Method, device and storage medium for quantitatively monitoring thermal barrier coating cmas damage spallation

Abstract

The application provides a thermal barrier coating CMAS damage shedding quantitative monitoring method, equipment and a storage medium, and the method comprises the steps of: constructing an initial supervised training set based on an annotated CMAS corrosion time sequence image sequence, and training an initial time sequence segmentation model based on a combined segmentation loss function; using the initial time sequence segmentation model to infer an unannotated CMAS corrosion time sequence image sequence, combining difference calculation to obtain a preliminary shedding area prediction probability graph and a physical prior graph, performing weighted fusion and binary processing, and then applying a physical constraint to obtain a pseudo-label sample set; a mixed training set containing the initial supervised training set and the pseudo-label sample set is constructed, the initial time sequence segmentation model is retrained based on a mixed loss function, an optimized final model is obtained, and a CMAS corrosion image sequence to be processed is segmented to obtain a shedding area mask. The method can realize high-precision segmentation of a large-scale unannotated image sequence only by a small amount of manual annotation, and can effectively extract corrosion kinetic parameters.

Description

Technical Field

[0001] This invention belongs to the field of material failure analysis and intelligent monitoring technology, specifically relating to a method, equipment and storage medium for quantitative monitoring of damage and detachment of thermal barrier coatings (CMAS). Background Technology

[0002] In high-temperature components such as aero engines and gas turbines, thermal barrier coatings (TBCs) are crucial barriers protecting the base metal from high-temperature oxidation and corrosion. However, during service, calcium magnesium aluminum silicate (CMAS) formed from the melting of environmental substances such as dust and volcanic ash can penetrate and erode the TBC, leading to irreversible spalling of the coating. This is one of the main mechanisms of TBC failure.

[0003] Currently, monitoring of the CMAS corrosion process mainly relies on the following two methods:

[0004] 1) Destructive testing: such as scanning electron microscopy (SEM), X-ray computed tomography (XCT), etc. Although these methods can provide microstructural information, they require interruption of the experiment and sample preparation, and cannot obtain continuous, in-situ kinetic data.

[0005] 2) Manual or semi-automatic image analysis: Time-series images of the corrosion process are recorded using an optical camera, and researchers then manually delineate the detachment area or use a simple thresholding method. This method is extremely inefficient, highly subjective, and difficult to guarantee consistency over time, failing to meet the need for accurate and objective quantification of key dynamic parameters such as detachment rate and boundary advancement speed.

[0006] Existing automated image segmentation methods (such as U-Net, ConvLSTM, and other deep learning models) typically require a large amount of high-quality manually labeled data for training. However, performing frame-by-frame fine-grained annotation on thousands of eroded image sequences is an extremely time-consuming and expensive task, constituting a significant bottleneck for practical applications. Furthermore, traditional optical flow methods (such as RAFT and TV-L1) assume pixel intensity conservation and rely on significant displacement fields, while the CMAS erosion process exhibits slow, monotonous texture and brightness gradients, lacking obvious pixel motion. This results in poor performance and high computational cost for optical flow methods in such scenarios.

[0007] Therefore, there is an urgent need for an intelligent monitoring method that can efficiently, accurately, and robustly identify and quantify TBC detachment areas from optical image sequences with limited labeled data. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, and storage medium for quantitative monitoring of damage and detachment of thermal barrier coatings (CMAS) based on self-evolving pseudo-labels. This method ingeniously integrates prior physical knowledge with deep learning technology. Through a self-evolving semi-supervised learning framework, it can achieve high-precision segmentation of large-scale unlabeled image sequences with only a small amount of manual annotation and effectively extract corrosion kinetic parameters.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The first aspect of this invention provides a method for quantitative monitoring of damage and detachment of thermal barrier coatings (CMAS), comprising:

[0011] Obtain multiple sets of labeled CMAS erosion time-series image sequences to construct an initial supervised training set;

[0012] The initial temporal segmentation model is trained using the initial supervised training set based on the combined segmentation loss function, which is composed of a weighted sum of binary cross-entropy loss, Dice loss, and monotonicity constraint loss.

