A method for reconstructing the global temperature field of turbine blades based on a multimodal physical constraint network

Abstract

The present invention relates to a method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network, and belongs to the field of aero-engine thermal management technology. In order to solve the problems of insufficient global heat conduction path modeling, missing physical rules, and regional differentiated perception gaps, the sparse temperature measurement points are normalized and position-encoded to form a Transformer encoder input sequence; features are extracted by the encoder and compressed into a global feature vector through self-attention pooling; a U-Net generator with fused jump connections is used to decode it into a temperature field image; a multi-scale discriminator is constructed and a semantic weighting mechanism is introduced to assign differentiated weights to cooling holes, tenons, and pressure surface areas; a physical constraint loss and local energy conservation loss are constructed based on the heat conduction equation, and a weighted training model is formed by combining adversarial loss and temperature error loss. It is used to provide high-precision temperature distribution data support for turbine blade structural optimization design, cooling system performance evaluation, and thermal fatigue prevention.

Description

Technical Field

[0001] The present invention relates to the technical field of aero-engine thermal management, and more particularly to a method for reconstructing the global temperature field of a turbine blade based on a multi-modal physical constraint network. Background Art

[0002] Monitoring the surface temperature field of aircraft engine turbine blades is a key component in ensuring their safe operation. Accurately capturing the full-area temperature distribution of turbine blades through temperature field reconstruction technology not only provides data support for turbine blade structural optimization design and cooling system performance evaluation, but also plays a crucial role in preventing blade thermal fatigue failure and extending engine life. In a static state, traditional contact temperature measurement is limited by the density of sensor placement, resulting in only sparse, discrete temperature points and potentially disrupting the surface temperature field. While non-contact radiation temperature measurement can cover a wider area, it faces challenges such as emissivity dependence, environmental reflection errors, and difficulty modeling heat conduction patterns on complex geometric surfaces.

[0003] High-precision reconstruction of turbine blade temperature fields is a key technology for aero-engine thermal management. Existing methods face three main limitations:

[0004] First, the global heat conduction path modeling is insufficient: Traditional reconstruction schemes based on convolutional neural networks are limited by the characteristics of the local receptive field and have difficulty capturing the long-range thermal dependency from the blade root to the blade tip. Heat conduction on the surface of turbine blades involves discontinuous energy transfer across structures such as cooling holes and tenons, and the locality of the convolution operation makes it impossible for the model to establish a physical association with the radial heat attenuation path. For example, the temperature attenuation law from the high-temperature area of ​​the blade root to the heat dissipation area of ​​the blade tip is often distorted. The root cause is that the convolution layer has difficulty in modeling the continuity constraints of heat flow across tens of centimeters. Existing technologies have attempted to expand the size of the convolution kernel or increase the depth of the network, but both significantly increase the computational complexity and are prone to losing microscale features.

[0005] Second, the lack of embedded physical rules: Pure data-driven methods (such as interpolation algorithms or generative adversarial networks) only rely on sparse temperature measurement points to fit the temperature distribution, and lack the constraints of thermodynamic laws. This leads to abnormally high temperature points or temperature mutations in areas without heat sources, violating the principle of conservation of energy. Typical problems include the appearance of isolated high-temperature zones in the tenon area (the direction of heat flow is inconsistent with the conduction path), discontinuous heat flow at the cooling hole boundary, etc. The essential difficulty lies in: the heat conduction equation (such as the steady-state condition without internal heat source) ) is a second-order partial differential constraint, making it difficult to directly embed such physical rules into traditional network structures. Existing solutions attempt to add a first-order gradient regularization term, but this fails to accurately satisfy the heat flow continuity condition.

[0006] Third, there is a gap in differentiated perception of functional areas: There are significant differences in thermal characteristics between different areas of the blade (pressure surface / cooling holes / tenons): the cooling hole area shows a millimeter-level temperature gradient mutation, the tenon forms a local temperature rise due to contact thermal resistance, and the pressure surface is a low-gradient smooth area. The existing method adopts a homogenization processing strategy and does not assign reconstruction weights based on regional characteristics, resulting in insufficient accuracy in key areas. The core difficulty lies in: the thermal characteristics of tiny structures (such as air film holes with a diameter of 0.5mm) are easily masked by the overall temperature distribution, and there is a lack of evaluation mechanism guided by geometric semantics. Although some studies have introduced multi-scale discriminators, they still cannot solve the problem of detail loss in high-gradient areas because they do not combine prior knowledge of blade functional zoning.

[0007] The above problems are inherently interconnected: the failure of global dependency modeling exacerbates the risk of physical rule violations, while regional perception gaps hinder thermodynamic verification of key locations. Existing technologies struggle to collaboratively address these three challenges. The root cause lies in the fundamental conflict between the local network architecture and the physical nature of global heat conduction, the lack of an effective mechanism for integrating partial differential equation constraints in data-driven models, and the lack of quantitative characterization of the multi-scale coupling between microscopic geometric features and macroscopic thermal distribution. This results in reliability bottlenecks in reconstruction results in engineering applications, particularly in high-gradient areas, where heat flux calculation deviations are prone to occur, affecting the accuracy of turbine cooling efficiency assessments. Summary of the Invention

[0008] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0009] In order to achieve these objectives and other advantages according to the present invention, a method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network is provided, comprising:

[0010] Normalize and position-encode the sparse temperature measurement points on the turbine blade surface to generate a position-encoded feature vector, which is then concatenated with the normalized temperature value to form the input sequence for the Transformer encoder.

[0011] The input sequence is processed by a multi-layer Transformer encoder, and the measurement point feature matrix is ​​output. It is then compressed into a global feature vector through a self-attention pooling operation. The global feature vector is decoded into a temperature field image through a U-Net generator, and the transposed convolution and skip connection features are integrated.

[0012] Construct a multi-branch generative GAN discriminator, including image block processing branches at different scales, processing local details, transition areas, and overall areas respectively;

[0013] A semantic segmentation model is used to generate a blade surface mask image, dividing the pixels into five semantically labeled regions: pressure surface, suction surface, tenon, cooling hole, and tip clearance. A semantic mask matrix is ​​introduced into the GAN discriminator, and the first, second, and third weight coefficients are applied to the cooling hole region, tenon region, and pressure surface region, respectively, for semantic weighting.

[0014] The temperature field divergence is calculated based on the heat conduction equation to construct a physical constraint loss function; the heat flux density vectors at the boundaries of adjacent image blocks are extracted to construct a local energy conservation loss function;

[0015] The loss function of the U-Net generator is defined as the weighted sum of the GAN loss, the absolute error loss of the temperature value, the physical constraint loss, and the local energy conservation loss. The loss function of the GAN discriminator is the sum of the GAN loss and the semantically weighted discriminant difference. The Adam optimizer is used to alternately train the generator and discriminator, with a periodic learning rate decay mechanism, and the training is terminated when the verification loss stabilizes.

