Blade material damage life prediction method based on SCConv-LiteMLA-VIT network

Abstract

The invention discloses a blade material damage life prediction method based on an SCConv-LiteMLA-VIT network. The method comprises the steps that S1, cavitation morphology pictures and weightlessness data of a blade material at each stage moment t are acquired; s2, a training set and a verification set are divided according to the proportion of 8: 2 through the data set obtained in the S1, training is conducted on the constructed improved SCConv-LiteMLA-VIT network, and a trained model is stored and used for cavitation damage degree judgment of a blade material; and S3, inputting a blade material damage area picture needing to be identified into the trained SCConv-LiteMLA-VIT model obtained in the S2, obtaining a cavitation life coefficient zeta of the image, and completing life prediction of the material according to the cavitation life coefficient zeta. According to the method, the cavitation degree of the blade material can be identified, the identification speed is high, and the maintenance efficiency can be greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of blade materials, and relates to the identification of the degree of material surface damage such as cavitation and water erosion, and is specifically a blade material damage life prediction method based on an SCConv-LiteMLA-VIT network. Background Art

[0002] Cavitation and water erosion are common forms of material damage in liquid power environments, and are widely found in flow-through components such as cooling water pumps, hydraulic systems, engine cylinder liners, and steam turbines. During the operation of these devices, pressure changes or high-speed impact of droplets cause cavitation collapse, and impact the surface of surrounding materials in the form of shock waves and microjets, which in turn cause plastic deformation and strength damage to the components, and a large number of honeycomb damage often appears on the surface. This not only affects the normal working efficiency of the equipment, but may also seriously endanger the reliability of the components and the service life of the equipment. Therefore, how to prevent cavitation and water erosion from damaging components has become an important issue that needs to be urgently solved in the design of flow-through components. Summary of the invention

[0003] In order to overcome the above-mentioned technical problems, the purpose of the present invention is to provide a life prediction method for blade material damage evolution based on an improved SCConv-LiteMLA-VIT network. This method does not rely on expert experience evaluation, nor does it require cumbersome modeling and theoretical calculations, and can achieve relatively accurate identification of the degree of post-corrosion damage of blade materials and prediction of remaining life.

[0004] The present invention is achieved by adopting the following technical solutions:

[0005] A life prediction method for blade material damage evolution based on an improved SCConv-LiteMLA-VIT network comprises the following steps:

[0006] Step S1, collecting and obtaining damage morphology images and weight loss data of blade materials at each stage and time t;

[0007] Step S2, the data set obtained in step S1 is divided into a training set and a validation set in a ratio of 8:2, and the training is performed on the constructed improved SCConv-LiteMLA-VIT network model, and the trained model is saved for the damage degree discrimination of the blade material;

[0008] Step S3: Input the image of the damaged area of ​​the blade material to be identified into the trained SCConv-LiteMLA-VIT network model obtained in step S2, obtain the life coefficient ζ of the image, and complete the life prediction of the blade material based on the life coefficient ζ.

[0009] The method of the present invention can predict the RUL of blade materials, avoid sudden failure thereof, and ensure stable and efficient operation of equipment.

[0010] The improved SCConv-LiteMLA-VIT network model described in step S2 is as follows:

[0011] SCConv-LiteMLA-VIT network model: The size of the original input image A is 3×224×224; after the channel increase module, the number of channels of the input image is increased from 3 to 16, and the output size becomes 224×224×16; then, the image passes through the spatial channel convolution module to enhance its spatial and channel information, and the output size remains at 224×224×16; then, the image enters the multi-scale linear attention module, and the output size is 224×224×3; before the image generates tiles, the data passes through the tile embedding module to divide the input image into fixed-size tiles, each of which is 16×16 in size, and the image is divided into 14×14 tiles in total. Each tile is expanded into a one-dimensional vector, and its dimension is increased to 192 through linear transformation; then the positional embedding (Positional Embeddings) are added to the tiles to form a tile sequence containing position information. The final input information is 1×198×192, where 198 includes 196 tiles, 1 classification tag and 1 distillation tag; the tile sequence passes through 12 Transformer encoder layers, each of which contains multi-head attention, normalization and multi-layer perceptron; the output information of each encoder layer is 1×198(196+1+1)×192. After being processed by 12 encoder layers, the CLS class tag represents the global features of the entire sequence in the entire model, and these features are extracted for the final classification task; the multi-layer perceptron layer further processes these global features, and the final input information is 1×1×192; finally, the model passes the summarized CLS class tag to the classifier, which consists of one or more Dense Layers and a softmax activation function. The softmax activation function converts the CLS class tag output into a probability distribution, that is, the category with the highest score calculated by the softmax function is the category that the model predicts the image is most likely to belong to.

