Mechanical equipment surface damage identification method based on improved U-Net

Abstract

The invention provides a mechanical equipment surface damage identification method based on improved U-Net, and the method comprises the steps: collecting a mechanical equipment surface image with damage features, carrying out the cleaning and cutting, and dividing the mechanical equipment surface image into a training set, a verification set and a test set; a training set is selected for data enhancement; inputting the enhanced training set into a mechanical equipment surface damage identification model for training; according to the mechanical equipment surface damage identification model, a color feature extraction module is introduced on the basis of a U-Net model, the color feature extraction module comprises a 1 * 1 convolution layer, a batch normalization layer and a ReLU activation function, and an encoder comprises four first feature extraction modules and a second feature extraction module which are connected in sequence. The decoder comprises four feature fusion modules which are connected in sequence, and the feature fusion modules are in one-to-one correspondence with the first feature extraction modules; verifying the improved U-Net model by using a verification set; and testing the final improved U-Net model by using the test set, and counting evaluation indexes.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment detection, and in particular to a mechanical equipment surface damage identification method based on improved U-Net. Background Art

[0002] In a complex industrial production environment, mechanical equipment will inevitably be affected by various external factors, resulting in complex damage such as cracks and corrosion on its surface. If these damages are ignored, they may gradually deteriorate and eventually cause damage to the internal structure and even cause equipment failure, seriously affecting production efficiency and safety. Therefore, timely and accurate identification and treatment of surface damage of mechanical equipment is crucial to maintaining the stability and safety of industrial production.

[0003] With the widespread adoption of machine learning in various fields, methods for using machine learning to identify surface damage of mechanical equipment have emerged and have achieved remarkable results. However, the existing machine learning models have limited training data and high requirements for technical personnel in model selection and improvement, resulting in uneven recognition effects.

[0004] Therefore, how to further improve the accuracy of surface damage identification of mechanical equipment is crucial to maintaining the production efficiency and safety of mechanical equipment. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method for identifying surface damage of mechanical equipment based on an improved U-Net, comprising: Step 1: Collect various mechanical equipment surface images with damage characteristics, clean and crop them, and divide the mechanical equipment surface images into a training set, a validation set, and a test set according to a preset ratio; Step 2: Select some data in the training set for data enhancement to obtain an enhanced training set; Step 3, input the enhanced training set into the mechanical equipment surface damage recognition model for training; the mechanical equipment surface damage recognition model adopts an improved U-Net model, including an encoder and a decoder, the encoder includes four first feature extraction modules and one second feature extraction module connected in sequence, the first feature extraction module includes 2 3×3 convolutional layers, 2 color feature extraction modules and a 2×2 maximum pooling layer, the two color feature extraction modules of each of the first feature extraction modules are connected in series and connected in parallel with the first 3×3 convolutional layer of the first feature extraction module, the features output by the two are summed in the channel dimension and then pass through the second 3×3 convolutional layer and the 2×2 maximum pooling layer of the first feature extraction module in sequence, the second feature extraction module includes 2 3×3 convolutional layers and 2 color feature extraction modules, the two color feature extraction modules of the second feature extraction module are connected in series and connected in parallel with the first 3×3 convolutional layer of the second feature extraction module, the features output by the two are summed in the channel dimension and then input into the second feature extraction module The decoder comprises four feature fusion modules connected in sequence, the feature fusion modules correspond to the first feature extraction module one by one, the feature fusion module comprises a 2×2 upsampling layer and two 3×3 convolutional layers, the features output by the second 3×3 convolutional layer of the second feature extraction module enter the 2×2 upsampling layer of the first feature fusion module, the 2×2 upsampling layer of the feature fusion module is connected in series with the first 3×3 convolutional layer of the feature fusion module, the 2×2 maximum pooling layer of the first feature extraction module is jump-connected to the corresponding 2×2 upsampling layer of the feature fusion module, the color features extracted by the color feature extraction module of the first feature extraction module and the output features of the first 3×3 convolutional layer of the feature fusion module are summed in the channel dimension and then input into the second 3×3 convolutional layer of the feature fusion module, and the output of the second 3×3 convolutional layer of the last feature fusion module passes through a 1×1 convolutional layer to obtain the recognition result; Step 4: verify the improved U-Net model trained in step 3 using the verification set, adjust and retrain the model parameters, and determine the final mechanical equipment surface damage recognition model; Step 5: Use the test set to test the final mechanical equipment surface damage identification model and calculate statistical evaluation indicators.