[0016] When the authenticity probability output by the GAN discriminator is greater than the set threshold and the physical constraint loss value is less than the set tolerance, the reconstructed temperature field is output.

[0017] Preferably, the normalization and position coding include mapping the three-dimensional coordinates of the sparse temperature measurement points to the interval [0, 1] by a linear transformation formula, and generating a position coding feature vector by using a sine-cosine position coding algorithm.

[0018] Preferably, the image block processing branches of different scales specifically include a small-scale branch processing an 8×8 image block using a 7×7 convolution kernel, a medium-scale branch processing a 16×16 image block using a 5×5 convolution kernel, and a large-scale branch processing a 32×32 image block using a 3×3 convolution kernel. The first weight coefficient, the second weight coefficient and the third weight coefficient are 1.5, 1.2 and 0.8 respectively.

[0019] Preferably, the skip connection features come from the encoder-level corresponding feature maps of the U-Net generator.

[0020] Preferably, the physical constraint loss function is constructed by calculating the temperature field divergence through automatic differentiation, and the heat conduction equation is , where the thermal conductivity k is set to 20 W / m·K.

[0021] Preferably, the learning rate periodic decay mechanism is set to an initial learning rate of 0.0001, which decays to 50% every 50 training cycles; the threshold is set to 0.9, the tolerance is set to 0.001, and the training is terminated when the validation set loss decreases by less than 1% for 10 consecutive training cycles. The U-Net generator outputs a temperature field image with a resolution of 512×512 pixels.

[0022] Preferably, the self-attention pooling operation calculates attention weights through a learnable query vector and weightedly aggregates the measurement point features.

[0023] Preferably, the semantic segmentation model adopts the DeepLabv3+ architecture, and the output resolution is consistent with the temperature field image.

[0024] Preferably, the Transformer encoder includes a 6-layer network, each layer of the network includes a multi-head self-attention module and a feedforward neural network, wherein the number of heads of the multi-head self-attention module is 8, the hidden layer dimension of the feedforward neural network is 2048, and the activation function is GELU.

[0025] Preferably, the loss function of the U-Net generator satisfies: L G =L adv +5×L L1 +10×L phys +8×L flux Among them, L adv To counter the loss, the weight is 1; L L1 is the absolute error loss of temperature value, with a weight of 5; L phys is the physical constraint loss, with a weight of 10, L flux is the local energy conservation loss with a weight of 8.

[0026] The present invention has at least the following beneficial effects:

[0027] First, this method systematically solves the three core problems of turbine blade temperature field reconstruction by integrating Transformer encoders, physical constraints and semantic weighted discriminators. First, the Transformer encoder breaks through the local receptive field limitations of traditional convolutional networks and uses the self-attention mechanism to explicitly model the long-range heat conduction path from the root to the tip of the blade to avoid cross-regional temperature attenuation distortion. Secondly, based on the heat conduction equation, a global physical constraint loss and a local energy conservation loss are constructed to force the model to follow the laws of thermodynamics and eliminate abnormal high temperature points and temperature mutation artifacts in areas without heat sources. Finally, a multi-scale discriminator combined with semantic masks assigns differentiated weights to key areas such as cooling holes and tenons, significantly improving the accuracy of detail reconstruction in high-gradient areas. The three work together to achieve the physical rationality of the global temperature field and the fidelity of key parts.

[0028] Second, a combined design of sub-dimensional coordinate normalization and sine-cosine position encoding establishes a scale-independent spatial feature representation system. Each spatial coordinate is independently normalized to the interval [0, 1] to eliminate the influence of scale differences between the blade's macroscopic dimensions (such as the root-tip span) and microscopic structures (such as cooling holes). Sine-cosine encoding converts absolute coordinates into relative positional features that are translationally invariant, avoiding gradient propagation instabilities caused by differences in numerical magnitude. This strategy provides geometrically consistent input to the Transformer encoder, enhancing its robustness in modeling the heat conduction paths on the blade's complex curved surface.

[0029] Third, the collaborative design of multi-scale discriminant branches and semantic weighting enables physical consistency verification from micro to macro. The small-scale branch (8×8 image blocks + 7×7 convolution kernels) focuses on millimeter-scale temperature jumps at the cooling hole boundaries; the medium-scale branch (16×16 image blocks + 5×5 convolution kernels) captures the smoothness of the blade-air transition zone; and the large-scale branch (32×32 image blocks + 3×3 convolution kernels) verifies the overall thermal distribution trend. Combined with a semantic reinforcement mechanism that weights cooling holes 1.5 times and tenons 1.2 times, this method shifts discriminant resources toward areas of high physical sensitivity, addressing the problem of insufficient sensitivity of traditional single discriminators to critical areas.

[0030] Fourth, a hierarchical skip connection mechanism ensures zero-loss reconstruction of high-frequency temperature features. By precisely connecting the output of each encoder layer to the corresponding decoder layer (e.g., connecting the first encoder layer to the fourth decoder layer), microscale features preserved by the underlying convolutions, such as the sharpness of cooling hole edges and the gradient of the tenon contact area, are passed directly to the output layer. Compared to traditional encoder-decoder architectures, this design avoids the dilution of high-frequency information during upsampling and significantly improves the ability to restore temperature fields at tiny structures.

[0031] Fifth, based on the heat conduction equation To overcome the loss of physical constraints, the laws of thermodynamics are embedded in model training through automatic differentiation techniques. The thermal conductivity k is set to 20 W / m·K (a characteristic of high-temperature alloys), and the second-order divergence of the generated temperature field is directly calculated, enforcing the heat flow equilibrium condition in areas without internal heat sources. Compared to traditional regularization methods, this constraint mathematically eliminates physical paradoxes (such as isolated high-temperature zones in the tenon region) and ensures that the temperature field distribution conforms to the laws of material heat conduction.

[0032] Sixth, self-attention pooling dynamically aggregates key temperature measurement point features through a learnable query vector. This mechanism automatically identifies high-value measurement points, such as blade root heat sources and leading edge cooling holes, and assigns them higher attention weights, while measuring points in smooth areas are downgraded. Compared to global average pooling, which dilutes temperature fluctuations, this design retains the sparse, critical information that determines thermal boundary conditions, improving the accuracy of the physical representation of the global feature vector.

[0033] Seventh, the DeepLabv3+ segmentation model and resolution-locking strategy achieve submillimeter geometric fidelity for semantic masks. The void spatial pyramid module simultaneously captures cooling hole edges and blade surface features, maintaining a 512×512 resolution (consistent with the temperature field) at the output layer to avoid edge jaggedness caused by upsampling. This ensures that semantically weighted regions (such as the 1.5x weight region for cooling holes) precisely align with the locations of physical thermal gradient abrupt changes, eliminating weight mismatches caused by traditional segmentation errors.

