Computer vision diagnosis method for fan blade surface faults

Abstract

The invention provides a computer vision diagnosis method for fan blade surface faults, and relates to the technical field of vision diagnos.The computer vision diagnosis method comprises the steps that firstly, an unmanned aerial vehicle carrying a high-definition camera is controlled to shoot the faults occurring on the fan blade surface, and fan blade fault pictures are obtained; performing image enhancement processing on the fan blade fault picture to form a fan blade fault data set in a VOC format; secondly, a Faster Net module formed by a local convolution PConv operator is integrated into a CSPDarkne-53 backbone network, and an OfficiCIoU loss function is added into a head network, so that the construction of an EV-YOLOv8 model is completed; a neural architecture search method is adopted to optimize training hyper-parameters of the EV-YOLOv8 model, the training hyper-parameters are trained through a fan blade fault data set, and an optimal parameter model is obtained; and finally, calling the optimal parameter model to carry out fan blade fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of visual diagnosis, and particularly to a computer vision diagnosis method for faults on the surface of fan blades. Background Art

[0002] With the development of computer vision technology, deep learning has become a promising solution. This technology is good at processing large datasets, enhancing detection efficiency and accuracy through deep learning algorithms. It can identify complex defect patterns without physical contact, reducing potential damage and safety risks. In addition, computer vision technology based on deep learning has characteristics of scalability, adaptability and automation, promoting continuous optimization and requiring less manual intervention. Therefore, it represents a comprehensive solution for efficient, accurate and safe fault detection of wind turbine blades. The emergence of region-aware convolutional neural network models and the innovation of algorithms such as the You Only Look Once (YOLO) series have brought revolutionary changes in object detection and recognition in deep learning.

[0003] In the existing document "Wei Gang, Ren Wei, Zhang Tao, etc. Research on Fault Monitoring and Diagnosis Method of Wind Turbine Blades Based on Empirical Mode Decomposition [J]. Electrical Technology and Economy, 2023(08): 28-31+37.", the method of homogeneous mean interpolation is used for the sampled signals of wind turbine blades to process missing values and outliers of the sampled signals; the method of decomposing the sampled signals by ensemble empirical mode decomposition is used to eliminate the influence of noise in the operating state signals of wind turbine blades under complex working conditions, the principal component analysis method is used to extract the state monitoring signals of wind turbine blades in multi-dimensional scales, and a convolutional neural network is used to identify fault features of the extracted blade operating state signals to monitor the blade operation in real time.

[0004] In the existing document "Wei Gang, Ren Wei, Zhang Tao, etc. Research on Fault Monitoring and Diagnosis Method of Wind Turbine Blades Based on Empirical Mode Decomposition [J]. Electrical Technology and Economy, 2023(08): 28-31+37.", fault identification and diagnosis are carried out by analyzing the characteristics of the unit operation monitoring signals in the fault state of the blade. The sensors installed on the wind turbine blades for collecting signals will affect the overall structure of the wind turbine blades, thus reducing the wind energy capture efficiency; when extracting signal features, the method of EEDM+PCA is adopted. The EEDM+PCA method needs to perform multiple steps such as event detection, matching and principal component analysis in the feature extraction process, and the computational complexity is relatively high. Especially when dealing with large-scale datasets, it may lead to large consumption of computing resources and long running time; this method uses a convolutional neural network model to identify the fault types of wind turbine blades, and the convolutional neural network model has problems of relatively complex hyperparameter adjustment and excessive requirements for computing resources. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a computer vision diagnosis method for faults on the surface of wind turbine blades;

[0006] A computer vision diagnosis method for faults on the surface of wind turbine blades specifically includes the following steps:

[0007] Step 1: Preparation of the wind turbine blade fault dataset;

[0008] Step 1.1: Shoot the faults on the surface of the wind turbine blade by controlling a drone equipped with a high-definition camera to obtain wind turbine blade fault pictures;

[0009] Step 1.2: Perform image enhancement processing on the wind turbine blade fault pictures;

[0010] The image enhancement processing includes: contrast enhancement, brightness enhancement, and motion blur processing;

[0011] Step 1.3: Manually label the fault categories of the wind turbine blade fault pictures before and after the image enhancement processing to form a VOC format wind turbine blade fault dataset;

[0012] Step 1.4: After converting the VOC format wind turbine blade fault dataset into the YOLO format, divide it into a training set, a validation set, and a test set according to a set ratio;

[0013] Step 2: Integrate the FasterNet module composed of local convolutional PConv operators into the CSPDarknet-53 backbone network;

[0014] Step 2.1: Define the PConv class; in the PConv class, two methods, forward_slicing and forward_split_cat, are defined to represent different forward propagation methods respectively;

[0015] Step 2.2: Define the FasterNet Block class based on the PConv module, specifically: Define the FasterNet Block class according to the PConv module and represent the FasterNet Block module.