[0012] The constructed SCConv-LiteMLA-VIT network model is trained as follows:

[0013] Under the ultra-depth microscope at 300 times magnification, the image pixels taken were 1600×1200, and the areas with the most serious damage to the leaf material were collected; then, each image in the training set and the test set was randomly cropped 20 times to a pixel size of 448×448; during the training process, the training set and the validation set were normalized uniformly, and the 448×448 pixel images were randomly cropped and scaled to 224×224 using data enhancement, and then Gaussian blur processing was performed on them with a blur kernel of 5×5, and randomly rotated within 15°, and Mixup and CutMix operations were performed on the images to simulate images collected in a real environment; the network was trained by setting the AdamW optimizer, variable learning rate, and cross entropy loss function.

[0014] The model structure designed by the present invention is based on the image classification architecture design of Transformer. In view of the problem that VIT is not able to capture local features, the model integrates a multi-scale linear attention (LiteMLA) module and a spatial channel convolution (SCConv) module. This design enables the model to effectively perform detailed predictions at high resolution, while also having the ability to learn global perception fields and multi-scale features, while maintaining a high hardware operation efficiency. First, the model includes a channel increase module, which increases the number of channels of the input image from 3 to 16 through a convolutional layer. Then, spatial channel convolution is performed through the SCConv component to enhance spatial and channel information. Next, the model introduces LiteMLA, a lightweight multi-scale linear attention mechanism for capturing long-distance dependencies. Before the image generates a tile, position embeddings are added to the input patch to provide position information for the model. In addition, a learnable tag (CLS token) representing the entire image is introduced. The core part of the model is the Transformer structure with multiple Transformer layers, which consists of 12 Transformer encoder layers, including multi-head attention mechanism, multi-layer perceptron (MLP) and normalization components. The multi-head attention mechanism allows the model to focus on multiple areas of the image patch in parallel, capturing diverse contextual associations; normalization stabilizes the learning process by standardizing the data; and the multi-layer perceptron processes the attention features and adds nonlinearity to recognize complex patterns. After feature extraction and transformation by the encoder, the model summarizes the output and inputs it into the classifier. The classifier part usually consists of a dense layer followed by a softmax activation function to complete the final classification task. This model structure reflects the current development trend of AI, which aims to take advantage of the advantages of different methods to improve performance and efficiency, especially when processing complex inputs such as images. This model is becoming increasingly important in fields that require sophisticated visual recognition capabilities, such as medical imaging, autonomous driving, and content review.

[0015] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0016] First, the blade material cavitation degree identification method described in the present invention can accurately judge the cavitation state of the blade material cavitation area. It does not rely on expert experience and relevant knowledge, does not require complex modeling and theoretical calculation processes, and only needs to be based on the SCConv-LiteMLA-VIT network model, by collecting its cavitation morphology pictures and weight loss data under cavitation for training, it can achieve accurate judgment of the cavitation state of the blade material and ensure the stable and efficient operation of the equipment.

[0017] Second, the SCConv-LiteMLA-VIT network model used in the present invention can realize the recognition of the microscopic surface cavitation morphology state of the cavitation area of ​​the blade material. By introducing the LiteMLA attention mechanism and the spatial channel module (SCConv), the recognition accuracy is improved; at the same time, the use of data enhancement technology improves the generalization and robustness of the model. The model has high accuracy and fast operation speed, and can apply the trained model without complex modeling and theoretical calculations to realize the recognition of the degree of cavitation damage of the blade material.