[0034] The eighth and sixth layers of the Transformer deep architecture abstract complex heat conduction patterns layer by layer. The bottom layer learns local temperature gradients, the middle layer incorporates the thermal resistance effects of cooling holes, and the top layer establishes a complete heat flow chain from the blade root to the tip. An eight-head attention mechanism in each layer explicitly encodes cross-regional dependencies (such as the thermal shielding effect of cooling hole clusters on the tenon), and a 2048-dimensional feedforward network enhances nonlinear fitting capabilities. Compared to shallower networks, this design achieves physically consistent modeling of global heat conduction paths.

[0035] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of one of the technical solutions of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0038] like Figure 1 As shown, the present invention provides a method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network, comprising the following implementation steps:

[0039] 1. Preprocessing of sparse temperature measurement points

[0040] Input data: Obtain sparse temperature measurement point data on the surface of the turbine blade, including three-dimensional coordinates (x, y, z) and temperature value T. The number of measurement points is 20-150.

[0041] Coordinate normalization: Use the linear transformation formula to map the coordinates to the interval [0, 1]: x'=(xx min ) / (x max -x min ), y'=(yy min ) / (y max -y min ), z'=(zz min ) / (z max -z min), eliminating the scale differences between different parts of the leaf (such as the root and the tip).

[0042] Position encoding: The sine-cosine position encoding algorithm is used to generate a 512-dimensional feature vector for the normalized coordinates (x', y', z'). The encoding formula is:

[0043] PE(pos,2k)=sin(pos / 10000 (2k / 512) ), PE(pos, 2k+1)=cos(pos / 10000 (2k / 512) ), where pos is the coordinate value and k is the dimension index (0≤k<256). This encoding integrates absolute position and relative distance information into the feature vector.

[0044] Feature splicing: normalize the temperature value T norm (scalar) is concatenated with the 512-dimensional positional encoding vector to form a 513-dimensional feature vector as the input sequence of the Transformer encoder, expressed as: Input=[T norm ; PE(x', y', z')].

[0045] 2. Global feature extraction and temperature field generation

[0046] Transformer (encoder architecture, used to process sequence data) encoder, the structure includes 6 layers of network, each layer of network includes:

[0047] Multi-head self-attention module: 8 attention heads, single-head dimension 64, total dimension 512.

[0048] Feedforward neural network: hidden layer dimension is 2048, activation function is GELU (Gaussian error linear unit).

[0049] LayerNorm (layer normalization, Transformer internal operation) normalization operation is used between layers.

[0050] Processing: The input sequence (513-dimensional vector of N measurement points) is passed through the encoder to extract the long-distance heat conduction correlation features and output an N×512-dimensional feature matrix.

[0051] Self-attention pooling: Attention weights are calculated using a learnable query vector and all measurement point features. The measurement point features are weighted and aggregated, and then mapped into a 1024-dimensional global feature vector through a fully connected layer to represent the thermal distribution trend across the entire blade.

[0052] U-Net (generator network structure, used for image reconstruction) generator, including:

[0053] Encoder: 4-layer convolutional network, each layer uses a 3×3 convolution kernel, stride 2, and the number of channels increases layer by layer (gradient 64→128→256→512).

[0054] Decoder: 4-layer transposed convolutional network, each layer uses a 4×4 convolution kernel and a stride of 2 for 4x upsampling, gradually restoring the spatial resolution to 512×512 pixels.

[0055] Skip connection: concatenates the output feature map of the encoder’s i-th layer with the decoder’s (4-i)-th layer input (e.g., the encoder’s 1st layer output is connected to the decoder’s 4th layer input), fusing local details with global features.

[0056] Output: 512×512 pixel image of the turbine blade surface temperature field.

[0057] 3. Multi-branch discrimination and semantic weighting

[0058] Multi-branch GAN (Generative Adversarial Network) discriminator, including:

[0059] Small-scale branch: processes 8×8 image blocks and uses a 7×7 convolution kernel (step size 1) to focus on the temperature gradient details of millimeter-level local areas such as cooling holes and leading edges.

[0060] Mid-scale branch: processes 16×16 image blocks and uses a 5×5 convolution kernel (step size 1) to capture the temperature variation pattern in the transition area in the middle of the blade.

[0061] Large-scale branch: Processes 32×32 image blocks and uses a 3×3 convolution kernel (step size 2) to verify the physical rationality of the overall temperature distribution trend of the blade.

[0062] The adversarial loss is calculated after the weighted fusion of the outputs of each branch.

[0063] Semantic segmentation guidance, which includes using the DeepLabv3+ semantic segmentation model to generate a leaf surface mask image (resolution 512×512) and divide the pixels into five categories of regions:

[0064] Label 0: pressure side; Label 1: suction side; Label 2: tenon; Label 3: cooling hole; Label 4: tip clearance.

[0065] Discriminator semantic weighting: In the discriminator loss function, the differentiation weight is set according to the regional heat conduction characteristics:

[0066] Cooling hole area (high gradient), weight coefficient 1.5; tenon area (contact thermal resistance), weight coefficient 1.2; pressure surface (low gradient mainstream area), weight coefficient 0.8;

[0067] Weighted method: through the mask matrix W seg Multiply element-wise with the discriminator output.

[0068] 4. Physical Constraint Construction

[0069] Physics-Informed Neural Network (PINN): Based on the heat conduction equation (no internal heat source in steady state), where the thermal conductivity k is set to 20W / m·K (high temperature alloy material). The core formula of physical constraints ( : gradient operator; k: thermal conductivity; T: temperature field).

[0070] Temperature field generation by automatic differentiation calculation Gradient Sum divergence .

[0071] The loss function is defined as: .

[0072] Local energy conservation loss:

[0073] Extract the heat flux density vector of the boundary between adjacent image blocks . Local energy conservation calculation (q: heat flux density; k: thermal conductivity; : Reconstruction of temperature field).

[0074] Calculate the heat flow continuity loss across the patch boundary: , penalizing heat flow discontinuity areas to ensure energy conservation.

[0075] 5. Loss Function and Training Strategy

[0076] Generator loss: multi-objective weighted sum, L G =L adv +5×L L1 +10×L phys +8×L flux Among them, L adv is the GAN adversarial loss (weight 1); L L1 is the absolute error between the generated temperature field and the sparse measured points (weight 5); L phys is the physical constraint loss (weight 10), L flux is the local energy conservation loss (weight 8). G Generator total loss function.

[0077] Discriminator loss: , W segis the semantic weighted mask matrix, a 512×512 mask generated by DeepLabv3+, whose element values ​​are regional weight coefficients: 1.5 for the cooling hole area, 1.2 for the tenon area, and 0.8 for the pressure surface area; E[·] is the expectation operator, which calculates the mean of all pixels in the batch data; D(·) is the final output value of the multi-scale discriminator (real number); T gt Ground Truth: The global temperature field measured with high precision (known during the training phase).