[0006] Optionally, the step 1 includes: After collecting various surface images of mechanical equipment with damage characteristics, the fuzzy detection algorithm is used for data cleaning to eliminate images with poor imaging quality. The images are cropped and divided into 224×224 non-overlapping small images, and the surface images of mechanical equipment are divided into training set, verification set and test set according to the preset ratio.

[0007] Optionally, step 2 includes: Select a certain proportion of data in the training set and perform rotation, flipping, occlusion, shifting and cropping to achieve data expansion; A certain proportion of data in the training set is selected, and the k-means clustering model is used to cluster pixels with similar colors in the image. Then, each category in the pixel clustering result is traversed, and a color-based data enhancement method is used for each category to convert the color of the pixels contained in it.

[0008] Optionally, the color feature extraction module includes a 1×1 convolution layer, a batch normalization layer and a ReLU activation function.

[0009] Optionally, step 5 includes: Using the test set to test the final mechanical equipment surface damage recognition model to obtain a mechanical equipment surface damage recognition result; According to the results of mechanical equipment surface damage identification, the four evaluation indicators of precision, accuracy, recall rate and intersection-over-intersection ratio are calculated and counted. The specific calculation formula is as follows: (1) (2) (3) (4) in, TP , FP , TN , FN They are the number of correct samples in the damaged area predicted by the model, the number of incorrect samples in the damaged area predicted by the model, the number of correct samples in the background area predicted by the model, and the number of incorrect samples in the background area predicted by the model. A and B They are the damage area predicted by the model and the actual damage area, respectively.

[0010] After adopting the above technical solution, the present invention has at least the following beneficial effects: (1) The present invention proposes a data enhancement method based on the k-means clustering algorithm. The k-means clustering algorithm is used to cluster pixels with similar colors in the mechanical equipment surface damage image, and a color-based data enhancement method is used to convert the colors of the pixels contained in each clustered category, so as to achieve the coupling of color features and texture features of the mechanical equipment surface damage image. Compared with the traditional data enhancement method, this method can take into account the color features and texture features of the damage area, which can not only partially alleviate the problem of underfitting of model training caused by less data, but also enable the mechanical equipment surface damage recognition model to better learn multiple features of the damage area; (2) The present invention proposes a new color feature extraction module that can realize the refined extraction of color features of mechanical equipment surface damage images. The color feature extraction module ensures that the model retains the complex damage edge features in the image as much as possible while extracting color features through the dimension increase-dimension reduction operation of the 1×1 convolution kernel, avoiding the problem of edge feature loss caused by the use of convolution kernels of other sizes. In addition, the composition of the color feature extraction module is relatively simple, and the number of parameters required by the module is relatively small. It can effectively shorten the calculation time of the model while ensuring the recognition accuracy of the model, simplify the complexity of the model, and improve the interpretability of the model to a certain extent; (3) The present invention proposes an improved U-Net mechanical equipment surface damage recognition model, which realizes the fusion of color features and texture features through a feature fusion mechanism. By summing the color features and texture features in the channel dimension in the encoder part, the coupling degree of texture features and color features is improved. The color features in the encoder are passed to the feature reconstruction layer of the corresponding level in the decoder through the jump connection strategy of U-Net, thereby realizing the feature fusion of the decoder part and better retaining and utilizing the color features. The use of this feature fusion mechanism can ensure that the model still has a high recognition accuracy when facing various types of equipment surface damage, which is conducive to the implementation of the model in the application of mechanical equipment surface damage inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] 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, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0012] Figure 1 This is a schematic diagram of the mechanical equipment surface damage data set; Figure 2 Schematic diagram of color space conversion and color channel combination results of mechanical equipment surface damage image, where a is the input image of mechanical equipment surface damage image, b is the schematic diagram of color space conversion results of mechanical equipment surface damage image, and c is the schematic diagram of color channel combination results of mechanical equipment surface damage image; Figure 3 This is a schematic diagram of the data enhancement results of the surface damage image of mechanical equipment based on the k-means clustering algorithm; Figure 4 This is a schematic diagram of the color feature extraction module structure; Figure 5 This is a schematic diagram of the structure of the mechanical equipment surface damage identification model; Figure 6A graph showing the learning rate adjustment based on cosine annealing. Figure 7 Output result diagram of the mechanical equipment surface damage identification model. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] Taking a certain manufacturer's mechanical equipment in operation as the experimental object, the steps of the mechanical equipment surface damage identification method are as follows: Step 1: Collect various mechanical equipment surface images with damage characteristics, clean and crop them, and divide the mechanical equipment surface images into a training set, a validation set, and a test set according to a preset ratio; After collecting various mechanical equipment surface images with damage characteristics on a certain marine platform in service, the fuzzy detection algorithm is used for data cleaning to remove images with poor imaging quality. In order to balance the model training speed and computer hardware performance, the images are cropped and divided into 224×224 non-overlapping small images. In view of the fact that there may be multiple types of damage on the surface of mechanical equipment at the same time, and the various damage characteristics vary greatly, the surface images of mechanical equipment with representative and significant damage characteristics are selected to form the mechanical equipment surface damage dataset, such as Figure 1 As shown, the mechanical equipment surface damage dataset is divided into training set, validation set and test set in a ratio of 8:1:1; Step 2: Select part of the data in the training set for data expansion and data enhancement to obtain an enhanced training set; In the mechanical equipment production environment, the damage identification task must be able to handle the diversity of damage in different positions, colors, backgrounds, quantities and types. However, in the currently available data sets of mechanical equipment surface damage, the damage types are often relatively single, and the data volume is generally small, so it cannot be used directly. In addition, after actual investigation, it was found that there are difficulties in collecting surface damage images of mechanical equipment. Some damage exists in the blind spot of the equipment's field of view, which is difficult to collect manually, resulting in the inability to obtain sufficient sample images. Due to the coupling of multiple damages in the collected damage images, the outline of a single category of damage is difficult to define, resulting in low annotation efficiency and poor annotation consistency for annotation personnel, further exacerbating the lack of data volume. The lack of training data sets can easily lead to insufficient model training, thereby affecting the normal reasoning ability of the model.