[0018] Third, the SCConv-LiteMLA-VIT network model used in the present invention has a network structure that can extract more and richer feature information without changing the network depth, so that it can accept a wider range of inputs and more spatial information. The model can achieve early warning and preventive maintenance of equipment from the cavitation damage degree and RUL of the blade material obtained by identification.

[0019] Fourth, the improved SCConv-LiteMLA-VIT network model used in the present invention has higher accuracy and lower loss value on the cavitation damage dataset and the CIFAR10 public dataset. The SCConv-LiteMLA-VIT network model achieved good results on the CIFAR10 dataset. Compared with the VIT-Distilled network model, the accuracy was improved by 1.1%, and the loss value was reduced from 0.8131 to 0.7625. At the same time, the model also achieved good results on the self-made cavitation damage dataset. Its test set accuracy reached 95.9%, and the loss value was reduced to 1.0786. Compared with the original model, its accuracy was improved by 4.1%, and the loss value was reduced by 0.2026.

[0020] The method of the invention is reasonably designed and can accurately identify the degree of cavitation damage of blade materials. The model has a fast recognition speed, which can greatly improve the maintenance efficiency and avoid equipment failure caused by sudden failure of blade materials, which may cause heavy casualties and economic losses. Therefore, it has good practical application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 A schematic flow diagram showing the method of the present invention.

[0024] Figure 2 A schematic diagram showing the division of cavitation characteristic curves used in the method of the present invention for predicting the cavitation damage life of blade materials.

[0025] Figure 3 A diagram showing the model design for predicting the cavitation damage life of blade materials using the method of the present invention.

[0026] Figure 4 A comparison chart showing the validation set accuracy of the SCConv-LiteMLA-VIT network model used in the method of the present invention on a self-made cavitation damage dataset.

[0027] Figure 5 A comparison chart showing the validation set loss values ​​of the SCConv-LiteMLA-VIT network model used in the method of the present invention on a self-made cavitation damage dataset.

[0028] Figure 6 A comparison chart showing the validation set accuracy of the SCConv-LiteMLA-VIT network model used in the method of the present invention on the CIFAR10 public dataset.

[0029] Figure 7 A comparison chart showing the validation set loss values ​​of the SCConv-LiteMLA-VIT network model used in the method of the present invention on the CIFAR10 public dataset.

[0030] Figure 8 A graph showing partial prediction results of the life coefficient ζ of the blade material using the SCConv-LiteMLA-VIT network model by the method of the present invention.

[0031] Fig. 9 Schematic diagram showing the attention visualization before improvement in an example of the present invention.

[0032] Fig.10 Schematic diagram showing the improved attention visualization in an example of the present invention. DETAILED DESCRIPTION

[0033] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0034] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0035] like Figure 1 As shown, the life prediction method of blade material damage evolution based on SCConv-LiteMLA-VIT network described in the present invention includes: performing an accelerated cavitation simulation experiment on the blade material to obtain a cavitation micromorphology image and its weight loss data, and drawing a cumulative mass loss curve; fitting the cumulative mass loss curve with a Logistic equation, and obtaining the turning point t2 through a mapping method, and thereby obtaining the cavitation life coefficient ζ; dividing the original collected images into a training set and a validation set in a ratio of 8:2, and expanding the data set through random cropping, and using the cavitation life coefficient ζ as the label of the training set and the validation set. The preprocessed data set is loaded into the improved SCConv-LiteMLA-VIT network model; the generalization and robustness of the model are improved by data enhancement; the corresponding training parameters are set, such as AdamW optimizer, variable learning rate and appropriate batch size, and finally the real cavitation photos on the blade material are loaded into the trained model file to obtain the cavitation life coefficient ζ. When ζ=1, it is considered that the cavitation damage of the blade material is serious and the material needs to be repaired or replaced; when ζ<1, it can be calculated according to the formula t r =t0(1-ζ) / ζ, calculate its RUL; where t0 is the time the blade material has been used in the actual operation process, t r is the RUL of the material. According to the measured RUL, early warning and preventive maintenance of the blade material are achieved. The present invention adopts the SCConv-LiteMLA-VIT network model, which can accurately identify the degree of cavitation damage of the blade material at a fast speed, does not require relevant knowledge and complex modeling and analysis processes, and does not require rich experience of experts, and can accurately identify the degree of cavitation damage of the blade material and predict its RUL.