[0016] Step 2.3: Implement the forward propagation mechanism in the FasterNet Block module by defining the forward propagation method; specifically: perform a convolution operation on the input, and add a residual connection, and then add the output after convolution to the original input to obtain the convolution result;

[0017] Step 2.4: Construct the FasterNet structure based on the FasterNet Block module; the FasterNet structure is a module in the neural network and is composed of several FasterNet Block modules;

[0018] Step 2.5: For the constructor of the Faster Net structure, initialize the Conv instances and assign them to cv1, cv2, and cv3, where cv1 is the PConv layer of the FasterNet Block module, and cv2 and cv3 are the two Conv layers of the FasterNet Block module; in addition, initialize a sequence containing several FasterNet Block modules using nn.Sequential; at the same time, define the forward method in the Faster Net class to implement the forward propagation process of FasterNet; first pass the input X through the Conv operation to get two results, then pass these two results into the sequence of FasterNet Block modules, and concatenate the outputs of these modules with the original input, and finally process through the convolution operation to get the final result;

[0019] Step 3: Add the EfficiCIoU loss function to the head network to complete the construction of the EV-YOLOv8 model;

[0020] The head network contains several detection heads, and each detection head is divided into two parts through the decoupled head structure. One part is the regression branch with the EfficiCIOU loss function added, and the other part is the classification branch that performs the classification function;

[0021] The calculation formula of the EfficiCIoU loss function is as follows:

[0022]

[0023] Among them, c represents the diagonal length of the target box; c w represents the height of the target box, that is, the vertical distance dimension; c h represents the width of the target box, that is, the horizontal distance dimension; IoU represents the intersection over union, b represents the predicted box, which consists of four elements: the x coordinate, the y coordinate, the width w of the predicted box, and the height h of the predicted box; b gt is the target box, which consists of the x coordinate, the y coordinate, the width w gt and the height h gt of the target box; ρ(b, b gt ) is the Euclidean distance between the center points of the target box and the predicted box, that is, the straight-line distance between the two center points on the plane; α represents the weight parameter; v is used to measure the similarity of the aspect ratio, and αv together represents the aspect ratio influence factor.

[0024] Step 4: Optimize the training hyperparameters of the EV-YOLOv8 model using the neural architecture search method;

[0025] Step 4.1: Evaluation strategy: Implement a population memory strategy based on similarity measurement, specifically: After the first execution of the training phase, restart the training again to evaluate each individual to determine its baseline accuracy; The architectures that have been trained and have their accuracy determined are then stored in the archive; When a new generation is generated, they are compared with the individuals in the archive according to their encoded structure data; If the similarity between the new individual and any individual in the archive exceeds the set threshold, the accuracy of the individual in the archive will be adopted as the fitness value of the new offspring; In the case where multiple archive architectures meet the similarity criteria, the fitness value will come from one of the archive architectures that exhibit the highest degree of similarity; In the case where the similarity does not reach the threshold, the accuracy of the new individual is determined through the process of retraining, and after obtaining this accuracy, the individual is added to the archive.

[0026] Step 4.2: Select the cosine similarity as the criterion for measuring similarity;

[0027] The calculation of the cosine similarity is as follows:

[0028]

[0029] A Q represents the encoding of the new generation architecture; A M represents the encoding of the architecture previously stored in the archive;

[0030] Step 4.3: Adopt a single-point crossover technique for crossover encoding, including the following steps:

[0031] Step 4.3.1: Independently select a crossover point for each parent encoding, marking the position where information exchange begins;

[0032] Step 4.3.2: Exchange the parts of the parent encodings after the crossover point to create offspring that fuse the features of the two parents;

[0033] Step 4.3.3: Adjust the encoding of the offspring to ensure compliance with the predefined architecture constraints.

