Cylinder inner wall wear detection method based on detail-aware sampling attention module

Abstract

The present invention discloses a cylinder inner wall wear detection method based on a detail perception sampling attention module, which relates to the field of intelligent manufacturing. The bilinear interpolation sampling process adopted by the existing model has limited perception ability for subtle wear. To this end, firstly, the present invention constructs a bilinear rearranged feature map, which is composed of the original feature points corresponding to each feature point after sampling; secondly, a perception attention module composed of a fully connected network is designed to learn a sampling weight map from the bilinear rearranged feature map; thirdly, a mean residual aggregation module is designed to process the bilinear rearranged feature map using a mean filter, and the aggregation results of the bilinear rearranged feature map and the sampling weight map are superimposed as the sampling feature map. Since the mean of the bilinear rearranged feature map represents the low-frequency information of the original feature map, when the residual is aggregated, the sampling attention learning tendency for high-frequency information can be strengthened, the model's perception ability for wear details can be improved, and the detection accuracy can be improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent equipment maintenance and diagnosis in intelligent manufacturing, and in particular to a method for detecting cylinder inner wall wear based on a detail perception sampling attention module. Background Art

[0002] The engine block is one of the core components of the engine. As the supporting structure of the combustion chamber and cylinder liner, it needs to withstand harsh working conditions such as high temperature and high pressure. During long-term use, the inside of the cylinder block is prone to wear, which seriously affects the engine performance. The generation and expansion of wear is a gradual process. The initial wear is small and has little effect on the engine performance. However, as the use time increases, the wear will gradually expand and cause damage to the engine. Therefore, the detection of engine cylinder inner wall wear has important application value in engine maintenance.

[0003] However, traditional detection models usually rely on non-learning static sampling methods such as nearest neighbor interpolation or bilinear interpolation for multi-scale processing. These methods lose a lot of details during the sampling process, resulting in limited perception of subtle wear and tear, making it difficult to meet the engine's needs for precision detection. Summary of the invention

[0004] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art, and propose a cylinder inner wall wear detection method based on a detail-aware sampling attention module, replace the upsampling module of the YOLO network with the constructed detail-aware sampling attention module and construct a cylinder inner wall wear detection model, thereby significantly improving the model's perception of wear details and ultimately improving the detection accuracy.

[0005] The present invention adopts the following technical solution:

[0006] A method for detecting cylinder inner wall wear based on a detail-aware sampling attention module comprises the following steps:

[0007] Constructing a detail-aware sampling attention module, wherein the detail-aware sampling attention module constructs a bilinear rearrangement feature map based on the input feature map, learns a sampling weight map from the bilinear rearrangement feature map, processes the bilinear rearrangement feature map using a mean filter, and superimposes an aggregated result of the bilinear rearrangement feature map and the sampling weight map as a sampling feature map;

[0008] The upsampling module of the YOLO network is replaced with the detail-aware sampling attention module and a cylinder inner wall wear detection model is constructed; and the cylinder inner wall wear detection is performed using the trained cylinder inner wall wear detection model.

[0009] The detail-aware sampling attention module includes a bilinear rearrangement feature map construction part, a perceptual attention module and a mean residual aggregation module; the bilinear rearrangement feature map construction part uses a bilinear transformation to determine the sampling grid, and uses the sampling grid as an index to obtain the original feature points corresponding to each feature point after sampling, and these original feature points form a bilinear rearrangement feature map; the perceptual attention module mainly includes a fully connected layer, which uses the fully connected layer to extract, transmit and enhance the context information of each original feature point layer by layer, learn the spatial correlation of features in the bilinear rearrangement feature map, and then obtain a sampling weight map; the mean residual aggregation module uses a mean filter to process the bilinear rearrangement feature map, and then superimposes the feature map processed by the mean filter with the aggregation result of the bilinear rearrangement feature map and the sampling weight map to generate the sampling feature map.