[0015] As a method to standardize the training process and reduce generalization errors, data augmentation methods are widely used in situations where data sets are scarce. The main purpose of this method is to expand the number of images in the data set and adopt more diverse and in-depth technical means to promote the neural network model to more effectively capture the essential characteristics of the image, thereby preventing the model from overfitting or falling into the local optimal solution problem.

[0016] Data augmentation methods create a large number of new samples with similar distribution to the original image by applying geometric and color adjustments to the original image to make up for the lack of sample quantity. In addition, data augmentation methods can also increase the diversity of image samples and improve the ability of neural network models to recognize features. For example, by rotating, mirroring and other transformations on the image, it can help the neural network learn features that are not affected by direction.

[0017] As the most intuitive visual feature in existing color images, color plays a vital role in the target feature learning of the model. In traditional data enhancement methods, the color feature information of the original image is disturbed by converting the color space, reorganizing the color channels, and adjusting the brightness, contrast, and saturation of the image to generate a large number of new samples with different color features. Figure 2 The data enhancement effect of color space conversion and color channel combination on the surface damage image of mechanical equipment is demonstrated. When the input image is an RGB color image, the color space conversion is achieved by converting the image from the RGB color space to the HSV color space. HSV represents hue, saturation and brightness, which is significantly different from the red, green and blue color space represented by RGB. In addition, by disrupting the order of the three RGB channels and changing them to read in GBR order, the color channel combination is changed, which changes the global color characteristics of the image. However, from Figure 2 From the changes between (b) and (c), it can be seen that the color space transformation further blurs the outline of the damaged area and does not reflect the texture characteristics of the damage well. The color channel combination operation only changes the global color and does not enhance the texture and color characteristics of the damaged area. Therefore, the traditional color-based data enhancement method is not suitable for mechanical equipment surface damage images.

[0018] It can be seen that only by coupling the color features and texture features can the features of the equipment surface damage area be better highlighted, which is conducive to the model's secondary learning of the same equipment surface damage image. Therefore, a data enhancement method based on the k-means clustering algorithm is proposed. The k-means algorithm is a distance-based clustering algorithm that aims to divide a set of data into several categories, where the internal data of each category is as similar as possible, and the data of different categories are as different as possible. The core idea of ​​the k-means algorithm is to distribute the data to k categories through an iterative process so that the sum of the distances between the data of each category and the center points of other categories is minimized. In image processing, the k-means algorithm clusters pixels according to the color features of the image to achieve the reduction of categories in the image. In this embodiment, the k-means clustering algorithm is used to capture the color features of the image with keen ability, and color space conversion or color channel combination is performed on each category after clustering. Color-based data enhancement operations such as color channel combination are performed to maintain the correlation of texture features between categories and the consistency of color features within the category, thereby strengthening the connection between color features and texture features of the damaged area in the mechanical equipment surface damage image.