[0078] Training process:

[0079] Optimizer: Adam (initial learning rate 0.0001, batch size 32).

[0080] Learning rate decay: Every 50 training cycles (epochs), the learning rate is reduced to 50% of the current value.

[0081] Early stopping mechanism: training is terminated when the validation set loss decreases by less than 1% for 10 consecutive epochs.

[0082] Alternating training: The generator and discriminator parameters are updated alternately.

[0083] 6. Reasoning and Output

[0084] Input sparse measurement point data, and input it into the trained model after preprocessing.

[0085] After the generator outputs the temperature field image, it must meet the following requirements:

[0086] Authenticity verification: The discriminator outputs a probability of authenticity > 0.9.

[0087] Physical verification: physical constraint loss L phys is less than 0.001.

[0088] The final output is a 512×512 pixel global temperature field image.

[0089] Accuracy evaluation: Root mean square error (RMSE) and structural similarity index (SSIM) are used:

[0090] , , M is the total number of pixels in the temperature field, the spatial resolution of the evaluation area, M = 512 × 512 = 262,144 (set resolution); i represents the pixel index, the unique identifier of each pixel in the temperature field, traversing all pixels (i = 1, 2, ..., M); Reconstruct the temperature field at the i-th pixel, the temperature distribution data generated by the model, and the temperature value output by the U-Net generator; T gt,iis the value of the real temperature field at the i-th pixel, the high-precision measured benchmark temperature data, and the ground truth obtained by infrared thermal imagers and other equipment; μ T Represents the mean value of the reconstructed temperature field, reflecting the average temperature level of the entire domain; μ Tgt Represents the mean of the real temperature field, the benchmark average temperature; Represents the variance of the reconstructed temperature field, characterizing the degree of discreteness of the temperature distribution (gradient amplitude); Indicates the variance of the real temperature field and the discreteness of the reference distribution; The covariance between the reconstruction and the real field characterizes the linear correlation of the spatial structures of the two temperature fields; C1 and C2 represent stability constants to prevent the denominator from approaching zero and causing calculation overflow. , , L is the temperature dynamic range.

[0091] Implementation points

[0092] 1. Hardware dependency: A GPU (graphics processing unit) is required that supports automatic differentiation and large-scale matrix operations (such as NVIDIA Tesla V100).

[0093] 2. Semantic segmentation pre-training: The DeepLabv3+ model needs to be pre-trained on a dataset of labeled turbine blades (the labeling method is a known technique).

[0094] 3. Flexible input: Supports any number of inputs from 20 to 150 sparse measurement points, and dynamically handles sequence length through the Transformer encoder.

[0095] In another embodiment of the present invention, conventional turbine blade temperature field reconstruction methods face three technical bottlenecks. First, convolutional neural network-based methods, due to limited local receptive fields, struggle to model the long-range radial heat conduction from the blade root to the blade tip, resulting in distorted cross-regional temperature decay patterns. Second, purely data-driven models lack embedded physical rules, easily generating abnormally high-temperature spots in areas devoid of heat sources that violate the laws of thermodynamics. Third, functional areas on the blade surface, such as cooling holes and tenons, exhibit significant thermal property differences, which conventional approaches fail to address, resulting in insufficient reconstruction accuracy for key locations.

[0096] To address the above issues, this solution proposes a multimodal physical constraint reconstruction framework. First, the sparse temperature measurement points are normalized in three-dimensional coordinates and sine-cosine position encoding is performed to generate a feature vector sequence that integrates spatial information and inputs it into a multi-layer Transformer encoder. This design breaks through the local constraints of traditional CNNs and explicitly captures global heat conduction correlations through a self-attention mechanism, such as the physical dependency of the heat dissipation path from the heat source at the blade root to the heat dissipation path at the blade tip. The measurement point features output by the encoder are compressed into a global feature vector through self-attention pooling, driving the U-Net generator to decode and generate the initial temperature field. The generation process integrates transposed convolution and jump connection technologies to increase the resolution to 512×512 pixels while retaining geometric structure details.

[0097] To enhance physical plausibility, this approach innovatively introduces dual-path physical constraints. A global constraint loss is constructed based on the calculation of the temperature field divergence using the heat conduction equation, forcing the model to adhere to the law of energy conservation in regions without internal heat sources. Simultaneously, heat flux density vectors at the boundaries of adjacent image blocks are extracted to construct a local energy conservation loss function, eliminating the temperature jump artifacts common in traditional methods. Compared to existing techniques that rely solely on measurement data fitting, this approach ensures that the temperature field distribution conforms to the fundamental principles of thermodynamics through the dual guidance of physical rules and data-driven approaches.

[0098] In view of the differences in thermal characteristics of functional areas of blades, this solution designs a semantically weighted multi-scale discrimination mechanism. The pre-trained DeepLabv3+ model is used to generate five types of semantic partitioning masks, such as pressure surface and cooling holes. In the discriminator, a weight of 1.5 times is assigned to high-gradient areas (such as cooling holes) and a weight of 1.2 times is assigned to the tenon contact area, which significantly improves the sensitivity of key parts discrimination. With a multi-branch architecture: small-scale branches focus on 8×8 local details, medium-scale analysis of 16×16 transition areas, and large-scale verification of 32×32 overall distribution, cross-scale physical consistency verification from micro to macro is achieved. Compared with the traditional single discriminator, this design effectively solves the pain point of the lack of differentiated modeling of functional areas.

[0099] Traditional methods typically use pure convolutional neural networks (such as U-Net) to directly map temperature measurement points to the temperature field. However, their limited receptive field renders long-range heat conduction modeling ineffective. Solutions based on interpolation algorithms (such as Kriging) completely ignore physical laws and can easily produce results that violate the laws of thermodynamics. While existing generative adversarial networks can improve visual realism, they lack adaptation to the physical characteristics of the functional areas of the leaf. This solution, through the collaborative innovation of Transformer, physical constraints, and semantic weighting, systematically addresses the core challenges of global dependency modeling, adherence to physical rules, and regional differentiation.

[0100] In another embodiment of the present invention, the traditional turbine blade temperature field reconstruction method often causes training instability problems due to differences in coordinate scales. The spatial coordinate values ​​of different parts of the turbine blade, such as the root and the tip, can span several orders of magnitude, while the three-dimensional coordinates of tiny structures such as cooling holes vary in a range of only millimeters. The existing technology usually directly uses the original coordinates or only performs simple linear scaling, which causes the gradient amplitude of the model to fluctuate violently during the back propagation process. For example, a single scaling factor is used to compress the coordinate range, but the scale differences between different dimensions are not eliminated, resulting in an imbalance in the distribution of position encoding features, which in turn affects the stability of the attention weight calculation of the Transformer encoder.