[0010] The constructed bilinear rearrangement feature map expression is as follows:

[0011] ;

[0012] in, represents the sampling grid obtained by the bilinear transformation algorithm, h, w represent the height and width of the output feature map, 4 represents the four original feature points corresponding to each feature point after sampling, and 2 represents that G records the x and y coordinates of each original feature point; represents the input feature map, C, H, W represent the number of channels, height and width of the input feature map; It represents the feature map that indexes the input feature map F according to the sampling grid G, that is, according to the coordinate information in the sampling grid, the pixel value of the original feature point corresponding to each feature point after sampling is obtained. Its dimension is ; Reshape2d represents Rearrange to have A matrix of rows and 4 columns, resulting in a bilinear rearrangement feature map .

[0013] The steps to build the perception attention module are as follows:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] in, represents the bilinear rearrangement feature map; , and Represent three fully connected layers respectively. are the learnable parameters corresponding to the three fully connected layers; is a batch normalization layer; ReLU represents a nonlinear activation function; the perceptual attention module inputs Bilinear rearrangement feature map of Mapped to a higher dimensional space to obtain a feature map , p represents the feature map The number of columns; feature map Input to , further integrate and refine features to obtain feature maps , q represents the feature map The number of columns; feature map go through Perform dimension alignment to obtain the feature map , 4 represents the feature map Has 4 columns; finally, the feature map Input to the Softmax function for normalization, where dim=1 means that the Softmax function Normalize by column to obtain the sampling weight map ; C, h, w represent the number of channels, height and width of the feature map respectively.

[0019] The steps to construct the mean residual aggregation module are as follows:

[0020] ;

[0021] ;

[0022] ;

[0023] in, Represents the bilinear rearrangement feature map The low-frequency feature map obtained by mean filtering, Mean represents mean filtering, and the corresponding dim=3 represents mean filtering in The dimension corresponding to 4 in the figure is carried out; represents the features learned by the perceptual attention module, represents the dot multiplication operation, Represents the bilinear rearrangement of the feature map and the sampling weight map Perform dot multiplication, SUM represents the summation operation, and the corresponding dim=3 represents the summation in Reshape3d represents a tensor shape transformation operation. represents the final sampling feature map, which is Sum the results and reshape them into the following dimensions using Reshape3d C, h, and w represent the number of channels, height, and width of the feature map, respectively.

[0024] The cylinder inner wall wear detection model is connected to the target detection loss function, and the parameters of the detail perception sampling attention module are regularized. The gradient descent method is used for training, so that the detail perception sampling attention module is trained collaboratively with other modules of the YOLO network to obtain the trained cylinder inner wall wear detection model.

[0025] The loss function expression of the cylinder inner wall wear detection model is as follows:

[0026] ;

[0027] in, represents the target detection loss function, which includes the target box classification loss and the target box regression loss; represents the sampling feature graph; represents other feature maps except the input feature map F of the detail-aware sampling attention module during training, represents parameters other than the detail-aware sampling attention module parameters; Represents the learnable parameters corresponding to the three fully connected layers in the perception attention module; represents the L2 regularization term, Represents the regularization coefficient, which is used to control the strength of the regularization term. By adding an L2 regularization term to each fully connected layer, the weight amplitude of each layer is limited to avoid overfitting.

[0028] The cylinder inner wall wear detection model includes a Backbone module, a Neck module and a Head module. The Backbone module is used to extract the basic features of the image; the Neck module is used to further process the basic features extracted by the Backbone module, and is provided with multiple sampling modules; the Head module is used to perform target detection and classification on the output of the Neck module; the detail-aware sampling attention module replaces the upsampling module in the Neck module.

[0029] It can be seen from the above description of the present invention that, compared with the prior art, the present invention has the following beneficial effects:

[0030] In the present invention, a detail-aware sampling attention module is constructed, which constructs a bilinear rearrangement feature map based on the input feature map, learns a sampling weight map from the bilinear rearrangement feature map, processes the bilinear rearrangement feature map using a mean filter, and superimposes the aggregation result of the bilinear rearrangement feature map and the sampling weight map as the sampling feature map, which can significantly enhance the ability to capture refined features and more accurately perceive subtle changes in the image; the detail-aware sampling attention module replaces the upsampling module in the YOLO network to construct a cylinder inner wall wear detection model. Compared with the traditional static sampling method, the trained cylinder inner wall wear detection model can provide higher accuracy and reliability for wear detection tasks.