[0036] like Figure 3As shown, the model structure designed by the present invention is based on the image classification architecture design of Transformer. In view of the problem that VIT is not able to capture local features, the model integrates a multi-scale linear attention (LiteMLA) module and a spatial channel convolution (SCConv) module. This design enables the model to effectively perform detailed predictions at high resolution, while also having the ability to learn global perception fields and multi-scale features, while maintaining a high hardware operation efficiency. First, the model includes a channel increase module, which increases the number of channels of the input image from 3 to 16 through a convolutional layer. Then, spatial channel convolution is performed through the SCConv component to enhance spatial and channel information. Next, the model introduces LiteMLA, a lightweight multi-scale linear attention mechanism for capturing long-distance dependencies. Before the image generates a tile, position embeddings are added to the input patch to provide position information for the model. In addition, a learnable tag (CLS) representing the entire image is introduced. The core part of the Transformer model is composed of multiple Transformer layers stacked together. Each Transformer structure consists of 12 Transformer encoder layers, including multi-head attention mechanism, multi-layer perceptron (MLP) and normalization components. The multi-head attention mechanism allows the model to focus on multiple areas of the image patch in parallel to capture diverse contextual associations; normalization stabilizes the learning process by standardizing the data; and the multi-layer perceptron processes the attention features and adds nonlinearity to recognize complex patterns. After feature extraction and transformation by the encoder, the model summarizes the output and inputs it into the classifier. The classifier part usually consists of a dense layer followed by a softmax activation function to complete the final classification task.

[0037] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0038] The embodiment of the present invention takes the cavitation process as an example and provides a life prediction method for blade material damage evolution based on an improved SCConv-LiteMLA-VIT network, comprising the following steps:

[0039] (1) Simulation experiment and data collection: Based on the ultrasonic cavitation tester, the cavitation process of the blade material is simulated; a number of test pieces of the blade material are made. During the accelerated simulation experiment, the test pieces are removed at regular intervals, cleaned with an ultrasonic cleaner, and then dried; then, the test pieces are placed on an electronic balance with tweezers, and their mass is weighed multiple times and the average value is recorded to obtain the discrete points of cumulative mass loss.

[0040] At the same time, at the same time interval, the cleaned sample was placed on the stage of the ultra-depth-of-field microscope and observed at 300 times to take multiple morphological images of the most serious cavitation area. In the cavitation process, each time t divided needs to correspond to a series of cavitation morphological images to enhance the robustness of the model; the collected weightlessness data is fitted to draw the cavitation characteristic curve and determine the turning point t2; through the turning point t2, the cavitation life coefficient ζ corresponding to each time t is determined, and a data set with the cavitation life coefficient ζ as the label is obtained.

[0041] Among them, the turning point is determined by drawing on the cavitation characteristic curve; the discrete points of the cumulative mass loss of the material during the experiment are fitted by the Logistic equation to draw the cumulative mass loss curve, and then the tangent F1 at the maximum slope of the curve and the tangent F2 at the end of the curve are fitted respectively. Then the intersection of F1 and the axis is taken as t1, and the intersection of F1 and F2 is taken as t2. After that, the different periods of the cavitation process can be quantitatively divided, where [0, t1], [t1, t2] and [t2, +∞) represent the incubation period, acceleration period and attenuation period respectively. Define the cavitation life coefficient ζ, which is defined as Where t is the cavitation test time during the experiment, t2 is the turning point between the cavitation acceleration period and the attenuation period of the blade material measured in the experiment, and the cavitation life coefficient obtained from this varies in the range of [0,1]; when ζ=1, it is considered that the blade material has been severely damaged and needs to be repaired or replaced; when ζ<1, it can be calculated according to the formula t r =t0×(1-ζ) / ζ, calculate the remaining useful life RUL (Remaining Useful Life) of the blade material; where t0 is the time the blade material has been used in actual operation, t r is the RUL of the blade material. Since the cavitation image changes most dramatically during the acceleration period, as many pictures as possible should be collected during the acceleration period, and the collection time interval should be shorter than that of the other two time periods in order to improve the subsequent image recognition accuracy.