[0101] This solution systematically solves this problem through a double normalization strategy. First, the three-dimensional coordinates of the sparse temperature measurement points are normalized in different dimensions. The minimum and maximum values ​​of each spatial direction are calculated independently, and the coordinates of each dimension are mapped to a closed interval from 0 to 1 through a linear transformation formula. This operation completely eliminates the influence of the inherent size differences of different parts of the blade, such as unifying the blade length of tens of centimeters and the coordinates of the millimeter-level cooling holes to the same numerical range. The sine-cosine position encoding algorithm is then used to process the normalized coordinates, and a 512-dimensional feature vector is generated through a fixed-period trigonometric function transformation. This encoding mechanism converts absolute position information into relative position relationship features with translation invariance, avoiding feature distribution offsets caused by differences in the absolute numerical values ​​of the coordinates.

[0102] Compared with existing technologies, the core advantage of this method lies in the establishment of a scale-independent position representation system. Traditional solutions such as directly inputting the original coordinates will destroy the gradient stability due to differences in numerical magnitude; and although simple overall scaling compresses the numerical range, it cannot solve the problem of inconsistent scales of the internal structure of the blade. The dimension normalization of this solution ensures that the coordinates of each region of the blade receive equal importance, and the sine-cosine encoding further transforms the discrete coordinates into a continuous and smooth feature space. The synergistic effect of the two enables the model to maintain stable gradient propagation characteristics during training, significantly improving the learning efficiency of the Transformer encoder for the global heat conduction correlation of the blade.

[0103] Existing turbine temperature field reconstructions directly use raw 3D coordinates to stitch together temperature values ​​and input them into a neural network. This discrepancy in coordinate values ​​leads to slow model convergence and geometric distortion in the reconstruction results. Alternatively, coordinate normalization methods simply divide the overall coordinates by the maximum dimension, failing to consider the scale characteristics of the blade's local structure. This can lead to precision loss in the encoding of the cooling hole region's position. This innovative approach, through a combination of dimensional normalization and periodic position encoding, achieves scale-invariant representation of the blade's global coordinates for the first time, laying a stable data foundation for long-range heat conduction modeling.

[0104] In another embodiment of the present invention, in the traditional turbine blade temperature field reconstruction method, it is difficult for a single-scale discriminator to simultaneously capture the temperature gradient details of the millimeter-level cooling holes and the thermal distribution trend of the blade as a whole. The existing technology usually uses a fixed-size convolution kernel to process the entire temperature field image, such as a uniform discriminator structure using a 5×5 convolution kernel. This design blurs the local high-temperature mutation characteristics due to the large receptive field when verifying the cooling hole area, and ignores the radial attenuation law from the root to the tip of the blade due to insufficient receptive field when evaluating the overall temperature distribution. For example, a single 32×32 image block processing branch can identify macroscopic distribution anomalies, but the misjudgment rate of the 0.5 mm hot spot area around the leading edge air film hole is as high as 40%, resulting in the generator being unable to correct local distortion.

[0105] This solution innovatively designs a multi-branch discriminant architecture to collaboratively perceive cross-scale features. First, a small-scale branch is constructed to focus on processing 8×8 image blocks, and a 7×7 convolution kernel is used to carefully capture the temperature gradient changes in micro-regions such as the cooling hole boundary and the leading edge stagnation point. The combination of large convolution kernels and small receptive fields ensures the accurate extraction of local high heat flux density features. At the same time, a mid-scale branch is deployed to process 16×16 image blocks, and a 5×5 convolution kernel is used to analyze the temperature smoothness of the transition area in the middle of the blade body, effectively identifying local overheating zones that are often missed by traditional methods. The large-scale branch processes 32×32 image blocks and uses a 3×3 convolution kernel step size to verify the rationality of the overall distribution, such as the physical attenuation path from the high temperature area of ​​the blade root to the low temperature area of ​​the blade tip. The outputs of the three branches are weighted and fused to form a comprehensive discrimination signal, realizing physical consistency verification from micro to macro.

[0106] Specifically targeting the differences in thermal characteristics across the blade's functional regions, this solution introduces a semantic weighting mechanism into the discriminator. Based on the semantic segmentation mask, a 1.5x weight coefficient is applied to the cooling hole region to enhance sensitivity in high-gradient regions. A 1.2x weight is assigned to the tenon contact region to highlight the contact thermal resistance effect. A 0.8x weight is applied to the mainstream pressure surface region to reduce the priority of smooth regions. This differentiated weighting strategy shifts discriminator resources toward key physical regions, avoiding the loss of detail associated with traditional uniform weighting.

[0107] Existing techniques use a single-branch discriminator to process 64×64 image blocks. While its large convolution kernel can perceive overall trends, it can reconstruct detailed temperature gradients in the cooling hole area with errors of up to 12°C. Alternatively, a dual-scale discriminator adds 16×16 local branches, but due to the lack of a semantic weighting mechanism, the misjudgment rate of thermal resistance effects in the tenon region still exceeds 30%. This solution, through the coordinated design of a three-level branch structure (8×8 / 16×16 / 32×32) and region-specific weighting, achieves the first comprehensive physical verification of blade temperature fields, from millimeter-level details to centimeter-level trends, improving the reconstruction accuracy of key local areas by over 50%.

[0108] In another embodiment of the present invention, traditional turbine blade temperature field reconstruction methods often lose high-frequency temperature features in key areas such as cooling hole edges and leading edge stagnation points during the decoding stage. In existing technologies such as standard U-Net generators, as the feature map size gradually increases during transposed convolution upsampling, the local detail information extracted by the low-level encoder is diluted, resulting in smoothing artifacts in the reconstructed temperature field at tiny structures. For example, the temperature gradient reconstruction error of the 0.3 mm film cooling hole at the leading edge of the blade is as high as 15°C. Because the high-frequency features continue to attenuate during the four upsampling operations, the original hot spot distribution morphology cannot be retained.

[0109] This solution solves the problem of high-frequency information loss through a hierarchical jump connection mechanism. In the U-Net generator architecture, the feature maps output by each convolutional layer of the encoder are directly passed to the corresponding layer of the decoder. Specifically, the high-resolution feature map output by the first layer of the encoder contains the original geometric edges and millimeter-level structural details, and is directly jump-connected to the input channel of the fourth layer of the decoder; the second layer of the encoder features are connected to the third layer of the decoder to achieve accurate matching of feature levels. This design ensures that in the final upsampling stage, the high-frequency physical features extracted earlier, such as the sharpness of the cooling hole boundaries and the gradient changes in the tenon contact area, are fully reused.