[0031] In the present invention, the detail-aware sampling attention module includes a bilinear rearrangement feature map construction part, a perceptual attention module and a mean residual aggregation module. The bilinear rearrangement feature map construction part uses bilinear transformation to determine the sampling grid, and uses the sampling grid as an index to obtain the original feature points corresponding to each feature point after sampling, and the bilinear rearrangement feature map is composed of these original feature points; the perceptual attention module uses a fully connected layer to extract, transmit and enhance the context information of each feature point layer by layer, learns the spatial correlation of features in the bilinear rearrangement feature map, and then obtains a sampling weight map; in conjunction with the mean residual aggregation module, the bilinear rearrangement feature map is processed using a mean filter, and the aggregation results of the bilinear rearrangement feature map and the sampling weight map are superimposed to generate a sampling feature map with more detailed prominence. This optimization strategy significantly enhances the ability to depict wear details and significantly enhances the ability to capture refined features. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a diagram of the cylinder inner wall wear detection model constructed by the present invention.

[0033] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. DETAILED DESCRIPTION

[0034] In order to more clearly understand the purpose, technical methods and advantages of the present invention, the present invention is further described below through specific implementation methods.

[0035] See also Figure 1 , a cylinder inner wall wear detection method based on detail-aware sampling attention module, comprising:

[0036] S1 constructs a detail-aware sampling attention module. The detail-aware sampling attention module constructs a bilinear rearrangement feature map based on the input feature map, learns a sampling weight map from the bilinear rearrangement feature map, processes the bilinear rearrangement feature map using a mean filter, and superimposes the aggregation result of the bilinear rearrangement feature map and the sampling weight map as the sampling feature map.

[0037] The detail-aware sampling attention module of the present invention is expanded on the basis of bilinear sampling, and a perceptual attention mechanism composed of a fully connected layer is introduced. The sampling weight map is learned from the bilinear rearrangement feature map, and a mean residual aggregation module is designed. The bilinear rearrangement feature map is processed by a mean filter, and the aggregation results of the bilinear rearrangement feature map and the sampling weight map are superimposed to generate a more expressive sampling feature map, thereby significantly enhancing the model's ability to capture complex details and local wear features, and improving the accuracy of wear detection.

[0038] For details, see Figure 1 The detail-aware sampling attention module includes a bilinear rearrangement feature map construction part, a perceptual attention module, and a mean residual aggregation module. The bilinear rearrangement feature map construction part uses a bilinear transformation to determine the sampling grid, and uses the sampling grid as an index to obtain the original feature points corresponding to each feature point after sampling. These original feature points form a bilinear rearrangement feature map. The constructed bilinear rearrangement feature map expression is as follows:

[0039] ;

[0040] in, represents the sampling grid obtained by the bilinear transformation algorithm, h, w represent the height and width of the output feature map, 4 represents the four original feature points corresponding to each feature point after sampling, and 2 represents that G records the x and y coordinates of each original feature point; represents the input feature map, C, H, W represent the number of channels, height and width of the input feature map; It represents the feature map that indexes the input feature map F according to the sampling grid G, that is, according to the coordinate information in the sampling grid, the pixel value of the original feature point corresponding to each feature point after sampling is obtained. Its dimension is ; Reshape2d represents Rearrange to have A matrix of rows and 4 columns, resulting in a bilinear rearrangement feature map .