[0042] After the experiment is completed, a series of weight loss data and a data set of cavitation morphology images at each time t can be obtained. The cavitation life coefficient ζ at each time t can be determined through the turning point, and the life coefficient ζ is used as the label of the data set.

[0043] (2) The collected original cavitation image dataset is randomly divided into a training set and a validation set in a ratio of 8:2, and each image is randomly cropped 20 times to obtain an image dataset with a pixel size of 448×448. It is then input into the constructed SCConv-LiteMLA-VIT network model for training. Each time t corresponds to a cavitation life coefficient ζ label.

[0044] Among them, SCConv-LiteMLA-VIT network model: The size of the input original image A is 3×224×224. After the channel increase module, the number of channels of the input image is increased from 3 to 16, and the output size becomes 224×224×16. Then, the image passes through the spatial channel convolution module to enhance its spatial and channel information, and the output size remains at 224×224×16. Next, the image enters the multi-scale linear attention module, which applies a lightweight multi-scale linear attention mechanism to capture long-distance dependencies, and the output size is 224×224×3. Before the image generates tiles, the data passes through the tile embedding module to divide the input image into tiles of a fixed size (can be specified), each of which is 16×16 in size. The image is divided into 14×14 tiles in total, each tile is expanded into a one-dimensional vector, and its dimension is increased to 192 through linear transformation. In addition, positional embeddings are added to the tiles to form a tile sequence containing position information. The final input information is 1×198×192, where 198 includes 196 tiles, 1 classification tag, and 1 distillation tag (after the image is divided into tiles, the classification tag indicates the original data name of these tiles; the distillation tag indicates the teacher model or student model, which is used to distinguish which are the teacher models and which are the student models during the distillation process. The student model will imitate the output or behavior of the teacher model to achieve knowledge transfer and model compression). The tile sequence passes through 12 Transformer encoder layers, each of which contains multi-head attention, normalization, and multi-layer perceptron. The multi-head attention design enables the model to focus on multiple parts of the image in parallel, thereby effectively capturing complex contextual dependencies; normalization stabilizes the training process through data normalization; the multi-layer perceptron processes features from the attention mechanism, increases nonlinearity, and helps recognize complex patterns. The output information of each encoder layer is 1×198(196+1+1)×192. After 12 encoder layers, CLS (class labeling, classification of data annotation, abbreviation of classification; for example, a series of pictures are annotated, the pictures in this folder represent name B, and the pictures in another folder represent name C) plays an important role in the entire model. It represents the global features of the entire sequence (containing rich information). These features are extracted for the final classification task. The multi-layer perceptron layer further processes these global features, and the final input information is 1×1×192.Finally, the model passes the summarized CLS (class label) to the classifier, which consists of one or more Dense Layers and a softmax activation function. The softmax activation function converts the CLS (class label) output into a probability distribution, that is, the category with the highest score calculated by the softmax function is the category that the model predicts the image is most likely to belong to. The entire calculation process is implemented through six main steps: data preprocessing, feature extraction and enhancement, image generation blocks and embedding, feature encoding and conversion, feature aggregation and classification, and result output. The model architecture combines the strengths of convolutional neural networks CNN and Transformer. Through the combination of multi-scale learning and attention mechanisms, it enhances the ability to extract and recognize complex features. This model performs particularly well in high-precision image classification tasks.

[0045] The above SCConv-LiteMLA-VIT network model is improved based on VIT, SCConv and LiteMLA attention mechanism; an improved SCConv-LiteMLA-VIT network model is constructed, and the data set is input into the network for training and learning; the entire network is mainly constructed based on VIT, and the LiteMLA attention mechanism module and the spatial channel convolution (SCConv) module are added, which mainly enhances the model's ability to represent the features of the input data, specifically: the importance of the features can be adaptively readjusted, so that the model can better capture the key features of the input data, thereby improving the model's ability to represent features. Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the improved network model converges faster and the loss value decreases faster, and the accuracy and loss value are significantly improved and decreased compared with the other three network model structures. Therefore, the network model structure proposed in the present invention has better performance.