[0013] The trained initial temporal segmentation model is used to infer the unlabeled CMAS erosion temporal image sequence to obtain a preliminary probability map of the detachment area, and a physical prior map is generated by calculating the absolute brightness difference between adjacent frames of the same unlabeled CMAS erosion temporal image sequence.

[0014] The preliminary detachment region prediction probability map and the physical prior map are weighted and fused in logit space to generate a fused probability map. After binarization, physical constraints are applied to obtain a pseudo-label sample set.

[0015] Construct a hybrid training set containing the initial supervised training set and the pseudo-label sample set, and use the hybrid training set to retrain the initial temporal segmentation model based on the hybrid loss function to obtain the optimized final model. The hybrid loss function is composed of a weighted sum of binary cross-entropy loss, Dice loss, mask segmentation loss and monotonicity constraint loss.

[0016] The final model is used to segment the CMAS erosion image sequence to be processed, and the detachment region mask of each frame is obtained.

[0017] Based on the mask of the detached area, the target corrosion kinetic parameters are fitted and extracted by calculating the curve of the detached area changing over time.

[0018] Furthermore, the initial temporal segmentation model adopts an encoder-decoder architecture, where the encoder is a ResNet-18 backbone network and the decoder integrates a ConvLSTM module to model temporal dependencies; the model takes a single frame or multiple consecutive frames of images as input and outputs a probability map of the same size to represent the confidence level of each pixel belonging to the missing region.

[0019] Furthermore, the expression for the combined segmentation loss function is as follows:

[0020]

[0021] in, and These are the weighting coefficients for the corresponding loss terms; For the corresponding initial supervised training set The combined segmentation loss is defined as the weighted sum of the binary cross-entropy loss and the Dice loss:

[0022]

[0023] Among them, the weighting coefficient This is used to balance the two losses;

[0024] The binary cross-entropy loss is defined as:

[0025]

[0026] The Dice loss is defined as:

[0027]

[0028] in For the first True label confidence per pixel This represents the predicted shedding probability from the model. The total number of pixels in the image. To prevent division by zero of small constants;

[0029] The monotonic constraint loss is defined as:

[0030]

[0031] in and These represent the model at the 1st and 2nd. The and the first Spatial position under thermal shock cycles Predicted detachment probability This represents the total number of loops.

[0032] Furthermore, calculating the absolute brightness difference between adjacent frames in the same unlabeled CMAS erosion time-series image sequence to generate a physical prior map includes:

[0033] For the same unlabeled CMAS corrosion time series image sequence, the first... Frame and the grayscale image of a frame and Calculate its pixel-wise absolute difference: This forms a difference graph;

[0034] The difference graph is subjected to morphological closing operation to connect neighboring response regions and normalized to the [0,1] interval to obtain the physical prior graph.

[0035] Furthermore, the preliminary predicted probability map of the detachment area and the physical prior map are weighted and fused in the logit space to generate a fused probability map, including:

[0036] The initial probability map of the detachment area is converted into a logit form, and the physical prior map is linearly mapped to the same scale and used as an auxiliary logit term.

[0037] The two are weighted and fused according to preset weight coefficients, and then transformed back into probability space by the sigmoid function to obtain the fused probability map.

[0038] Furthermore, the physical constraints include time smoothing constraints and monotonicity constraints.

[0039] Furthermore, the expression for the hybrid loss function is as follows:

[0040]

[0041] in, These are the weighting coefficients for the corresponding loss terms; The masking segmentation loss corresponds to the pseudo-label sample set. The combined segmentation loss is calculated only at pixel locations where the false label confidence is higher than a preset threshold one or lower than a preset threshold two, while the remaining pixels are masked.

[0042] Furthermore, the target corrosion kinetic parameters include any or all of the following: total detachment area, local corrosion rate, or boundary average propagation velocity.