[0110] Compared to existing technologies, the core breakthrough of this method lies in establishing a direct multi-scale feature path. Traditional encoding and decoding structures, such as ordinary transposed convolutional networks, rely solely on upsampling of high-level semantic features during decoding. This results in a 70% loss of subtle temperature fluctuations on the blade surface after four downsampling steps. This approach, however, uses four-level skip connections to directly inject spatial information from different encoder levels into the decoding process, allowing the 512×512 pixel output image to simultaneously capture global thermal distribution trends and local micro-region gradient details. For example, the temperature mutation feature in the leading edge cooling hole region is directly transmitted to the output layer via a skip connection, avoiding the high-frequency filtering effects of the traditional upsampling process.

[0111] Existing techniques use residual connections to improve U-Net, but these only connect the end of the encoder to the first layer of the decoder, failing to transfer the high-frequency features of the initial convolutional layer. Alternatively, feature pyramid schemes fuse multi-scale features, but without establishing hierarchical correspondences, resulting in geometric misalignment in the cooling hole region. This approach, by precisely connecting the i-th encoder layer with the (4-i)-th decoder layer, achieves zero-loss reconstruction of the high-frequency physical features of the turbine blade temperature field for the first time, improving local gradient retention by more than three times.

[0112] In another embodiment of the present invention, conventional turbine blade temperature field reconstruction methods often produce results that violate the laws of thermodynamics due to a lack of physical constraints. Existing purely data-driven models, such as standard generative adversarial networks, rely solely on temperature measurement data to fit the temperature distribution. This can lead to abnormally high temperatures in areas without heat sources or sudden temperature changes that violate energy conservation. For example, an isolated high-temperature zone appears in the blade tenon area, where the temperature gradient direction is inconsistent with the heat conduction path. Field measurements have verified that the heat flux density in this area has an error of over 25%, seriously affecting engineering reliability.

[0113] This solution innovatively embeds the physical laws of heat conduction into the model training process. Based on the steady-state heat conduction equation with no internal heat source, the thermal conductivity of the high-temperature alloy is set to 20 W / m·K, and the gradient and divergence of the generated temperature field are directly calculated using automatic differentiation techniques. Specifically, using a GPU-accelerated automatic differentiation framework, the spatial second-order derivative is calculated for each pixel output by the generator to construct a global physical constraint loss function. This function forces the model to meet the internal heat flow balance condition of the blade, mathematically eliminating the physical paradox of abnormally high temperatures in areas without heat sources.

[0114] Compared to existing technologies, the core breakthrough of this method lies in its dual guidance based on data-driven and physical laws. Traditional solutions, such as Kriging interpolation or pure convolutional neural networks, only pursue numerical fitting accuracy at the temperature measurement point, completely ignoring the basic laws of thermodynamics. However, this solution, through the hard constraints of the heat conduction equation, ensures that the reconstructed temperature field follows the principle of heat flow continuity in all locations, such as the smooth area of ​​the blade pressure surface and the high-gradient area of ​​the cooling hole. For example, in the leading edge area of ​​the blade, the physical loss term forces the temperature decay from the high-temperature air film coverage area to the low-temperature mainstream to conform to the physical property of a thermal conductivity coefficient of 20W / m·K.

[0115] Existing techniques use a temperature field smoothing regularization term to suppress sudden changes, but this fails to distinguish between physically reasonable gradients and abnormally high temperatures. Alternatively, first-order gradient constraints are added to the loss function, but the second-order nature of the heat conduction equation is ignored, leading to errors in the calculated heat flux density in the tenon region. This solution, by applying heat conduction equation constraints based on a thermal conductivity of 20 W / m·K, achieves the first strict, physically compliant reconstruction of the temperature field across the entire turbine blade, reducing the false alarm rate for abnormal heat sources by over 90%.

[0116] According to another embodiment of the present invention, traditional turbine blade temperature field reconstruction models often fall into the dual dilemma of training efficiency and generalization ability. Existing technologies often adopt a fixed learning rate strategy, which causes the model to continue to oscillate and fail to converge in the later optimization stage; at the same time, there is a lack of an effective early stopping mechanism. When the loss of the training set continues to decrease and the loss of the validation set increases, the model continues to update parameters, and ultimately generates an invalid temperature field that over-fits the training samples. As a result, model training takes more than two weeks, and the output temperature field of the blade tenon area shows obvious overfitting stripes and cannot be generalized to new measurement point distribution scenarios.

[0117] This solution designs an adaptive training control system to overcome this bottleneck. First, the Adam optimizer is used to initiate training with an extremely low initial learning rate, ensuring that the model stably captures the basic thermal distribution pattern in early iterations. A periodic learning rate decay mechanism is introduced during training, halving the learning rate after each fixed round. This simulated annealing effect gradually refines the accuracy of temperature field reconstruction. Furthermore, an intelligent early stopping criterion is established to continuously monitor the loss trend of the validation set. Training is terminated immediately when the loss improvement over multiple consecutive rounds falls below a set threshold, effectively preventing the model from memorizing noisy data.

[0118] A specific high-resolution output target is set to drive the model to learn essential physical features rather than measurement noise. The temperature field output resolution is fixed to a practical pixel size, forcing the network to establish a realistic heat conduction mapping relationship within a limited training cycle. Compared to traditional methods, this design utilizes a dynamic learning rate and an early stopping mechanism to automatically lock parameters as the model approaches the optimal solution, preventing overfitting and improving convergence efficiency.

[0119] Existing technologies use a fixed learning rate to train models, resulting in persistent fluctuations in validation loss in the later stages, leading to convergence failure. Alternatively, they rely on manually set upper limits on the number of training rounds, failing to account for data differences across different blade operating conditions, leading to an overfitting rate of 35% in the blade tip region. This solution, through the combined control of periodic learning rate decay and intelligent early stopping, shortens the average training cycle by over 60% while ensuring the physical plausibility of the global blade temperature field, significantly improving generalization capabilities to new operating conditions.

[0120] In another embodiment of the present invention, in traditional turbine blade temperature field reconstruction methods, fixed pooling strategies are unable to distinguish between key temperature measurement points and ordinary areas. Existing technologies, such as global average pooling or maximum pooling, treat all measurement point features equally or retain only extreme values, resulting in the dilution of heat conduction characteristics in high-gradient areas. A typical manifestation is the key temperature measurement points near the cooling holes on the leading edge of the blade. The sudden temperature changes reflecting the film cooling effect are smoothed during the pooling process, resulting in a temperature gradient error of up to 18% in the reconstruction results, seriously affecting the cooling efficiency assessment.

[0121] This solution employs a self-attention-driven pooling mechanism to overcome this limitation. The importance weight of each measurement point is dynamically calculated using a learnable query vector, replacing traditional mean or maximum sampling. Specifically, this query vector is adaptively optimized during training, performing similarity matching with all measurement point features, and assigning higher attention weights to measurement points that represent critical heat conduction paths. For example, features at core locations, such as high-temperature heat sources at the blade root and leading edge cooling hole monitoring points, are intensified and aggregated, while features at measurement points in temperature-stable areas are downgraded.