[0046] The improved SCConv-LiteMLA-VIT network model has higher accuracy and lower loss value. The SCConv-LiteMLA-VIT network has achieved good results on the CIFAR10 public dataset. Compared with the VIT-Distilled network, the accuracy has increased from 93.1% to 94.2%, and the loss value has dropped from 0.8131 to 0.7625. Figure 4 , Figure 5As shown in the figure, the experimental results on the self-made cavitation damage dataset show that after adding LiteMLA, the accuracy of the model increased from 91.8% to 94.9%, an increase of 3.1%, and the loss decreased from 1.2812 to 1.1723, further improving the performance. After combining the LiteMLA and SCConv modules, the improved SCConv-LiteMLA-VIT network model test set accuracy reached 95.9%, and the loss was reduced to 1.0786, verifying the effectiveness of these improvements and adapting to the self-made cavitation damage dataset. Figure 8 Shown are the prediction results of some pictures in the test set. It can be seen that the prediction accuracy reaches more than 90%, which has a very high prediction accuracy.

[0047] The interpretability of deep learning models involves the ability to understand and explain the internal mechanisms and capabilities of the model's decision-making, which is crucial for the design and optimization of deep models. In the pursuit of model interpretability, researchers can start mathematical theory analysis at the model construction stage, or after the model is trained, use techniques such as class activation maps to perform interpretative analysis of the model. In order to intuitively display the changes before and after the model is improved, a visualization method of the attention mechanism (saliency map) is used to verify the effectiveness of the improved model.

[0048] LiteMLA module and SCConv module can help VIT extract high-level semantic features and achieve more accurate learning; these modules can also improve the network's extraction effect on QKV, thereby enhancing the ability of the multi-head attention mechanism. The experiment selected data from the test set to more intuitively demonstrate the improvement in classification effect. Q (Query): query vector, used to query information at other locations; K (Key): key vector, used to represent information at other locations for matching with Query; V (Value): value vector, representing the value of information at other locations, and ultimately used to generate weighted output. The main function is to recalculate the attention weights in the attention mechanism and perform weighted summation. In computer vision tasks, such as image description generation and visual question answering, QKV is also used to build cross-modal attention mechanisms.

[0049] like Fig. 9 , Fig.10As shown in the figure, before the improvement, the original image shows a pattern with obvious texture and different brightness. The mean attention map reflects the distribution focus of the attention mechanism in the image. The attention distribution is more scattered, with higher attention concentration in some areas (bright areas) and lower attention in other areas (dark areas), showing the important areas recognized by the model. After the improvement, the original image remains consistent and shows the same texture and pattern. The mean attention map shows that the improved attention distribution is more uniform, and the high-intensity areas are more obvious. The concentrated attention areas are clearer, indicating that the improved mechanism better identifies the important areas. In terms of attention distribution, the improved attention mechanism shows a more consistent and concentrated attention distribution, indicating that the model has improved performance in identifying important areas of the image. In terms of intensity and clarity, the attention map in the second figure shows more obvious bright and concentrated areas, indicating that the improved model can better capture and highlight key features.

[0050] The constructed SCConv-LiteMLA-VIT network model has the following specific training methods:

[0051] Under the ultra-depth microscope at 300 times, the image pixels taken were 1600×1200, and the most serious cavitation damage areas of the blade material were collected. Subsequently, each image in the training set and the test set was randomly cropped 20 times to a pixel size of 448×448 to increase the training data and enhance the generalization of the model. Since the pixels of the input image of the network model are 224×224 pixels, if the cropped pixels are too large, such as cropping to 960×960 pixels, a large amount of image information will overlap during multiple cropping processes, increasing the risk of overfitting, and may cause the model to fail to converge, the accuracy and loss value fluctuate too much, and reduce the generalization and robustness of the model. It should be noted that the original image must first be divided into a training set and a validation set in a ratio of 8:2, and the cavitation life coefficient ζ is used as a label before the cropping operation is performed, otherwise the model will overfit, resulting in training failure. Therefore, when making a data set, the data distribution problem is particularly important.