[0034] Step 4.4: Implement three different mutation operations: addition, deletion, and modification;

[0035] The addition is: Randomly generate a new unit and insert it into a randomly selected position in the parent architecture;

[0036] The deletion is: Randomly select a unit to remove from the architecture;

[0037] The modification is as follows: change the details within the unit, specifically including the type, the number of units, and the number of input and output channels;

[0038] Step 5: After setting the hyperparameters of the EV-YOLOv8 model, train it with the fan blade fault dataset to obtain the optimal parameter model;

[0039] Step 6: Enter the detection instruction in the terminal, select the storage path of the fault picture to be detected, and then call the optimal parameter model to perform fan blade fault diagnosis.

[0040] The beneficial effects of adopting the above technical solutions are as follows:

[0041] The present invention provides a computer vision diagnosis method for fan blade surface faults. A new object detection network model is established based on the YOLOv8n network model, which reduces the number of parameters to be calculated during the training process of the network model, simplifies the complexity of the network model, and improves the detection accuracy of the network model. The hyperparameters are set by using the neural architecture search method, and a user interaction interface is made according to the trained optimal weight parameters, which not only saves the time and resource consumption of training the model, but also reduces the manual intervention in hyperparameter setting, realizing a fan blade fault diagnosis method with simple operation and excellent performance.

[0042] Compared with the prior art, the average detection accuracy of the fan blade fault detection method proposed by the present invention reaches 99.5%, and at the same time, the number of parameters to be calculated required for training is only 2,072,695, and the complexity of the network model is only 7.0, which greatly reduces the requirement for the computer hardware computing power to execute the fan blade fault detection method while ensuring that the detection accuracy meets the industrial requirements. In addition, since the method proposed by the present invention uses a drone equipped with a high-definition camera to collect fan blade fault images, there is no need to install additional sensors, which ensures the structural integrity of the fan blade and reduces the cost. Description of the Drawings

[0043] Figure 1 It is the fan blade fault diagnosis method in the embodiment of the present invention;

[0044] Figure 2 It is the working process and structural schematic diagram of the decoupling head in the embodiment of the present invention;

[0045] Figure 3 It is the structural diagram of the EV-YOLOv8 network model in the embodiment of the present invention. Detailed Embodiment

[0046] The following combines the drawings and embodiments to further describe in detail the specific embodiments of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0047] In this embodiment, as Figure 1 shown, the fan blade fault diagnosis method proposed by the present invention mainly includes two parts. First, a drone equipped with a high-definition camera is manually controlled to conduct microscopic inspections inside the wind farm, and during the inspection process, fault pictures are taken of the faults on the wind turbine blades, and then the fault pictures are uploaded to the cloud server to complete the collection of fan blade fault pictures. Then, the staff downloads the fan blade fault pictures from the cloud server through a computer in the control center, and the control center computer completes the object detection task.

[0048] A computer vision diagnosis method for fan blade surface faults specifically includes the following steps:

[0049] Step 1: Dataset preparation;

[0050] Step 1.1: Operate a drone equipped with a high-definition camera to take pictures of the faults on the fan blade surface to obtain fan blade fault pictures;

[0051] Step 1.2: Perform image enhancement processing on the fan blade fault pictures; In order to simulate various weather factors and fan blade motion states that may occur during daily inspections, image enhancement processing is performed on each obtained fan blade fault picture. The image enhancement processing includes: contrast enhancement, brightness enhancement, and motion blur processing;

[0052] Step 1.3: Manually label the fault categories of the fan blade fault pictures before and after image enhancement processing to form a VOC format fan blade fault dataset;

[0053] Step 1.4: After converting the VOC format fan blade fault dataset into the YOLO format, divide it into a training set, a validation set, and a test set according to a set ratio; In this embodiment, the set ratio is 7:2:1;

[0054] Step 2: Integrate the FasterNet module composed of local convolutional PConv operators into the CSPDarknet-53 backbone network; used to reduce the number of training parameters and reduce the network complexity;

[0055] In this embodiment, the improved backbone network architecture is shown in Table 1.

[0056] Table 1

[0057]

[0058] Step 2.1: Define the PConv class; in the PConv class, two methods, forward_slicing and forward_split_cat, are defined to represent different forward propagation methods respectively;

[0059] Step 2.2: Define the FasterNet Block class based on the PConv module, thus laying a structural foundation for the construction of the FasterNet Block module. Specifically: Define the FasterNet Block class according to the PConv module and represent the FasterNetBlock module.