[0019] The k-means clustering algorithm keenly captures the color features of the damaged image and realizes the clustering of pixels with similar colors. After clustering, the image is cut into several blocks, each of which represents a color feature category. However, after clustering, the damaged area and the background area are still in a state of similar color, and the texture features are still not well reflected. Further random color data enhancement is performed on each category, such as Figure 3 As shown in the figure, this change in color features destroys the original similar color features of the damaged area and the background area, causing a huge color feature mutation at the outline of the damaged area, which makes the outline of the damaged area clearer and achieves dual enhancement of color features and texture features, which is conducive to the subsequent recognition model to achieve sufficient model training with a small number of samples.

[0020] The data in the training set are selected according to the preset ratio for rotation, flipping, occlusion, shifting, cropping and data enhancement based on the k-means algorithm, as shown in Table 1.

[0021] Table 1

[0022] Step 3: Input the enhanced training set into the mechanical equipment surface damage recognition model for training; In order to take into account both the complexity of the model and the accuracy of recognition, the U-Net semantic segmentation model was selected as the benchmark model for surface damage recognition of mechanical equipment. As a classic "encoder-decoder" architecture model, the encoder part extracts the features of the input image layer by layer through the alternating use of convolutional layers and pooling layers. In this process, the size of the image gradually decreases and the number of channels of the feature map gradually increases, thereby realizing the compression and encoding of the image information; the decoder part is responsible for restoring the feature map extracted by the encoder to a prediction result with the same size as the input image. The decoder part uses upsampling operations to increase the size of the feature map, and uses convolutional layers to refine and adjust the detail information of the feature map. In addition, the U-Net model uses a skip connection strategy to splice the feature map of the encoder when it is downsampled with the feature map of the corresponding feature reconstruction layer in the decoder, thereby realizing the recovery of position information. This design not only helps to improve the accuracy and robustness of the model, but also can effectively solve the problem of position information lost in the encoder due to downsampling operations.

[0023] Inspired by the encoder-decoder architecture and skip connection strategy in U-Net, and considering that the surface damage of mechanical equipment has significant texture features and color features, the mechanical equipment surface damage recognition model introduces a color feature extraction module based on the U-Net model, such as Figure 4 As shown in the figure, the color feature extraction module includes a 1×1 convolution layer, a batch normalization layer and a ReLU activation function, wherein the 1×1 convolution kernel is the key to guiding the model to realize color feature extraction. Different from convolution kernels of other sizes, the 1×1 convolution kernel only changes the number of channels of the feature map without changing the size of the feature map. Considering that in color images, color features are reflected in the number of channels of the image, the model can focus on the refined extraction of color features through multiple dimensionality increase and dimensionality reduction operations based on the 1×1 convolution kernel. The batch normalization layer normalizes a set of input feature data to restore the distribution space required for color features. Finally, the ReLU activation function introduces nonlinear transformation to enhance the nonlinear fitting ability of the color feature extraction module.