[0052] During the training process, the training set and the validation set are uniformly normalized (according to the given mean and standard deviation), and the 448×448 pixel images are randomly cropped and scaled to 224×224 using data augmentation. Then, they are Gaussian blurred with a blur kernel of 5×5 and randomly rotated within 15°. Mixup and CutMix operations are performed on the images to simulate cavitation images collected in real environments. The purpose of such operations is to improve the generalization ability and robustness of the model, because it allows the model to learn features from different perspectives of the same image. Gaussian blurring is to convolve the input image by setting the Gaussian kernel size and standard deviation to achieve the effect of blurring the image. Gaussian blurring can make the image smoother, reduce noise and details, and make the image clearer and easier to process. The network is trained by setting the AdamW (Adam with Weight Decay, a variant of the Adam (Adaptive Moment Estimation) optimizer), variable learning rate (cosine annealing strategy) and cross entropy loss function (Cross Entropy Loss). AdamW optimizer is a gradient descent algorithm for training neural networks. It combines the momentum algorithm and the adaptive learning rate algorithm. It achieves faster convergence and better generalization by calculating different adaptive learning rates for each parameter. Its main advantage is that it can adaptively adjust the learning rate of each parameter, thereby improving the convergence speed and generalization ability of the model. The setting of variable learning rate can effectively reduce the numerical fluctuations of the accuracy and loss value of the later model. According to the number of collected samples, the appropriate batch size and number of training rounds are set. The smaller the value of the cross entropy loss function, the better the prediction effect of the model. Its advantage is that it can improve the prediction accuracy.

[0053] (3) Life prediction of blade materials: When inspecting and maintaining blade materials, take photos of the area where cavitation occurs and load them into the trained SCConv-LiteMLA-VIT network model to obtain the cavitation life coefficient ζ test When test =1, it is considered that the blade material has been seriously damaged and needs to be repaired or replaced; test <1, then according to the formula t r =t0(1-ζ) / ζ, calculate the RUL of the blade material; where t0 is the time the blade material has been used in actual operation, t r is the RUL of the blade material.

[0054] During the latent period, the RUL of the blade material is relatively long, indicating that it still has a high use value at this stage; during the acceleration period, the RUL drops rapidly, indicating that the damage rate of the blade material is accelerated at this stage and needs to be closely monitored and evaluated. During the decay period, the RUL is close to zero, indicating that the blade material is approaching the end of its service life and needs to be replaced or repaired in time. According to the measured RUL, early warning and preventive maintenance of the equipment can be achieved.

[0055] The method for identifying the degree of cavitation of blade materials described in the present invention does not rely on expert experience and relevant knowledge, and does not require model construction and complex numerical calculation processes. It only needs to be based on the SCConv-LiteMLA-VIT network model. By collecting cavitation micromorphology images and weightlessness data to train the network model, it is possible to accurately judge the cavitation state of the blade material and ensure stable and efficient operation of the equipment.

[0056] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A blade material damage life prediction method based on SCConv-LiteMLA-VIT network, characterized by: The steps include: Step S1, collecting and obtaining damage morphology images and weight loss data of blade materials at each stage and time t; Step S2, the data set obtained in step S1 is divided into a training set and a validation set in a ratio of 8:2, and the training is performed on the constructed improved SCConv-LiteMLA-VIT network model, and the trained model is saved for the damage degree discrimination of the blade material; Step S3: Input the image of the damaged area of ​​the blade material to be identified into the trained SCConv-LiteMLA-VIT network model obtained in step S2, obtain the life coefficient ζ of the image, and complete the life prediction of the blade material based on the life coefficient ζ.