[0043] Furthermore, the method also includes experimentation and image acquisition steps, specifically including:

[0044] Test samples were prepared, and high-temperature thermal shock tests were conducted based on a pre-set experimental procedure. Sample surface images were collected to obtain a time-series image sequence reflecting the coating corrosion and peeling process.

[0045] Several sets of acquired time-series image sequences were manually labeled to obtain labeled CMAS corrosion time-series image sequences.

[0046] A second aspect of the present invention provides an electronic device comprising:

[0047] Memory, used to store computer programs;

[0048] A processor for executing a computer program stored in the memory to implement the method described in the first aspect above.

[0049] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0050] Compared with the prior art, the present invention has the following outstanding advantages:

[0051] 1) High accuracy and strong robustness: By fusing brightness difference physical priors with deep learning predictions and introducing a self-evolutionary optimization mechanism, this invention significantly outperforms traditional deep learning models, optical flow methods, and simple thresholding methods in terms of Dice coefficients and boundary IoU. Furthermore, the obtained model exhibits excellent generalization ability across temperature ranges (1200℃, 1250℃, 1300℃) and under different CMAS composition conditions.

[0052] 2) High efficiency and low cost: The core of this invention is a semi-supervised learning framework that requires only a small number (approximately 30 sets) of manually labeled sequences to start, greatly reducing the cost and time of data labeling. The pseudo-label generation process is computationally efficient (approximately 0.45 seconds / sequence), making it suitable for large-scale data analysis.

[0053] 3) Physical consistency guarantee: By explicitly introducing the key physical constraint of "monotonicity", the model prediction results are ensured to strictly conform to the physical nature of irreversible CMAS corrosion, so that the extracted kinetic parameters (such as shedding rate) have a high degree of scientific credibility.

[0054] 4) Automation and Quantifiability: It realizes end-to-end fully automated analysis from raw images to corrosion kinetic parameters, providing materials scientists with objective, accurate and continuous quantitative tools, completely eliminating the dependence on subjective and inefficient manual measurement, and providing strong technical support for in-depth research on TBC failure mechanisms and accelerating the development of new materials. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an embodiment of the monitoring method of the present invention.

[0056] Figure 2This is a schematic diagram of the model training and quantization analysis process in an embodiment of the present invention. Detailed Implementation

[0057] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0058] See Figure 1 and Figure 2 This invention provides a method for quantitative monitoring of damage and detachment of thermal barrier coatings (CMAS), comprising the following steps:

[0059] S1. Obtain multiple sets of labeled CMAS erosion time-series image sequences to construct an initial supervised training set.

[0060] In one example, under a high-temperature thermal shock test at 1250℃, optical images of the specimen surface were acquired at the end of each thermal cycle cooling stage using a high-speed industrial camera, forming a time-series image sequence. Then, 30 complete sequences were randomly selected from the entire time-series image sequence, and professionals used pixel-level annotation tools to meticulously delineate the CMAS-induced ablation regions in each frame, generating a binary mask as the ground truth label set. , and the corresponding image set Constructing the initial supervised training set .

[0061] S2. Train the initial temporal segmentation model using the initial supervised training set based on the combined segmentation loss function.

[0062] In one example, the initial temporal segmentation model employs a three-stage temporal encoding-memory-decoding architecture, specifically including: the encoder, with ResNet-18 as the backbone network, processes the input single-frame or multi-frame continuous image sequence. Spatial feature extraction is performed to obtain the coded features of each frame. The temporal memory unit encodes the features of each frame. The input is fed into a convolutional long short-term memory (ConvLSTM) network to model dynamic evolutionary relationships in the spatiotemporal domain, and outputs a temporally consistent sequence of hidden states. The decoder receives the hidden state of the ConvLSTM output via a U-Net-style upsampling path. By combining the skip connections of the corresponding encoder levels, a high-resolution probability map of the missing regions is reconstructed frame by frame. .

[0063] The input to the initial temporal segmentation model described above is of length [length missing]. Image sequences (e.g.) (frames), the output is a frame-by-frame segmentation probability map of the same length; the entire network is trainable end-to-end, and the ConvLSTM convolutional kernel size is... The number of hidden channels matches the number of output channels of the last layer of the encoder (e.g., 256).