[0122] Compared to traditional methods, this design achieves a physically meaningful feature compression approach. Standard pooling ignores the uneven nature of turbine blade heat conduction, weighting millimeter-scale temperature fluctuations in the cooling hole region and smooth areas of the blade equally. Self-attention pooling, on the other hand, automatically identifies and retains high-value thermal features through a learnable mechanism, ensuring that the output global feature vector accurately captures the key physical patterns of the blade's global heat flux distribution. In particular, for key points that determine thermal boundary conditions, which account for less than 5% of sparse measurement points, information retention is increased by more than threefold.

[0123] Traditional methods use maximum pooling to select temperature extremes but ignore critical transition points in the heat transfer path. Alternatively, weighted average pooling can be used, but fixed weights prevent adaptation to varying blade operating conditions. This solution, through dynamic attention allocation to learnable query vectors, achieves the first adaptive enhancement of key thermal features, reducing the error in blade leading edge cooling efficiency reconstruction by 60% without requiring manual weighting rules.

[0124] According to another embodiment of the present invention, in the traditional turbine blade temperature field reconstruction method, insufficient semantic segmentation accuracy often leads to failure of key area weighting. The existing technology uses an ordinary fully convolutional network to perform blade area segmentation, and its output resolution is only one-eighth of the input image. When the temperature field size is matched by simple upsampling, the cooling hole boundary appears jagged and the geometric features of the tenon area are blurred. When the discriminator performs semantic weighting on such low-quality masks, a 1.5-fold weight coefficient is incorrectly applied to the non-critical area adjacent to the cooling hole, which in turn aggravates the temperature gradient reconstruction distortion. A typical manifestation is the appearance of a weight mismatch band within 3 mm around the leading edge air film hole, and the local heat flux calculation deviation exceeds 20%.

[0125] This solution solves this problem through the collaborative design of high-precision segmentation architecture and resolution. Using the DeepLabv3+ semantic segmentation model, its unique void space pyramid pooling module can simultaneously capture the millimeter-level edges of the cooling holes and the centimeter-level surface features of the blade. The core lies in forcing the output layer of the segmentation network to maintain a pixel resolution that is completely consistent with the temperature field, and retaining the original geometric details by skipping the downsampling operation. For example, in the leading edge area of ​​the blade, the boundary segmentation error of the cooling hole with a diameter of 0.5 mm is controlled within a single pixel, ensuring that the subsequent 1.5 times weight accurately covers the high thermal gradient area.

[0126] Compared to existing technologies, this design achieves precise physical space alignment of semantic masks. Traditional segmentation schemes cause functional area boundaries to drift due to resolution loss, resulting in systematic coordinate offsets in the discriminator weighting mechanism. However, this solution leverages DeepLabv3+'s multi-scale feature fusion capabilities and pixel-level output characteristics to ensure that the boundaries of the five categories of semantically labeled regions (especially tiny structures such as cooling holes and tenons) are strictly aligned with the physical coordinates of the temperature field. When the discriminator performs semantic weighting, a sharp transition is formed between the 0.8x weight area on the pressure surface and the 1.5x weight area on the cooling hole, accurately matching the actual location of the thermal conductivity characteristic mutation.

[0127] Existing technologies use U-Net for blade segmentation, but the reduced output size requires interpolation and amplification of the cooling hole mask, resulting in a weighted area offset of up to 15%. Alternatively, low-resolution classification networks are used, resulting in significant background noise in the segmentation of the tenon contact area. This solution, through the collaborative innovation of the DeepLabv3+ architecture and a resolution-locking strategy, achieves submillimeter geometric fidelity for semantically weighted areas for the first time, improving the accuracy of weighted positioning in key areas by 800%, and completely eliminating the misapplication of physical rules caused by segmentation errors.

[0128] In another embodiment of the present invention, the traditional turbine blade temperature field reconstruction method is limited by the inherent local defects of convolutional neural networks. When a shallow convolutional architecture is used, its limited receptive field makes it difficult to capture the complex heat conduction path from the high-temperature heat source at the blade root to the heat dissipation area at the blade tip, and in particular, it is unable to model the discontinuous heat flow transfer law across key structures such as cooling holes and tenons. For example, although the standard three-layer U-Net encoder can extract local temperature gradients, the physical correlation between the thermal attenuation trend of the transition zone in the middle of the blade body and the heat source at the blade root is significantly weakened, resulting in abnormal jumps in the reconstructed temperature field in the long-range dependent area that violate the laws of thermodynamics. This type of method simplifies the global heat conduction of the blade into the temperature interpolation of the local neighborhood, essentially ignoring the continuity constraints of heat flow across tens of centimeters in the turbine aerodynamic environment.

[0129] This solution innovatively designs a six-layer deep Transformer encoder architecture to overcome the above limitations. Each layer of the network integrates an eight-head self-attention mechanism and a high-dimensional feedforward neural network, capturing the multimodal dependencies of heat conduction on the blade surface through multi-head parallel computing. The self-attention module dynamically establishes global correlation weights between sparse temperature measurement points. For example, the energy transfer path between the heat source point at the blade root and the heat dissipation area at the blade tip is explicitly encoded as an attention probability distribution; the feedforward neural network performs nonlinear transformations on thermal features through a 2048-dimensional hidden layer and a GELU activation function, enhancing the model's ability to fit complex thermal boundary conditions. The six-layer stacked structure abstracts heat conduction features step by step. The bottom-layer network learns the local temperature gradient pattern, the middle-layer network integrates the thermal resistance effects of the cooling holes and the tenon, and the top-layer network finally constructs a complete heat flow transfer chain from the blade root to the blade tip.

[0130] This deep architecture demonstrates key advantages in typical turbine blade operating conditions. Regarding the thermal interaction between the leading edge cooling hole group and the tenon mounting surface, the multi-head attention mechanism automatically identifies the thermal shielding effect of the cooling airflow on the tenon area, and models the correlation between the cooling hole temperature mutation characteristics and the tenon contact thermal resistance; and for the heat diffusion process from the smooth area of ​​the blade pressure surface to the separation area of ​​the suction surface, the six-layer network accurately reconstructs the radial temperature attenuation curve through cross-regional feature fusion. Compared with the traditional three-layer convolutional network that can only process local neighborhoods, this design achieves physical consistency modeling of the global heat conduction path through cascaded self-attention layers, completely eliminating the temperature field distortion caused by the lack of long-range dependencies.

