30 seconds without any non-physical results such as divergence, negative temperatures, or abnormal hot spots. Peak temperature and kinetic energy ratio coefficient for each group of operating conditions Positive correlation: R It increased from 0.70 to 1.45.

[0080] The peak temperature monotonically increases from approximately 1100K to approximately 1800K. Simulation results for representative operating conditions are shown in Table 4.

[0081] Table 3. Statistics of Simulation Results for Representative Working Conditions

[0082] As the initial kinetic energy increases, the peak time is advanced from approximately 12 seconds to 8 seconds. This is because the frictional power is higher and the temperature rises faster under high kinetic energy conditions, reaching the heat generation-dissipation equilibrium point earlier. The residual temperature at 30 seconds also increases from 760K to 1180K with increasing kinetic energy, indicating that more heat is retained under high-energy conditions within the same cooling time, which conforms to the law of conservation of energy. All examples converge stably, providing a reliable and comprehensive data foundation for training the PINN model.

[0083] Accuracy evaluation of the clamped weighted PINN model: After training, the weighted model demonstrated good performance on a test set including all 18 working conditions. The model's mean absolute error (MAE) was 13.12K, root mean square error (RMSE) was 19.90K, and coefficient of determination was [missing information]. R The value reached 0.9793. The main performance indicators are listed in Table 5.

[0084] Table 4 Performance metrics of the weighted PINN model test set

[0085] Global statistical indicators and goodness of fit: In engineering applications: RA mean absolute error (MAO) exceeding 0.97 indicates that the neural network has accurately captured over 97% of the thermal evolution characteristics of the brake disc assembly. The MARCH value of 13.12K, with a relative error of approximately 0.9% across the entire temperature range of 1500K, is suitable for structural safety assessments in engineering applications. Although there is a local bias of 218.10K at the maximum temperature gradient, this region has sparse samples, accounting for less than 0.5% of the entire dataset, and exhibits strong physical nonlinearity, such as... Figure 9 As shown, these random residuals do not change the physical consistency of the temperature field reconstruction.

[0086] Error Spatial Evolution: Statistical analysis of the prediction errors on the test set shows that the residuals as a whole follow an approximately normal distribution with a mean of zero. Approximately 68% of the sample errors are within ±15K, and approximately 95% are within ±40K (see [link to relevant documentation]). Figure 10 The symmetric distribution centered at zero proves that the model has no systematic overestimation or underestimation bias.

[0087] Figure 11 As shown, the MAE values ​​for the low-temperature region (<500K), the medium-temperature region (500–1000K), and the high-temperature region (>1000K) are 11K, 21K, and 41K, respectively, with the error increasing with increasing temperature. This is attributed to two factors: first, the low-temperature region has abundant data and a gradual temperature change, resulting in the highest fitting accuracy; second, the high-temperature region is concentrated near the friction surface, with temperature gradients reaching hundreds of thousands of K / m, where even small spatial differences correspond to temperature differences of tens of K, and the training samples in this region account for less than 5%, naturally limiting accuracy. Nevertheless, the MAE of 41K in the high-temperature region, relative to the range of 1000–1800K, still maintains a relative error within 2%.

[0088] Spatially, the error is mainly concentrated in the region with the largest temperature gradient near the friction surface. Beyond 5 mm from the friction surface, the error quickly drops below 10 K. Inside the disk, in areas far from the friction surface, the error generally decreases to below 5 K. This spatial error variation clearly shows that the high-gradient region limits further improvement in overall accuracy. This also demonstrates the necessity of a temperature-weighted strategy, which, while maintaining a stable fit across the entire field, consciously pushes the optimization weights towards the main monitoring areas with large temperature gradients and of primary engineering concern.

[0089] Table 5 shows that after adding the temperature-weighted scheme, the MAE in the high-temperature region decreased from approximately 65K to 28K, a reduction of 57%, and the overall MAE decreased from approximately 34K to 13.12K, demonstrating a significant optimization effect. This method solves the problem of poor fitting effect and large prediction error of high-temperature data points caused by uneven data distribution, ensuring that safety critical points can be accurately captured.