[0064] As a preferred implementation, the combined segmentation loss function used during training consists of a weighted sum of binary cross-entropy loss (BCE), Dice loss, and monotonicity constraint loss, in order to simultaneously optimize pixel classification accuracy and overall region overlap performance.

[0065] In one example, the expression for the combined segmentation loss function is as follows:

[0066]

[0067] in, and These are the weighting coefficients for the corresponding loss terms; For the corresponding initial supervised training set The combined segmentation loss is defined as the weighted sum of the binary cross-entropy loss and the Dice loss:

[0068]

[0069] Among them, the weighting coefficient This is used to balance the two losses;

[0070] The binary cross-entropy loss is defined as:

[0071]

[0072] Dice loss is defined as:

[0073]

[0074] in For the first True label confidence per pixel This represents the predicted shedding probability from the model. The total number of pixels in the image. To prevent division by zero of small constants;

[0075] The monotonic constraint loss is defined as:

[0076]

[0077] in and These represent the model at the 1st and 2nd. The and the first Spatial position under thermal shock cycles Predicted detachment probability The total number of cycles is represented by this loss term, which penalizes the local decrease in the predicted damage probability between any adjacent thermal cycles, thereby forcing the coating peeling area to remain constant with the number of thermal cycles, reflecting the irreversible physical characteristics of the CMAS corrosion process.

[0078] S3. Use the trained initial temporal segmentation model to infer the unlabeled CMAS erosion temporal image sequence to obtain a preliminary probability map of the detachment region, and generate a physical prior map by calculating the absolute brightness difference between adjacent frames of the same unlabeled CMAS erosion temporal image sequence.

[0079] In one example, the remaining thousands of unlabeled images selected from S1 are input into the trained initial temporal segmentation model in their original temporal order. The model outputs a preliminary probability map of the missing regions frame by frame. This is recorded as the preliminary prediction result of the detachment area.

[0080] In one illustrated example, for the first image in the same image sequence... Frame and the grayscale image of a frame and Calculate its pixel-wise absolute difference: This forms a difference graph;

[0081] The difference map is then subjected to a morphological closing operation (structuring element radius of 3 pixels) to connect neighboring response regions and normalized to the [0,1] interval to obtain the physical prior map. It is used to characterize the local dynamic changes induced by the CMAS reaction.

[0082] S4. The above preliminary detachment area prediction probability map and physical prior map are weighted and fused in logit space to generate a fused probability map. After binarization, physical constraints are applied to obtain a pseudo-label sample set.

[0083] In one example, the initial probability map of the detachment area is first predicted. Convert to logit format:

[0084]

[0085] Then the physical prior diagram After linear mapping to the same scale, it is used as an auxiliary logit term;

[0086] The two are weighted by coefficients. Weighted fusion:

[0087]

[0088] The probability map is then transformed back into probability space using the sigmoid function, resulting in the fused probability map. This fusion strategy combines data-driven statistical confidence with physical interpretability.

[0089] In one example, the fused probability map Binarization is performed using a threshold of 0.5 to generate initial pseudo-labels. Two physical constraints were then imposed: (i) Temporal smoothing constraint: Perform 3-frame median filtering on short-time jitter regions to eliminate non-physical flicker; (ii) Monotonicity constraint: mandatory satisfaction That is, the detached area of ​​the current frame must contain all the detached pixels of the previous frame to ensure that the damage evolution is irreversible (that is, the detached area can only expand or remain unchanged over time, and cannot shrink), so as to ensure that it conforms to the physical laws of CMAS erosion.

[0090] Finally, a high-quality pseudo-label sample set was obtained. .

[0091] S5. Construct a hybrid training set containing the initial supervised training set and the pseudo-label sample set, and use the hybrid training set to perform end-to-end joint retraining of the initial temporal segmentation model based on the hybrid loss function to obtain the optimized final model.