[0041] The perceptual attention module uses the fully connected layer to extract, transfer and enhance the context information of each original feature point layer by layer, learns the spatial correlation of features in the bilinear rearrangement feature map, and then obtains the sampling weight map, which is expressed as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] in, represents a bilinear rearrangement feature map; , and Represent three fully connected layers respectively. are the learnable parameters corresponding to the three fully connected layers; is a batch normalization layer; ReLU represents a nonlinear activation function; the perceptual attention module inputs Bilinear rearrangement feature map of Mapped to a higher dimensional space to obtain a feature map , p represents the feature map The number of columns; feature map Input to , further integrate and refine features to obtain feature maps , q represents the feature map The number of columns; feature map go through Perform dimension alignment to obtain the feature map , 4 represents the feature map Has 4 columns; finally, the feature map Input to the Softmax function for normalization, where dim=1 means that the Softmax function Normalize by column to get the sampling weight map ; C, h, w represent the number of channels, height and width of the feature map respectively.

[0047] The mean residual aggregation module uses a mean filter to process the bilinear rearrangement feature map, and then superimposes the feature map processed by the mean filter with the aggregation result of the bilinear rearrangement feature map and the sampling weight map to generate a sampling feature map, as follows:

[0048] ;

[0049] ;

[0050] ;

[0051] in, Represents the bilinear rearrangement feature map The low-frequency feature map obtained by mean filtering, Mean represents mean filtering, and the corresponding dim=3 represents mean filtering in The dimension corresponding to 4 in the figure is carried out; represents the features learned by the perceptual attention module, represents the dot multiplication operation, Represents the bilinear rearrangement of the feature map and the sampling weight map Perform dot multiplication, SUM represents the summation operation, and the corresponding dim=3 represents the summation in Reshape3d is a tensor shape transformation operation that is specifically used to adjust the dimensions of three-dimensional data while ensuring that the total number of elements remains unchanged. represents the final sampling feature map, which is Sum the results and reshape them into the following dimensions using Reshape3d C, h, and w represent the number of channels, height, and width of the feature map, respectively.

[0052] S2 replaces the upsampling module of the YOLO network with the detail-aware sampling attention module DSA, that is, the detail-aware sampling attention module DSA replaces the function of the original upsampling module in the YOLO network, thereby constructing a cylinder inner wall wear detection model.

[0053] The cylinder inner wall wear detection model is connected to the target detection loss function, and the parameters of the detail-aware sampling attention module are regularized. The gradient descent method is used for training, so that the detail-aware sampling attention module is trained collaboratively with other modules of the YOLO network to obtain a trained cylinder inner wall wear detection model.

[0054] The loss function expression of the cylinder inner wall wear detection model is as follows:

[0055] ;

[0056] in, represents the target detection loss function, which includes the target box classification loss and the target box regression loss; represents other feature maps except the input feature map F of the detail-aware sampling attention module during training, represents the parameters other than the detail-aware sampling attention module parameters; Represents the learnable parameters corresponding to the three fully connected layers in the perceptual attention module; represents the L2 regularization term, Represents the regularization coefficient, which is used to control the strength of the regularization term. By adding an L2 regularization term to each fully connected layer, the weight amplitude of each layer is limited to avoid overfitting.

[0057] S3 uses the trained cylinder inner wall wear detection model to perform cylinder inner wall wear detection. Figure 1The cylinder inner wall wear detection model of the present invention is a DSA-YOLO model, which includes a Backbone module, a Neck module and a Head module. The Backbone module includes modules such as C2f, SCDown and SPPF, which are used to extract basic features of the image. The Neck module is used to further process the basic features extracted by the Backbone module, which includes Concat, C2f / CIB, Conv, SCDown and the above-mentioned detail perception sampling attention module, namely DSA. The Head module is used to detect and classify targets on the output of the Neck module.

[0058] The engine cylinder wear images are collected to construct a training set, where each image is equipped with a label indicating whether it shows wear. The collected engine cylinder images are preprocessed, including resizing and normalization, to meet the input requirements of the DSA-YOLO model. The training set is used to train the constructed cylinder inner wall wear detection model.

[0059] In this step, the cylinder inner wall wear detection model with detail perception sampling attention module is used to detect the cylinder inner wall wear, as follows:

[0060] The generator cylinder wear image to be detected is preprocessed and then input into the Backbone module, and the basic features of the image are extracted through modules such as C2f, SCDown and SPPF.