[0060] Step 2.3: Implement the forward propagation mechanism in the FasterNet Block module by defining the forward propagation method; specifically: Perform a convolution operation on the input, and add a residual connection, then add the output after convolution to the original input to obtain the convolution result;

[0061] Step 2.4: Construct the FasterNet structure according to the FasterNet Block module; the FasterNet structure is a module in the neural network, consisting of several FasterNet Block modules; the constructor of the FasterNet structure receives parameters c1 and c2 representing the number of input and output channels, and n representing the number of FasterNet Block modules, and other parameters also represent relevant parameters.

[0062] Step 2.5: For the constructor of the Faster Net structure, initialize the Conv instance and assign it to cv1, cv2, and cv3, where cv1 is the PConv layer of the FasterNet Block module, and cv2 and cv3 are the two Conv layers of the FasterNet Block module; in addition, initialize a sequence containing several FasterNet Block modules using nn.Sequential; at the same time, define the forward method in the Faster Net class to implement the forward propagation process of FasterNet; first pass the input X through the Conv operation to obtain two results, then pass these two results into the sequence of FasterNet Block modules, and concatenate the outputs of these modules with the original input, and finally obtain the final result after convolution operation processing;

[0063] Step 3: Add the EfficiCIoU loss function to the head network to complete the construction of the EV-YOLOv8 model, as Figure 3 shown;

[0064] In this embodiment, the head network architecture based on the decoupled head structure is shown in Table 2.

[0065] Table 2

[0066]

[0067] The head network contains several detection heads. Each detection head is divided into two parts through a decoupling head structure. One part is a regression branch with the EfficiCIOU loss function added, and the other part is a classification branch that performs the classification function. In this embodiment, since the neck network outputs three feature maps, namely T1, T2, and T3, three detection heads are correspondingly required in the head network. The working process of each detection head is as follows Figure 2 shown

[0068] The EfficiCIoU loss function first uses the CIoU concept to calculate the aspect ratio of the predicted bounding box until it converges to a suitable range. Then, it gradually adjusts each edge according to the EIoU concept until convergence. The calculation formula of the EfficiCIoU loss function is as follows

[0069]

[0070] where c represents the diagonal length of the target bounding box; c w represents the height of the target bounding box, that is, the vertical distance dimension; c h represents the width of the target bounding box, that is, the horizontal distance dimension; IoU represents the intersection over union ratio, which measures the similarity between the predicted bounding box and the ground truth box by selecting positive and negative samples and is the most popular metric in bounding box regression; b represents the predicted bounding box, which consists of four elements: the x coordinate, the y coordinate, the width w of the predicted bounding box, and the height h of the predicted bounding box; b gt is the target bounding box, which consists of the x coordinate, the y coordinate, the width w gt of the target bounding box and the height h gt of the target bounding box; ρ(b, b gt ) is the Euclidean distance between the center points of the target bounding box and the predicted bounding box, that is, the straight-line distance between the two center points on the plane; α represents the weight parameter; v is used to measure the similarity of the aspect ratio, and αv together represents the aspect ratio influence factor

[0071] When applying the EfficiCIoU loss function, first, the bbox_effciou function should be newly added to the metrics.py file, then the newly added function should be imported in the loss.py file, and finally, the original CIoU=true should be changed to EfficiCIoU=true in the corresponding class BboxLoss class to complete the modification. The construction of the EV-YOLOv8 model is completed according to steps 2 - 3

[0072] Step 4: Optimize the training hyperparameters of the EV-YOLOv8 model using the neural architecture search method. (Hyperparameters are the parameters set during the training process of a machine learning model. Their values cannot be optimized through the training process and need to be manually adjusted and set before training. These hyperparameters directly affect the training process and the final performance of the model.)

[0073] Step 4.1: Evaluation strategy: To reduce the computational burden, implement a population memory strategy based on similarity measurement, specifically: After the first execution of the training phase, restart the training again to evaluate each individual to determine its baseline accuracy; The architectures that have been trained and had their accuracy determined are then stored in the archive; When a new generation is generated, they are compared with the individuals in the archive based on their encoded data; If the similarity between the new individual and any individual in the archive exceeds the set threshold, the accuracy of the individual in the archive will be adopted as the fitness value of the new offspring; In the case where multiple archive architectures meet the similarity criteria, the fitness value will come from one of the archive architectures showing the highest degree of similarity; In the case where the similarity does not reach the threshold, the accuracy of the new individual is determined through the retraining process, and after obtaining this accuracy, the individual is added to the archive.