[0092] In one example, the initial supervised training set is shown. With pseudo-label sample set The sets are merged into a hybrid training set; a hybrid loss function is then used for retraining.

[0093] In one example, the expression for the hybrid segmentation loss function is as follows:

[0094]

[0095] in, , The meaning of has been explained above and will not be repeated here. These are the weighting coefficients for the corresponding loss terms; The mask combination segmentation loss corresponds to the pseudo-label sample set. The combined segmentation loss is expressed as follows:

[0096]

[0097]

[0098]

[0099] in For the first The pseudo-label confidence score of each pixel, and the combined segmentation loss only applies to the pseudo-label confidence score. >0.9 or Loss is calculated on pixels with a value <0.1, and the remaining pixels are masked to suppress noise interference from pseudo-labels in low-confidence regions, thus forming a "self-evolving" optimization loop that gradually improves the quality of pseudo-labels and the performance of the model.

[0100] After the above self-evolutionary training, a more robust final temporal segmentation model can be obtained.

[0101] S6. Use the final temporal segmentation model trained in S5 to segment any new CMAS erosion image sequence to obtain the detachment region mask for each frame.

[0102] In one example, newly acquired CMAS erosion time-series images are input into the final time-series segmentation model, and the model outputs a frame-by-frame high-precision binary mask. It accurately identifies the spatial distribution of the detached areas.

[0103] S7. Based on the high-precision binary mask obtained in S6, which is the mask of the detached area, the target corrosion kinetic parameters are fitted and extracted by calculating the curve of the detached area changing over time.

[0104] In one example shown, the mask can be applied to each frame. Count the number of white pixels and convert them into the actual total area of ​​peeling. And plot the detachment area-number of thermal cycles curve. .

[0105] Further methods can be employed, such as numerical differentiation or curve fitting (e.g., power-law models). Extract key kinetic parameters, including: corrosion rate constant. Response index The average advance rate at the boundary is used to quantitatively evaluate the coating's resistance to CMAS and predict its lifetime.

[0106] The image data used in this invention comes from high-temperature thermal shock tests. The embodiments of this invention further illustrate the high-temperature thermal shock test and sample preparation, as well as the image acquisition steps.

[0107] The experiment was conducted on a dedicated thermal cycling test and image data acquisition platform, and the test specimen was a standard circular thermal barrier coating sample.

[0108] Before the test, CMAS powder was dissolved in anhydrous ethanol at a surface density of 10 mg / cm², coated evenly on the surface of the test piece, and heat-treated at 1200℃ for a short time (5 minutes) to enhance adhesion, and then cooled to room temperature.

[0109] During the formal test, the test piece is fixed in the fixture, and parameters such as the target temperature (1250℃, with a temperature range of 1200–1300℃), holding time, cooling method, and number of cycles are set. Each thermal cycle includes three stages: heating, holding, and cooling.

[0110] At the end of the cooling phase of each cycle, an industrial camera at a fixed position automatically captures images of the test specimen surface, forming a CMAS corrosion time sequence image reflecting the coating corrosion and peeling process.

[0111] Another embodiment of the present invention discloses an electronic device, comprising:

[0112] Memory, used to store computer programs;

[0113] The processor is used to execute the program stored in the memory to implement the steps of the above embodiments of the quantitative monitoring method for damage and desquamation of thermal barrier coatings (CMAS).

[0114] The memory of the electronic device mentioned in this embodiment may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device.

[0115] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0116] Another embodiment of the present invention discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above embodiments of the quantitative monitoring method for damage and detachment of thermal barrier coatings (CMAS).

[0117] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and the multiple devices will interact with each other to complete the method described.