[0061] The extracted basic features are further processed by the Neck module. In the Neck module, the DSA module replaces the traditional upsampling module Upsample. The DSA module constructs a bilinear rearrangement feature map, learns the sampling weight map, and processes the bilinear rearrangement feature map using a mean filter. It superimposes the aggregation results of the bilinear rearrangement feature map and the sampling weight map to generate a more detailed sampling feature map, thereby enhancing the ability to capture the wear details of the cylinder inner wall. The feature map processed by the DSA module is fused with other feature maps in the Neck module. This step combines feature maps from different sources through operations such as Concat to form a rich feature representation, providing more comprehensive information for subsequent detection tasks.

[0062] The fused feature map is passed to the Head module for target detection and classification. The Head module is responsible for predicting the location and category of the wear area on the cylinder wall. The output of the Head module may contain multiple prediction boxes, and post-processing techniques need to be applied to filter and optimize the prediction results to reduce redundancy and improve accuracy.

[0063] The model outputs the final wear area detection results, including location (bounding box), category (such as normal or worn), and confidence score, which can be used for further analysis or directly used for engine maintenance decisions.

[0064] Existing detection models usually rely on traditional upsampling methods, such as nearest neighbor interpolation or bilinear interpolation, which are prone to losing detail information during multi-scale processing, especially in low-resolution images, resulting in the inability to effectively capture subtle wear features. In addition, traditional methods may fail to detect low contrast and fine structures when processing complex scenes, resulting in reduced detection accuracy. The cylinder inner wall wear detection model DSA-YOLO of the present invention can enhance the ability to capture subtle wear while retaining image details by introducing a detail-aware sampling attention module (DSA), thereby improving detection accuracy and robustness, and is particularly suitable for complex and low-quality endoscopic images. Through this innovation, DSA-YOLO overcomes the shortcomings of traditional models and provides higher accuracy and real-time performance.

[0065] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0066] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not disclosed in the present disclosure.

[0067] The above is only a specific implementation of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A cylinder inner wall wear detection method based on detail perception sampling attention module, characterized in that: The following steps are involved: Constructing a detail-aware sampling attention module, wherein the detail-aware sampling attention module constructs a bilinear rearrangement feature map based on the input feature map, learns a sampling weight map from the bilinear rearrangement feature map, processes the bilinear rearrangement feature map using a mean filter, and superimposes an aggregated result of the bilinear rearrangement feature map and the sampling weight map as a sampling feature map; The upsampling module of the YOLO network is replaced with the detail-aware sampling attention module and a cylinder inner wall wear detection model is constructed; and the cylinder inner wall wear detection is performed using the trained cylinder inner wall wear detection model; The detail-aware sampling attention module includes a bilinear rearrangement feature map construction part, a perception attention module and a mean residual aggregation module; the bilinear rearrangement feature map construction part uses a bilinear transformation to determine a sampling grid, and uses the sampling grid as an index to obtain the original feature points corresponding to each feature point after sampling, and the bilinear rearrangement feature map is composed of these original feature points; the perception attention module mainly includes a fully connected layer, which uses the fully connected layer to extract, transfer and enhance the context information of each original feature point layer by layer, learn the spatial correlation of features in the bilinear rearrangement feature map, and then obtain a sampling weight map; the mean residual aggregation module uses a mean filter to process the bilinear rearrangement feature map, and then superimposes the feature map processed by the mean filter with the aggregation result of the bilinear rearrangement feature map and the sampling weight map to generate the sampling feature map; The steps to build the perception attention module are as follows: Among them, F R represents the bilinear rearrangement feature map; FC U , FC K and FC W Represent three fully connected layers, α U ,α K and α W are the learnable parameters corresponding to the three fully connected layers; BN is the batch normalization layer; ReLU represents the nonlinear activation function; The perceptual attention module inputs the FC U The bilinear rearrangement feature map F R Mapped to a higher dimensional space to obtain a feature map p represents the feature map Z U The number of columns; feature map Z U Input to FC K , further integrate and refine features to obtain feature maps q represents the feature map Z K The number of columns; feature map Z K Via FC W Perform dimension alignment to obtain the feature map 4 represents the feature map Z W Has 4 columns; Finally, the feature map Z W Input to the Softmax function for normalization, where dim = 1 means that the Softmax function is normalized to Z W Normalize by column to obtain the sampling weight map C, h, and w represent the number of channels of the input feature map and the height and width of the output feature map, respectively.