[0024] After designing the color feature extraction module, the module is integrated into the benchmark U-Net model to achieve the fusion of color features and texture features. The architecture of the mechanical equipment surface damage recognition model is as follows: Figure 5As shown, the encoder includes four first feature extraction modules and one second feature extraction module connected in sequence. The first feature extraction module includes two 3×3 convolutional layers, two color feature extraction modules and a 2×2 maximum pooling layer. The two color feature extraction modules of each first feature extraction module are connected in series and in parallel with the first 3×3 convolutional layer of the first feature extraction module. The former is responsible for extracting color features, and the latter is responsible for extracting texture features. The features output by the two are summed in the channel dimension to achieve the fusion of the two, which can alleviate the gradient explosion and gradient diffusion problems to a certain extent, so that the fusion of color features and texture features is achieved. The reason for not using feature splicing is to avoid the feature map size in the model after splicing being too large, which will have an adverse effect on model training. At the same time, feature splicing will significantly increase the number of model parameters. The fused features pass through the second 3×3 convolution layer and 2×2 maximum pooling layer of the first feature extraction module in turn. The second feature extraction module includes 2 3×3 convolution layers and 2 color feature extraction modules. The 2 color feature extraction modules of the second feature extraction module are connected in series and in parallel with the first 3×3 convolution layer of the second feature extraction module. The features output by the two are summed in the channel dimension and input into the second feature extraction module. The decoder includes four feature fusion modules connected in sequence. The feature fusion modules correspond to the first feature extraction module one by one. The feature fusion module includes a 2×2 upsampling layer and two 3×3 convolutional layers. The features output by the second 3×3 convolutional layer of the second feature extraction module enter the 2×2 upsampling layer of the first feature fusion module. The 2×2 upsampling layer of the feature fusion module is connected in series with the first 3×3 convolutional layer of the feature fusion module. The 2×2 maximum pooling layer of the first feature extraction module is jump-connected to the 2×2 upsampling layer of the corresponding feature fusion module. The first feature extraction module The color features extracted by the color feature extraction module and the output features of the first 3×3 convolutional layer of the feature fusion module are summed in the channel dimension to achieve the fusion of color features and texture features in the decoder stage, improve the mechanical equipment surface damage recognition model in the decoder stage, better retain and utilize color features, and thus improve the accuracy of mechanical equipment surface damage recognition. The fused features are input into the second 3×3 convolutional layer of the feature fusion module. The output of the second 3×3 convolutional layer of the last feature fusion module passes through a 1×1 convolutional layer to obtain the recognition result. The detailed structural parameters of the model are shown in Table 2. Table 2 (a) The first feature extraction module 1

[0025] (b) First feature extraction module 2

[0026] (c) First feature extraction module 3

[0027] (d) First feature extraction module 4

[0028] (e) Second feature extraction module

[0029] (f) Feature fusion module 1

[0030] (g) Feature fusion module 2

[0031] (h) Feature fusion module 3

[0032] (i) Feature fusion module 4

[0033] The hyperparameters of the mechanical equipment surface damage identification model are set as follows: the learning rate is 0.01, the batch size is 10, the number of iterations is 5000, the optimizer uses NAdam, and the learning rate adjustment strategy uses the cosine annealing adjustment strategy. The maximum learning rate is 0.1, the minimum learning rate is 0.0001, the learning rate adjustment cycle is 200, and the learning rate adjustment curve is as follows: Figure 6 shown.

[0034] Step 4: Use the validation set to validate the improved U-Net model trained in step 3, adjust and retrain the model parameters, and determine the final mechanical equipment surface damage recognition model; Step 5: Use the test set to test the final mechanical equipment surface damage identification model and calculate the evaluation indicators.

[0035] The final mechanical equipment surface damage recognition model is tested with the test set to obtain the mechanical equipment surface damage recognition results, such as Figure 7 As shown in the figure, it can be seen that the recognition results cover diverse features such as damage location, damage edge, and damage color. The recognition results show that the mechanical equipment surface damage recognition model can accurately identify the damaged area in the image; In order to quantitatively analyze the accuracy of the model in identifying surface damage of mechanical equipment, according to the confusion matrix of mechanical equipment surface damage identification, as shown in Table 3, the four evaluation indicators of precision, accuracy, recall rate and intersection-over-union ratio are calculated and statistically analyzed. The specific calculation formula is as follows: Table 3

[0036] (1) (2) (3) (4) in, TP , FP , TN , FN They are the number of correct samples in the damaged area predicted by the model, the number of incorrect samples in the damaged area predicted by the model, the number of correct samples in the background area predicted by the model, and the number of incorrect samples in the background area predicted by the model. A and B They are the model predicted damage area and the real damage area respectively; The quantitative test results of the mechanical equipment surface damage recognition model are shown in Table 4. It can be seen from the data in the table that the model shows good performance. The key indicator of the average accuracy rate is 0.9710, and the maximum accuracy rate is 0.9997, which means that the model can effectively identify the damaged area in the mechanical equipment surface damage image, which strongly proves that the model can accurately divide the background area and the damaged area. In addition, the minimum value of the accuracy of the model is 0.8971, which means that the model performance will fluctuate when facing some small number of composite damage situations. Combined with the standard deviation of the model in accuracy of 2.9842E-02, it means that the model has good robustness and can effectively handle the diversity of mechanical equipment surface damage images.