[0090] Table 5 Model Performance Comparison

[0092] Three interpolation conditions (mass 18,000–28,000 kg, initial velocity 67–77 m / s) not used in training were selected for testing. The MAE remained stable at 13–16 K, and the R² was higher than 0.975, comparable to the accuracy of the training conditions. This indicates that the model did not overfit the training samples, but effectively captured the continuous variation of the temperature field with the operating conditions through the six-dimensional input layer, demonstrating adaptive inference capability across operating conditions.

[0093] When the input exceeds the range of the training set, the model can still perceive the changing trend, but the accuracy decreases. The MAE for heavy-load extrapolation (30000 kg, 80 m / s) is approximately 35 K, with an R² of approximately 0.92; the MAE for light-load extrapolation (16000 kg, 65 m / s) is approximately 25 K, with an R² of approximately 0.94. The decrease in accuracy is due to the extrapolated operating temperature range exceeding the training coverage area, increasing the uncertainty of the model when facing unknown temperature ranges.

[0094] In actual deployment, PINN should be applied within the designed operating range. If it needs to be extended to more extreme scenarios, the step size of the kinetic energy ratio parameter can be refined and corresponding operating conditions can be added to broaden the training set coverage, while keeping the model structure and training strategy unchanged.

[0095] In terms of inference speed, OpenFOAM takes about 4 hours to complete a 30-second transient simulation (4 cores in parallel), while PINN only takes about 20 seconds to complete a full-field inference of 4.46 million units, a speedup of 720 times.

[0096] Table 6 Comparison of computational efficiency between PINN inference and OpenFOAM simulation

[0097] Ablation experiments and methodological discussion: To quantify the independent contribution of the temperature-weighted loss strategy and its supporting measures to the model performance, this example uses the weighted PINN full configuration model as the baseline, designs three ablation experiments, removes each improvement measure one by one and records the amount of accuracy degradation, and the results are shown in Table 7.

[0098] Table 7 Ablation Experiment Results

[0099] This invention uses ablation experiments to verify that the three proposed improvement measures can produce a progressive synergistic effect, and proves that these measures are not effective in isolation. (1) Increasing the temperature weight can become the main driving force for improving accuracy. If this strategy is removed, the mean absolute error (MAE) of the model will increase from 13.12K to about 21K, and the performance will decrease by about 60%. The results show that setting high-temperature samples as high weights will significantly improve the situation of severe data skewness. (2) When the normalization range is insufficient (upper bound 1411K vs 1808K), the MAE rises to 34K, a regression of 159%, because the high-temperature samples fall into the Tanh saturation region, causing the gradient to disappear. Temperature weighting must be used in conjunction with a sufficient normalization range to be effective. (3) Appropriately reducing the weight of physical loss can eliminate the contradiction between improvement objectives. When the weight of physical residual is changed from 10-3 to 0.1, the MAE increases to 16K, a degradation of 22%. Analysis shows that if the physical loss from uniform sampling across the entire area has a larger weight than the weighted data loss from high-temperature regions, the two objectives will conflict in their direction of improvement. Keeping the physical loss at a low level better meets the engineering objective of this invention: prioritizing accuracy in the main high-temperature regions.

[0100] This invention uses the scientific computing post-processing platform ParaView as a visualization tool to convert the discrete numerical results output by the PINN model into 3D mesh data readable by ParaView. The first step, model inference and feature registration, utilizes a self-written prediction script to read the mesh information and cell center coordinate files from the OpenFOAM case directory. The cell center coordinates serve as the spatial components of the input vector. A pre-trained parametric weighted PINN weight file is then loaded, and the data is applied to a specified aircraft weight. m and initial landing velocity v The first step involves synchronous inference of the full-field temperature at 0°C. A single prediction can reconstruct a complex temperature field of 4.46 million grid cells, with each inference taking approximately 20 seconds. The second step is to encapsulate the VTK mesh file. The scalar temperature field obtained from the inference needs to correspond with the original geometric topology and be encapsulated together. Using the foamToVTK tool in OpenFOAM software, the temperature field data obtained from the model prediction is integrated and then converted into a VTK mesh file. This encapsulation method preserves the contact surface information inside the assembly, ensuring consistency between the visualization results and the original numerical solutions. The third step is to organize the time series. Throughout the entire dynamic braking cycle, the time series prediction script generates VTK sequence files with timestamp suffixes and uniformly names these files. They are then placed in the same directory. ParaView can automatically identify all time steps and combine the various dispersed transient sections into a three-dimensional temperature field model including the time series, which provides data support for the subsequent analysis of the overall dynamic evolution process.