2. The blade material damage life prediction method based on the SCConv-LiteMLA-VIT network according to claim 1 is characterized in that: In step S2, the improved SCConv-LiteMLA-VIT network is as follows: SCConv-LiteMLA-VIT network model: The original input image size is 3×224×224; after the channel increase module, the number of channels of the input image is increased from 3 to 16, and the output size becomes 224×224×16; then, the image passes through the spatial channel convolution module to enhance its spatial and channel information, and the output size remains at 224×224×16; then, the image enters the multi-scale linear attention module, and the output size is 224×224×3; before the image generates tiles, the data passes through the tile embedding module to divide the input image into fixed-size tiles, each of which is 16×16 in size, and the image is divided into 14×14 tiles in total. Each tile is expanded into a one-dimensional vector, and its dimension is increased to 192 through linear transformation; The position embedding is added to the tile to form a tile sequence containing position information. The final input information is 1×198×192, where 198 includes 196 tiles, 1 classification tag and 1 distillation tag; the tile sequence passes through 12 Transformer encoder layers, each of which contains multi-head attention, normalization and multi-layer perceptron; the output information of each encoder layer is 1×198×192. After processing by 12 encoder layers, the CLS class tag represents the global features of the entire sequence in the entire model, and these features are extracted for the final classification task; the multi-layer perceptron layer further processes these global features, and the final input information is 1×1×192; finally, the model passes the summarized CLS class tag to the classifier, which consists of one or more Dense Layers and a softmax activation function. The softmax activation function converts the CLS class tag output into a probability distribution, that is, the category with the highest score calculated by the softmax function is the category that the model predicts the image is most likely to belong to.

3. The blade material damage life prediction method based on the SCConv-LiteMLA-VIT network according to claim 2 is characterized in that: In step S2, the constructed SCConv-LiteMLA-VIT network model is trained as follows: Under the ultra-depth microscope at 300 times magnification, the image pixels were 1600×1200, and the areas with the most severe damage to the leaf material were collected; then, each image in the training set and the test set was randomly cropped 20 times to a pixel size of 448×448; During the training process, the training set and the validation set are normalized uniformly, and the 448×448 pixel images are randomly cropped and scaled to 224×224 using data augmentation. They are then Gaussian blurred with a blur kernel of 5×5 and randomly rotated within 15°. Mixup and CutMix operations are performed on the images to simulate images collected in a real environment. The network is trained by setting the AdamW optimizer, variable learning rate, and cross entropy loss function.

4. The blade material damage life prediction method based on SCConv-LiteMLA-VIT network according to claim 1 is characterized in that: In step S1, the damage morphology of the blade material is photographed by an ultra-depth-of-field microscope at 300 times magnification.

5. The blade material damage life prediction method based on SCConv-LiteMLA-VIT network according to claim 1 is characterized in that: In step S1, the collected weight loss data is fitted, a cavitation characteristic curve is drawn, and a turning point t2 is determined; through the turning point t2, the cavitation life coefficient ζ corresponding to each time t is determined, and a data set with the cavitation life coefficient ζ as a label is obtained.

6. The blade material damage life prediction method based on SCConv-LiteMLA-VIT network according to claim 5 is characterized in that: The turning point is determined by drawing on the cavitation characteristic curve; the cumulative mass loss discrete points of the blade material during the experiment are fitted by the Logistic equation to draw the cumulative mass loss curve, and then the tangent F1 at the maximum slope of the curve and the tangent F2 at the end of the curve are fitted respectively; then the intersection of F1 and the axis is taken as t1, and the intersection of F1 and F2 is taken as t2; then the different periods of the cavitation process can be quantitatively divided, where [0, t1], [t1, t2] and [t2, +∞) represent the latent period, acceleration period and attenuation period respectively; the cavitation life coefficient ζ is defined, which is defined as Where t is the cavitation test time during the experiment, t2 is the turning point between the cavitation acceleration period and the attenuation period of the blade material measured in the experiment, and the cavitation life coefficient obtained from this varies in the range of [0,1]; when ζ=1, it is considered that the blade material has been severely damaged and needs to be repaired and replaced; when ζ<1, according to the formula t r =t0×(1-ζ) / ζ to calculate the remaining service life RUL of the blade material; where t0 is the time the blade material has been used in actual operation, t r is the RUL of the blade material.

Images