[0131] In another embodiment of the present invention, the traditional turbine blade temperature field reconstruction method often uses a multi-objective loss function with fixed weights, but the artificially preset weight coefficients are difficult to adapt to the dynamic changes of different loss items during the training process. The existing technology usually superimposes the adversarial loss, temperature fitting loss and physical constraint loss in a fixed proportion based on experience, for example, uniformly assigning equal weights to each item. This static combination can still maintain balance in the early stage of training, but when the model optimization enters the middle and late stages, the gradient amplitude of the physical constraint loss may be dozens of times that of the adversarial loss, resulting in the gradient update direction being dominated by a single loss. Typical manifestations are that the blade tenon area over-satisfies the heat conduction equation but loses the authenticity of the temperature details, or the cooling hole area accurately fits the temperature of the measuring point but violates the local energy conservation law, which ultimately causes the model to oscillate between different physical rules and cannot converge to a stable solution.

[0132] This solution innovatively utilizes the natural attenuation characteristics of the loss function and the synergistic mechanism of periodic learning rate adjustment to achieve a dynamic balance of weight effects. Although this solution explicitly stipulates fixed weight coefficients for the four losses, in the actual training process, the high-weight design of the physical constraint loss plays a dominant role in the early iteration stage, forcing the model to quickly satisfy the basic physical laws of the heat conduction equation; and as the learning rate is periodically halved, the gradient amplitude of the high-weight loss exhibits exponential decay, and the relative influence of the adversarial loss and the temperature fitting loss gradually increases. This differentiated emphasis on timing enables the model to focus on compliance with physical rules in the early stages of training and turn to optimization of temperature field details in the later stages, essentially avoiding the convergence path conflict caused by fixed weights.

[0133] Compared with the traditional scheme of manually adjusting weights, the core breakthrough of this method is to naturally resolve the contradiction of multi-objective optimization through training mechanism design rather than modifying the weight formula. Existing technologies attempt to use adaptive weight algorithms, such as dynamically adjusting coefficients according to the size of the loss value, but such methods require the introduction of additional hyperparameters and increase training uncertainty; while this scheme strictly adheres to the fixed weight combination and only achieves the temporal distribution of loss influence through the initial learning rate setting and periodic attenuation strategy. For example, when the learning rate first decays to 50% of the original value, the gradient contribution of the physical constraint loss is simultaneously reduced to 25% of the early strength. At this time, the generator automatically enhances the ability to depict micro-region details such as the air film holes on the leading edge of the blade against the loss, while keeping the thermal resistance effect of the tenon contact area in line with the laws of physics.

[0134] This strategy demonstrates significant advantages in reconstructing the complex thermal environment of turbine blades. In view of the specific distribution of the coexistence of smooth areas on the blade pressure surface and high-gradient areas on the cooling holes, high-multiple physical constraints under fixed weights ensure the continuity of the overall heat flow, while the periodically decaying learning rate enables the model to gradually release its microscale generation capabilities to combat losses in later iterations. Traditional methods, due to their difficulty in balancing the contradiction between global physical rules and the authenticity of local details due to fixed weights, often lead to excessive smoothing of the pressure surface area or artifacts at the edges of the cooling holes. This solution achieves stable convergence of temperature field reconstruction without modifying the loss formula through the inherent synergy between the training mechanism and fixed weights.

[0135] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network, characterized in that: include: Normalize and position-encode the sparse temperature measurement points on the turbine blade surface to generate a position-encoded feature vector, which is then concatenated with the normalized temperature value to form the input sequence for the Transformer encoder. The input sequence is processed by a multi-layer Transformer encoder, and the measurement point feature matrix is ​​output. It is then compressed into a global feature vector through a self-attention pooling operation. The global feature vector is decoded into a temperature field image through a U-Net generator, and the transposed convolution and skip connection features are integrated. Construct a multi-branch generative GAN discriminator, including image block processing branches at different scales, processing local details, transition areas, and overall areas respectively; A semantic segmentation model is used to generate a blade surface mask image, dividing the pixels into five semantically labeled regions: pressure surface, suction surface, tenon, cooling hole, and tip clearance. A semantic mask matrix is ​​introduced into the GAN discriminator, and the first, second, and third weight coefficients are applied to the cooling hole region, tenon region, and pressure surface region, respectively, for semantic weighting. The temperature field divergence is calculated based on the heat conduction equation to construct a physical constraint loss function; the heat flux density vectors at the boundaries of adjacent image blocks are extracted to construct a local energy conservation loss function; The loss function of the U-Net generator is defined as the weighted sum of the GAN loss, the absolute error loss of the temperature value, the physical constraint loss, and the local energy conservation loss. The loss function of the GAN discriminator is the sum of the GAN loss and the semantically weighted discriminant difference. The Adam optimizer is used to alternately train the generator and discriminator, with a periodic learning rate decay mechanism, and the training is terminated when the verification loss stabilizes. When the authenticity probability output by the GAN discriminator is greater than the set threshold and the physical constraint loss value is less than the set tolerance, the reconstructed temperature field is output.

2. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, characterized in that: Normalization and position coding include mapping the three-dimensional coordinates of the sparse temperature measurement points to the interval [0, 1] through a linear transformation formula, and using the sine-cosine position coding algorithm to generate a position coding feature vector.

3. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, characterized in that: The image block processing branches of different scales specifically include small-scale branches processing 8×8 image blocks with a 7×7 convolution kernel, medium-scale branches processing 16×16 image blocks with a 5×5 convolution kernel, and large-scale branches processing 32×32 image blocks with a 3×3 convolution kernel. The first, second, and third weight coefficients are 1.5, 1.2, and 0.8, respectively.

4. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The skip connection features come from the corresponding feature maps of the encoder level of the U-Net generator.

5. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The physical constraint loss function is constructed by calculating the temperature field divergence through automatic differentiation, and the heat conduction equation is , where the thermal conductivity k is set to 20 W / m·K.

6. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The learning rate periodic decay mechanism is set to an initial learning rate of 0.0001, which decays to 50% every 50 training cycles; the threshold is set to 0.9, the tolerance is set to 0.001, and the training is terminated when the validation set loss decreases by less than 1% for 10 consecutive training cycles. The U-Net generator outputs a temperature field image with a resolution of 512×512 pixels.

7. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The self-attention pooling operation calculates attention weights through a learnable query vector and weightedly aggregates the features of the measurement points.

8. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The semantic segmentation model adopts the DeepLabv3+ architecture, and the output resolution is consistent with the temperature field image.

9. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The Transformer encoder contains a 6-layer network. Each layer includes a multi-head self-attention module and a feedforward neural network. The number of heads in the multi-head self-attention module is 8, the hidden layer dimension of the feedforward neural network is 2048, and the activation function is GELU.

10. The method for reconstructing the global temperature field of a turbine blade based on a multimodal physical constraint network according to claim 1, wherein: The loss function of the U-Net generator satisfies: L G =L adv +5×L L1 +10×L phys +8×L flux Among them, L adv To counter the loss, the weight is 1; L L1 is the absolute error loss of temperature value, with a weight of 5; L phys is the physical constraint loss, with a weight of 10, L flux is the local energy conservation loss with a weight of 8.

Images