[0118] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0119] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A method for quantitatively monitoring CMAS damage spallation of thermal barrier coatings, characterized in that, include: Obtain multiple sets of labeled CMAS erosion time-series image sequences to construct an initial supervised training set; The initial temporal segmentation model is trained using the initial supervised training set based on the combined segmentation loss function, which is composed of a weighted sum of binary cross-entropy loss, Dice loss, and monotonicity constraint loss. The trained initial temporal segmentation model is used to infer the unlabeled CMAS erosion temporal image sequence to obtain a preliminary probability map of the detachment area, and a physical prior map is generated by calculating the absolute brightness difference between adjacent frames of the same unlabeled CMAS erosion temporal image sequence. The preliminary detachment region prediction probability map and the physical prior map are weighted and fused in logit space to generate a fused probability map. After binarization, physical constraints are applied to obtain a pseudo-label sample set. Construct a hybrid training set containing the initial supervised training set and the pseudo-label sample set, and use the hybrid training set to retrain the initial temporal segmentation model based on the hybrid loss function to obtain the optimized final model. The hybrid loss function is composed of a weighted sum of binary cross-entropy loss, Dice loss, mask segmentation loss and monotonicity constraint loss. The final model is used to segment the CMAS erosion image sequence to be processed, and the detachment region mask of each frame is obtained. Based on the mask of the detached area, the target corrosion kinetic parameters are fitted and extracted by calculating the curve of the detached area changing over time.

2. The method of claim 1, wherein, The initial temporal segmentation model adopts an encoder-decoder architecture, where the encoder is a ResNet-18 backbone network and the decoder integrates a ConvLSTM module to model temporal dependencies. The model takes a single frame or multiple consecutive frames of images as input and outputs a probability map of the same size to represent the confidence that each pixel belongs to the missing region.

3. The method of claim 2, wherein the method is characterized by, The expression for the combined segmentation loss function is as follows: wherein, and are the weighting coefficients for the corresponding loss terms; is the combined segmentation loss for the initial supervised training set is defined as the weighted sum of binary cross-entropy loss and Dice loss: Among them, the weighting coefficient This is used to balance the two losses; The binary cross-entropy loss is defined as: The Dice loss is defined as: in For the first True label confidence per pixel This represents the predicted shedding probability from the model. The total number of pixels in the image. To prevent division by zero of small constants; The monotonic constraint loss is defined as: in and These represent the model at the 1st and 2nd. The and the first Spatial position under thermal shock cycles Predicted detachment probability This represents the total number of loops.

4. The method for quantitative monitoring of CMAS damage and detachment of thermal barrier coatings according to claim 1, characterized in that, Calculating the absolute brightness difference between adjacent frames of the same unlabeled CMAS erosion temporal image sequence to generate a physical prior map includes: For the same unlabeled CMAS corrosion time series image sequence, the first... Frame and the grayscale image of a frame and Calculate its pixel-wise absolute difference: This forms a difference graph; The difference graph is subjected to morphological closing operation to connect neighboring response regions and normalized to the [0,1] interval to obtain the physical prior graph.

5. The method for quantitative monitoring of CMAS damage and detachment of thermal barrier coatings according to claim 1, characterized in that, The preliminary predicted probability map of the detachment area and the physical prior map are weighted and fused in logit space to generate a fused probability map, including: The initial probability map of the detachment area is converted into a logit form, and the physical prior map is linearly mapped to the same scale and used as an auxiliary logit term. The two are weighted and fused according to preset weight coefficients, and then transformed back into probability space by the sigmoid function to obtain the fused probability map.

6. The method for quantitative monitoring of CMAS damage and detachment of thermal barrier coatings according to claim 1, characterized in that, The physical constraints include time smoothness constraints and monotonicity constraints.

7. The method for quantitative monitoring of CMAS damage and detachment of thermal barrier coatings according to claim 3, characterized in that, The expression for the hybrid loss function is as follows: in, These are the weighting coefficients for the corresponding loss terms; The masking segmentation loss corresponds to the pseudo-label sample set. The combined segmentation loss is calculated only at pixel locations where the false label confidence is higher than a preset threshold one or lower than a preset threshold two, while the remaining pixels are masked.

8. The method for quantitative monitoring of CMAS damage and detachment of thermal barrier coatings according to claim 1, characterized in that, The target corrosion kinetic parameters include any or all of the following: total detachment area, local corrosion rate, or boundary average propagation velocity.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory to implement the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images