[0101] Core visualization functionality includes volume rendering and internal thermal accumulation diagnosis. Volume rendering is crucial for detecting the temperature field distribution inside a disk. The basic idea is to convert each sample point in the 3D volume data into a combination of color and opacity based on its temperature. As light passes through the volume data, the color accumulates along the path, ultimately generating a semi-transparent rendered image (e.g., ...). Figure 12 ).

[0102] ParaView's non-linear transparency transfer function is used to differentiate the display effects of each temperature zone: high transparency is set for the low-temperature zone to make the background semi-transparent without obscuring the internal structure; high opacity is set for the high-temperature core zone above 1200K to clearly present the high-temperature annular band under the friction interface; and a smooth transition is used for the intermediate temperature zone to maintain spatial continuity and achieve good visual differentiation. In addition, ParaView's built-in time series function is used to organize multi-time-step VTK files into a continuous playback animation to fully present the dynamic evolution of the temperature field throughout the braking process. At the beginning of braking, a thermal shock zone quickly forms near the friction surface, and the color mapping suddenly changes from dark blue at room temperature to orange-red at high temperature; during the peak maintenance phase of 10 to 15 seconds, the high-temperature zone expands to its widest extent, but the interior of the disc remains predominantly low-temperature, and the steep temperature gradient along the axial direction is still clearly visible; after 15 seconds, as the rotation speed gradually decreases, the input of friction power further weakens, and the focus of visualization shifts to the process of heat diffusion towards the axial heat dissipation edge and radial disc center. The high-temperature zone slowly shrinks, and the overall color begins to transition from warm to cool tones.

[0103] This invention performs slice analysis in multiple directions, including axial slices and radial slices. The axial slice is shown below. Figure 13 As shown in (a), observing the heat decay process from the friction surface into the disk body reveals the heat diffusion along the thickness direction. The cross-section shows good symmetry in temperature rise at the contact interface between the moving and stationary disks in key areas of the assembly, indicating that the volumetric heat source scheme used is accurate in multi-disk coupled modeling. Radial cross-sections are shown below. Figure 13 As shown in (b), the temperature difference between the outer and inner diameters of the disc is quantified. It can be seen that the temperature on the outer diameter side is much higher than that on the inner diameter side, indicating that the friction linear velocity is different at the outer diameter. There is a linear relationship between the temperature and the radius of the brake disc, and the outer diameter area is the boundary of high temperature load.

[0104] Spatial field visualization diagnosis of prediction residuals: To verify the actual correction effect of the temperature weighting strategy, this invention constructs a spatial scalar field of prediction errors. In ParaView, two types of VTK files are loaded simultaneously: one is the temperature field predicted by PINN, and the other is the ground truth data obtained from OpenFOAM simulation. Both sets of data share the same mesh, and are directly overlaid. The temperature difference at each node is calculated using the Calculator filter to obtain the error field. Volume rendering of the error field then allows direct visualization of how the error is distributed in the high-gradient region of the friction surface and the low-gradient region inside the disk. Figure 14 ).

[0105] Visual diagnostic results show that large residuals greater than 100K only appear in the high-temperature gradient zone within a 2-3 mm range near the friction surface, and are most concentrated on the outer diameter side at the highest temperature t=10 s. Beyond 5 mm from the friction surface, the error field quickly exhibits a uniform deep blue color, mainly because the errors are mostly below 5K. In the overall temperature distribution, over 95% of the predicted and simulated values ​​match very well. The spatial error distribution characteristics demonstrate the effectiveness of the algorithm. Furthermore, these diagnostic results also suggest that the density of training data can be increased in the high-temperature gradient region, or more refined adaptive sampling points can be arranged.