[0074] Step 4.2: Select cosine similarity as the criterion for measuring similarity. This method simplifies the evaluation process by avoiding redundant calculations for architectures with significant similarity to architectures that have already been explored, thus optimizing the overall efficiency of the evolutionary search. The calculation of cosine similarity is as follows:

[0075]

[0076] A Q represents the encoding of the new generation architecture; A M represents the encoding of the architecture previously stored in the archive. Flexible encoding techniques are adopted so that the lengths of these encodings can vary between different architectures. Shorter encodings are extended with zero padding to make their lengths equal to the length of the longer encoding. This standardization ensures that the similarity calculation is consistent and accurate in all comparisons. This method significantly accelerates the evolutionary search by concentrating computational resources on exploring novel and unique architectures by leveraging the insights obtained from previous evaluations.

[0077] Step 4.3: Since the encoding lengths of the parent architectures and the offspring architectures may be different due to the adoption of flexible encoding techniques, traditional crossover methods cannot meet the requirements of this method. As a solution, a single-point crossover technique is adopted for crossover encoding, including the following steps:

[0078] Step 4.3.1: Independently select a crossover point for each parent encoding, marking the position where information exchange begins;

[0079] Step 4.3.2: Exchange the parts of the parental codes after the crossover point to create an offspring that combines the features of the two parents;

[0080] Step 4.3.3: Adjust the code of the offspring to ensure compliance with the predefined architecture constraints.

[0081] Step 4.4: Implement three different mutation operations: addition, deletion, and modification, each aimed at introducing changes and exploring within the architecture space.

[0082] The addition is as follows: Randomly generate a new unit (e.g., MBU) and insert it at a randomly selected position in the parental architecture; new architectural features are introduced and then adjusted to ensure that the offspring complies with the constraints;

[0083] The deletion is as follows: Randomly select a unit to be removed from the architecture; once removed, the offspring architecture is reconstructed and necessary adjustments are made to comply with the constraints, which may simplify or refine the architecture.

[0084] The modification is as follows: Change the details within the unit, specifically including the type, the number of units, and the number of input and output channels; its purpose is to fine-tune the architectural features without significantly changing its overall structure.

[0085] Step 5: After setting the hyperparameters of the EV-YOLOv8 model, train it with the fan blade fault dataset to obtain the best parameter model;

[0086] Step 6: Enter the detection instruction (mode = detect) in the terminal (the terminal area in the Pycharm software), and after selecting the storage path of the fault pictures to be detected, call the best parameter model for fan blade fault diagnosis.

[0087] The above description is only the preferred embodiments of the present disclosure and the description of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A computer vision diagnosis method for faults on the surface of a fan blade, characterized in that, It includes the following steps: Step 1: Prepare the fan blade fault dataset; Step 2: Integrate the FasterNet module composed of local convolutional PConv operators into the CSPDarknet-53 backbone network; Step 3: Add the EfficiCIoU loss function to the head network to complete the construction of the EV-YOLOv8 model; Step 4: Optimize the training hyperparameters of the EV-YOLOv8 model using the neural architecture search method; Step 5: After setting the hyperparameters of the EV-YOLOv8 model, train it with the fan blade fault dataset to obtain the optimal parameter model; Step 6: Enter the detection instruction in the terminal, select the storage path of the fault picture to be detected, and then call the optimal parameter model to diagnose the fan blade fault.

2. The computer vision diagnosis method for faults on the surface of a fan blade according to claim 1, wherein, The specific steps of Step 1 include the following steps: Step 1.1: Operate a drone equipped with a high-definition camera to take pictures of the faults on the surface of the fan blade to obtain fan blade fault pictures; Step 1.2: Perform image enhancement processing on the fan blade fault pictures; The image enhancement processing includes: contrast enhancement, brightness enhancement, and motion blur processing; Step 1.3: Manually label the fault categories of the fan blade fault pictures before and after image enhancement processing to form a VOC-format fan blade fault dataset; Step 1.4: After converting the VOC-format fan blade fault dataset into YOLO format, divide it into a training set, a validation set, and a test set according to a set ratio.