2. A cylinder inner wall wear detection method based on detail perception sampling attention module as claimed in claim 1, characterized in that: The constructed bilinear rearrangement feature map expression is as follows: in, represents the sampling grid obtained by the bilinear transformation algorithm, h, w represent the height and width of the output feature map, 4 represents the four original feature points corresponding to each feature point after sampling, and 2 represents that G records the x and y coordinates of each original feature point; represents the input feature map, C, H, W represent the number of channels, height and width of the input feature map; It represents the feature map that indexes the input feature map F according to the sampling grid G, that is, according to the coordinate information in the sampling grid, the pixel value of the original feature point corresponding to each feature point after sampling is obtained. Its dimension is Reshape2d represents Rearrange it into a matrix with C×h×w rows and 4 columns, and get the bilinear rearrangement feature map F R .

3. The cylinder inner wall wear detection method based on detail perception sampling attention module according to claim 1 is characterized in that: The steps to construct the mean residual aggregation module are as follows: Among them, F L Represents the bilinear rearrangement feature map F R The low-frequency feature map obtained by mean filtering, Mean represents mean filtering, and the corresponding dim=3 represents mean filtering in The dimension corresponding to 4 in the figure is as follows; F H represents the features learned by the perceptual attention module, ⊙ represents the dot product operation, and F R ⊙Z S Represents the bilinear rearrangement feature map F R and the sampling weight map Z S Perform dot multiplication, SUM represents the sum operation, and the corresponding dim=3 represents the sum in F R ⊙Z S The corresponding dimensions of 4 are performed; Reshape3d represents a tensor shape conversion operation; F O represents the final sampling feature map, which is F L and F H The summation result is rearranged into a feature map of dimension C×h×w via Reshape3d, where C, h, and w represent the number of channels of the input feature map and the height and width of the output feature map, respectively.

4. The cylinder inner wall wear detection method based on detail perception sampling attention module according to claim 1 is characterized in that: The cylinder inner wall wear detection model is connected to the target detection loss function, and the parameters of the detail perception sampling attention module are regularized. The gradient descent method is used for training, so that the detail perception sampling attention module is trained collaboratively with other modules of the YOLO network to obtain the trained cylinder inner wall wear detection model.

5. The cylinder inner wall wear detection method based on detail perception sampling attention module according to claim 4 is characterized in that: The target detection loss function expression of the cylinder inner wall wear detection model is as follows: L det represents the target detection loss function, which includes the target box classification loss and the target box regression loss; F O represents the sampling feature graph; represents other feature maps except the input feature map F of the detail-aware sampling attention module during training, Ω represents parameters except the parameters of the detail-aware sampling attention module; α U , α K , α W represents the learnable parameters corresponding to the three fully connected layers in the perceptual attention module; ‖‖2 represents the L2 regularization term, λ>0 represents the regularization coefficient, which is used to control the strength of the regularization term. By adding the L2 regularization term to each fully connected layer, the weight amplitude of each layer is limited to avoid overfitting.

6. The cylinder inner wall wear detection method based on detail perception sampling attention module according to claim 1 is characterized in that: The cylinder inner wall wear detection model includes a Backbone module, a Neck module and a Head module. The Backbone module is used to extract the basic features of the image; the Neck module is used to further process the basic features extracted by the Backbone module, and is provided with multiple sampling modules; the Head module is used to perform target detection and classification on the output of the Neck module; the detail-aware sampling attention module replaces the upsampling module in the Neck module.

Images