[0106] The analysis of time-series animation and volume rendering revealed the following key patterns: (1) Temperature distribution characteristics: The brake disc assembly is hot in the middle and cold at both ends. The cooling rate of the static / dynamic disc in the middle is significantly lower than that of the disc at the end due to double-sided friction heating and limited heat dissipation. This verifies the physical rationality of setting the mixing weight of the central area to 0.4 in this invention, and provides an engineering basis for optimizing ventilation and heat dissipation in the middle. (2) Damage sensitive area identification: The high temperature peak is concentrated in a narrow area of ​​about 10mm in the outer diameter friction ring. The temperature gradient is large here, and the thermal stress and mechanical load are superimposed. It is a high-risk area for oxidation degradation and fatigue failure of carbon / carbon composite materials. Based on the spatial coordinates and temperature information provided by visualization, it can guide the sensors to be deployed in this area or formulate a precise ground inspection plan. (3) Technical closed loop: The PINN lightweight inference engine realizes a complete technical closed loop from numerical prediction to engineering graphic diagnosis by connecting the VTK data with the mature rendering capabilities of ParaView.

[0107] The results show that the improved PINN model of this invention has a global mean absolute error of 13.12K and a coefficient of determination of [missing information]. R The accuracy reached 0.9793, which is a significant improvement compared to the basic model.

Claims

1. A method for predicting the state of landing gear brake discs based on clamp-weighted PINN, characterized in that, Includes the following steps: Step 1: Establish a physical model of the brake disc assembly, perform conformal treatment on the contact surfaces between the discs, and set the unsteady-state heat conduction equation, zoned heat transfer boundaries, and material parameters; Step 2: Construct the temperature field simulation control equations and the simulation volume heat source system. The time-varying volume heat source is implemented using a semi-implicit source model. The power time-varying curve is defined through an external dictionary file, and an over-boundary clamping function is set. The volume heat source acts on a thin layer region composed of one unit on each side of the friction surface, forming a dual volume heat source system. Step 3: Construct a braking condition matrix dataset through three-dimensional temperature field simulation. Design the condition matrix with landing mass and initial landing velocity as variables, complete the transient simulation, extract the temperature field at key moments and downsample it to form a training set, a validation set and a test set. Step 4: Establish the clamped weighted PINN model, specifically including: forming a six-dimensional parameterized feature vector and performing parameter normalization on each dimension; constructing the basic framework of the PINN model, using the established six-dimensional parameterized feature vector, setting the hidden layer to a deep multilayer perceptron, using Tanh as the activation function, and making the final output temperature... T Mathematically, it conforms to the requirements of the thermophysical equations; the physical residual terms of the unsteady-state heat conduction equations are added to the model to restrict the function expressed by the PINN model so that it conforms to the physical laws; a weighted loss function with a clamping mechanism is set to ensure that the fitting accuracy of the high-temperature region is not affected by a large number of low-temperature samples; Step 5: Complete model training. Use the Adam adaptive moment estimator and StepLR step decay scheduler to adjust the parameters, ensuring the training process is stable and the model converges to the optimal solution.

2. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 1, characterized in that, In step one, the multi-disc brake system comprises a brake disc assembly consisting of one pressure disc, five moving discs, and four stationary discs. All discs are hollow cylindrical in structure, and the material parameters of the brake discs are set as follows: density ρ 1700 kg / m³, specific heat capacity at constant pressure C p 1700 J / (kg·K), basic thermal conductivity k =50 W / (m·K), then the thermal diffusivity is obtained. ; The unsteady heat conduction equation in cylindrical coordinates is shown below: (2), In the formula: T The transient temperature at any position on the brake disc; r These are radial coordinates, in meters (m). ρ This refers to the material density, expressed in kg / m³. The volumetric heat source intensity is expressed in W / m³. θ The circumferential angle is expressed in rad. x This is the axial coordinate, in meters (m). k r 、 k θ 、k x These are the radial, circumferential, and axial thermal conductivity coefficients, respectively, in W / (m·K).

3. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 2, characterized in that, In step one, the heat transfer boundary of each region adopts a mixed boundary condition. The mixed boundary is controlled by the valueFraction, i.e., the mixing weight coefficient, which controls the mixing ratio of the Dirichlet boundary condition (fixed temperature) and the Neumann boundary condition (fixed gradient). The heat transfer characteristics of each region are as follows: Strong convection region at the end: Mixing weighting coefficient is set to 0.8, equivalent convective heat transfer coefficient... ≈80 W / (m²·K); Axial transition region: The mixing weighting factor is set to 0.6, and the equivalent convective heat transfer coefficient is... ≈55 W / (m²·K); Central weak heat transfer zone: The mixing weighting coefficient is set to 0.4, and the equivalent convective heat transfer coefficient is... ≈35 W / (m²·K); Insulation boundary: The fine structure of the moving disk end face and spline groove is set as the insulation boundary; Total heat power generated during braking The kinetic energy consumption curve is shown in equation (5): (5), In the formula: η The heat-to-work conversion efficiency is set to 0.9, meaning that 90% of the kinetic energy is converted into heat energy in the brake disc. m For aircraft landing quality; v This refers to instantaneous velocity.

4. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 3, characterized in that, Step four involves forming a six-dimensional parameterized feature vector, and performing parameter normalization on each dimension. Specifically, this involves processing the input vector... Expanded to a six-dimensional form including operating condition parameters, it is as follows: (6), where: x, y, z The coordinates (m) describe the three-dimensional position within the brake disc assembly. t The time coefficient (s) represents the time progression of unsteady heat transfer. m Landing mass (kg); v 0 Let m be the initial landing velocity (m / s), where m and v 0 together constitutes the parameterized variable; The following normalization calibration method was used: Input scaling: The six-dimensional input is linearly scaled to a specified range using Min-Max mapping; Temperature calibration: This involves setting the maximum true temperature value from the training set. T max =1808.31K is used as the upper limit for global normalization, and the output temperature is mapped to the interval [-1, 1].

5. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 4, characterized in that, In step four, the basic framework of the PINN model adopts a seven-layer architecture: one input layer, six hidden layers, and one output layer. The hidden layers use the Tanh activation function.

6. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 5, characterized in that, Based on the isotropic assumption, the thermal conductivity is taken as follows during the simulation stage. k =50 W / (m·K) to construct the physical residual, and use the automatic differentiation technique of equation (7) to construct the unsteady-state heat conduction residual. : (7), In the formula: ∇² T For the Laplace operator, describe the thermal diffusion process within the disk; The temperature prediction value (K) output by the neural network; The dynamic volumetric heat source load is determined by the operating parameters; ρ The density of the material is (kg / m³). C p Specific heat capacity at constant pressure (J / (kg·K)); t Time (s); k The thermal conductivity is denoted as W / (m·K), and the isotropic constant is taken as 50 W / (m·K) during the simulation stage.

7. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 6, characterized in that, A weighted MSE loss function based on the actual temperature gradient was set, and the weight function was defined as Equation (8): (8), In the formula: clamp This is a clamping function; w min This is a bottom clamp to prevent the model from losing the entire background when fitting high temperatures; w max This is a top clamp, used to prevent excessive gradients at extreme points during aborted takeoff conditions, which could cause training oscillations. The weighted data loss expression is given by equation (9): (9), In the formula: L data This refers to the weighted data loss. N data This represents the total number of observed sample points; w i The temperature weight of the i-th sample point is calculated by equation (8); T̂ i Let be the predicted temperature (K) of the i-th sample point; T i For the first i The simulated true temperature (K) for each sample point.

8. The landing gear brake disc state prediction method based on clamp-weighted PINN according to claim 7, characterized in that, The Adam optimizer was chosen, and the training was divided into two phases: Phase 1 Global Search: Initial learning rate set to 1×102 -3 Set batch size to 8192; The second stage of fine-tuning: the StepLR scheduling strategy is enabled, the learning rate is decayed to 0.9 times the original value every 10,000 iterations, and the gradient norm of each parameter update is limited to a preset threshold of 1.0.