[0037] Table 4

[0038] Embodiments of the present invention So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. A method for identifying surface damage of mechanical equipment based on improved U-Net, characterized in that: include: Step 1: collect various mechanical equipment surface images with damage characteristics, clean and crop them, and divide the mechanical equipment surface image data set into a training set, a validation set, and a test set according to a preset ratio; Step 2: Select some data in the training set for data enhancement to obtain an enhanced training set; Step 3, input the enhanced training set into the mechanical equipment surface damage recognition model for training; the mechanical equipment surface damage recognition model adopts an improved U-Net model, including an encoder and a decoder, the encoder includes four first feature extraction modules and one second feature extraction module connected in sequence, the first feature extraction module includes 2 3×3 convolutional layers, 2 color feature extraction modules and a 2×2 maximum pooling layer, the two color feature extraction modules of each of the first feature extraction modules are connected in series and connected in parallel with the first 3×3 convolutional layer of the first feature extraction module, the features output by the two are summed in the channel dimension and then pass through the second 3×3 convolutional layer and the 2×2 maximum pooling layer of the first feature extraction module in sequence, the second feature extraction module includes 2 3×3 convolutional layers and 2 color feature extraction modules, the two color feature extraction modules of the second feature extraction module are connected in series and connected in parallel with the first 3×3 convolutional layer of the second feature extraction module, the features output by the two are summed in the channel dimension and then input into the second feature extraction module The decoder comprises four feature fusion modules connected in sequence, the feature fusion modules correspond to the first feature extraction module one by one, the feature fusion module comprises a 2×2 upsampling layer and two 3×3 convolutional layers, the features output by the second 3×3 convolutional layer of the second feature extraction module enter the 2×2 upsampling layer of the first feature fusion module, the 2×2 upsampling layer of the feature fusion module is connected in series with the first 3×3 convolutional layer of the feature fusion module, the 2×2 maximum pooling layer of the first feature extraction module is jump-connected to the corresponding 2×2 upsampling layer of the feature fusion module, the color features extracted by the color feature extraction module of the first feature extraction module and the output features of the first 3×3 convolutional layer of the feature fusion module are summed in the channel dimension and then input into the second 3×3 convolutional layer of the feature fusion module, and the output of the second 3×3 convolutional layer of the last feature fusion module passes through a 1×1 convolutional layer to obtain the recognition result; Step 4: verify the improved U-Net model trained in step 3 using the verification set, adjust and retrain the model parameters, and determine the final mechanical equipment surface damage recognition model; Step 5: Use the test set to test the final mechanical equipment surface damage identification model and calculate statistical evaluation indicators.

2. The mechanical equipment surface damage identification method based on improved U-Net according to claim 1 is characterized in that: The step 1 comprises: After collecting various surface images of mechanical equipment with damage characteristics, the fuzzy detection algorithm is used for data cleaning to eliminate images with poor imaging quality. The images are cropped and divided into 224×224 non-overlapping small images, and the surface images of mechanical equipment are divided into training set, verification set and test set according to the preset ratio.

3. The mechanical equipment surface damage identification method based on improved U-Net according to claim 1 is characterized in that: The step 2 comprises: Select a certain proportion of data in the training set and perform rotation, flipping, occlusion, shifting and cropping to achieve data expansion; A certain proportion of data in the training set is selected, and the k-means clustering model is used to cluster the pixels with similar colors in the image. Then, each category in the pixel clustering result is traversed, and a color-based data enhancement method is used to convert the colors of the pixels contained in each category, thereby realizing data enhancement based on the coupling of color features and texture features of the damaged image.

4. The mechanical equipment surface damage identification method based on improved U-Net according to claim 1 is characterized in that: The color feature extraction module includes a 1×1 convolution layer, a batch normalization layer and a ReLU activation function.

5. The mechanical equipment surface damage identification method based on improved U-Net according to claim 1 is characterized in that: The step 5 comprises: Using the test set to test the final mechanical equipment surface damage recognition model to obtain a mechanical equipment surface damage recognition result; According to the results of mechanical equipment surface damage identification, the four evaluation indicators of precision, accuracy, recall rate and intersection-over-intersection ratio are calculated and counted. The specific calculation formula is as follows: （1） （2） （3） （4） in, TP , FP , TN , FN They are the number of correct samples in the damaged area predicted by the model, the number of incorrect samples in the damaged area predicted by the model, the number of correct samples in the background area predicted by the model, and the number of incorrect samples in the background area predicted by the model. A and B They are the damage area predicted by the model and the actual damage area, respectively.

Images