3. A computer vision diagnosis method for faults on the surface of a fan blade according to claim 1, characterized in that, The specific steps of Step 2 include the following steps: Step 2.1: Define the PConv class; in the PConv class, two methods, forward_slicing and forward_split_cat, are defined to represent different forward propagation methods; Step 2.2: Define the FasterNet Block class based on the PConv module, specifically: Define the FasterNet Block class according to the PConv module and represent the FasterNet Block module; Step 2.3: Implement the forward propagation mechanism in the FasterNet Block module by defining the forward propagation method; specifically: Perform a convolution operation on the input, add a residual connection, and then add the output after convolution to the original input to obtain the convolution result; Step 2.4: Construct the FasterNet structure according to the FasterNet Block module; the FasterNet structure is a module in the neural network and is composed of several FasterNet Block modules; Step 2.5: For the constructor of the Faster Net structure, initialize the Conv instances and assign them to cv1, cv2, and cv3, where cv1 is the PConv layer of the FasterNet Block module, and cv2 and cv3 are the two Conv layers of the FasterNet Block module; in addition, initialize a sequence containing several FasterNet Block modules using nn.Sequential; at the same time, define the forward method in the Faster Net class to implement the forward propagation process of FasterNet; first pass the input X through the Conv operation to obtain two results, then pass these two results into the sequence of FasterNet Block modules, concatenate the outputs of these modules with the original input, and finally obtain the final result through the convolution operation.

4. A computer vision diagnosis method for faults on the surface of a fan blade according to claim 1, characterized in that, The head network described in Step 3 contains several detection heads, and each detection head is divided into two parts through the decoupled head structure. One part is the regression branch with the EfficiCIOU loss function added, and the other part is the classification branch that performs the classification function. The calculation formula of the EfficiCIoU loss function is as follows: Among them, c represents the diagonal length of the target box; c w represents the height of the target box, that is, the vertical distance dimension; c h represents the width of the target box, that is, the horizontal distance dimension; IoU represents the intersection over union, b represents the predicted box, which consists of four elements: the x coordinate, the y coordinate, the width w of the predicted box, and the height h of the predicted box; b gt is the target box, which consists of the x coordinate, the y coordinate, the width w of the target box gt and the height h of the target box gt ; ρ(b, b gt ) is the Euclidean distance between the center points of the target box and the predicted box, that is, the straight-line distance between the two center points on the plane; α represents the weight parameter; v is used to measure the similarity of the aspect ratio, and αv together represents the aspect ratio influence factor.

5. A computer vision diagnosis method for faults on the surface of a fan blade according to claim 1, wherein, Step 4 specifically includes the following steps: Step 4.1: Evaluation strategy: Implement a population memory strategy based on similarity measurement. Specifically: After the first execution of the training phase, start training again to evaluate each individual to determine its baseline accuracy; the architectures that have been trained and have their accuracy determined are then stored in the archive; when a new generation is generated, they are compared with the individuals in the archive according to their encoded data; if the similarity between the new individual and any individual in the archive exceeds the set threshold, the accuracy of the individual in the archive will be adopted as the fitness value of the new offspring; in the case where multiple archive architectures meet the similarity criteria, the fitness value will come from one of the archive architectures that exhibits the highest degree of similarity; in the case where the similarity does not reach the threshold, the accuracy of the new individual is determined through the retraining process, and after obtaining this accuracy, the individual is added to the archive. Step 4.2: Select the cosine similarity as the criterion for measuring similarity. The calculation of the cosine similarity is shown as follows: A Q Represents the code of the SYNTEC architecture; A M Represents the code of the architecture previously stored in the archive; Step 4.3: Adopt a single-point crossover technique for cross-coding. Step 4.4: Implement three different mutation operations: addition, deletion, and modification.

6. A computer vision diagnosis method for faults on the surface of a fan blade according to claim 5, characterized in that Step 4.3 specifically includes the following steps: Step 4.3.1: Independently select a crossover point for each parent encoding, which marks the position where information exchange starts. Step 4.3.2: Exchange the parts of the parent encodings after the crossover point to create an offspring that combines the features of the two parents. Step 4.3.3: Adjust the encoding of the offspring to ensure compliance with the predefined architecture constraints.

7. A computer vision diagnosis method for faults on the surface of a fan blade according to claim 5, characterized in that, In Step 4.4, the addition is: randomly generate a new unit and insert it into a randomly selected position in the parent architecture. The deletion is: randomly select a unit to remove from the architecture. The modification is as follows: Change the details within the unit, specifically including the type, the number of units, and the number of input and output channels.

Images