Small and special electric machine armature defect detection method based on improved YOLOv11n

Abstract

The invention relates to the technical field of microelectronics, and discloses a small and special electric machine armature defect detection method based on improved YOLOv11n, which comprises the following steps: S1, acquiring small and special electric machine armature defect image data; s2, constructing an armature defect network model, wherein an original YOLOv11n model structure is improved by constructing a C3K2-MA module; a lightweight feature extraction module is used to replace a convolution module in a trunk feature extraction and down-sampling process, and a down-sampling module is improved; the CIoU of the original YOLOv11n model is improved by using DIOU, so that the model precision is improved, and a loss function is improved; s3, using the training set and the test set obtained in the step S1 to train through the armature defect network model constructed in the step S2; and S4, using the trained armature defect network model to carry out field detection on the armature defects of the small and special electric machine. According to the invention, a partial multi-scale feature aggregation module C3K2-MA is provided, the detection capability of different scale targets is significantly improved, and the detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of microelectronics, and in particular to a method for detecting armature defects of micro motors based on an improved YOLOv11n. Background Art

[0002] Micro motors are essential components for industrial, office, and home automation, and enjoy broad market demand. With the continuous emergence of new materials, processes, and technologies, the micro motor industry is undergoing continuous technological innovation and product upgrades. Currently, brushless DC motors, due to their simple structure, strong intelligent control capabilities, low noise, and high energy efficiency, are becoming a key trend in the industry's future development. However, due to the maturity of brushed DC motor technology and their relatively low price, traditional brushed DC motors still dominate the market. As a core component of micro motors, the quality of the armature significantly impacts motor performance. Therefore, quality control of the armature is particularly important.

[0003] Traditional armature defect detection relies primarily on manual labor. These methods, however, suffer from significant bias in human judgment, high labor costs, and slow processing speeds, making them inadequate for the rapid expansion of production capacity in the micromotor sector. Despite extensive research into target detection technologies, micromotor armature surface defect detection often suffers from omissions of key attributes and insufficient feature learning when processing samples with subtle defects.

[0004] In recent years, research on target detection of micro motor armature defects has also been innovative, such as:

[0005] Prior art 1

[0006] An existing Chinese patent, CN116468719A, provides a method for detecting armature defects in micro vibration motors based on semantic segmentation. The method uses a semantic segmentation network to perform semantic segmentation on images of the micro vibration motor to be tested, obtaining a feature map. The segmented feature map is then visualized and the feature map is then identified to determine whether an armature defect exists. The existing technology suffers from the high complexity of the multi-stage model, which affects speed, insufficient detection of small defects, reliance on large amounts of labeled data, insufficient application of semantic segmentation, high hardware requirements, and limited dynamic adjustment capabilities.

[0007] Prior Art 2: A Chinese patent with the publication number CN113674242A discloses a rotor defect detection method based on a generative adversarial network. First, data collection of solder joint characteristics is carried out to generate a solder joint feature template library. According to the significant features of qualified solder joints, over-soldering, under-soldering, welding offset, missing solder, and false soldering, classification is performed, and samples are analyzed. The image shape of the solder joints of the motor rotor is compared with a preset detection template image to obtain the matching degree between the solder joint image and the standard image, and different types of solder joint features are identified. Finally, a verification is carried out. If the verification is available, it is reserved as a sample for use as a feature template. The defect of this prior art is that the detection of tin stripping defects depends on complex auxiliary features and combined logic conditions, which are prone to misjudgment and missed judgment; and it can only detect specific defects such as false soldering, solder balls, and tin stripping, and other defects such as winding short circuits cannot be detected. Summary of the Invention

[0008] In view of the problems existing in the prior art of armature defect detection algorithms for micro and special motors, such as low detection accuracy, especially easy misjudgment of similar workpieces, the present invention proposes a method for detecting armature defects of micro and special motors based on improved YOLOv11n. By using a feature extraction module and a multi-scale feature fusion method, the network structure is optimized and streamlined, reducing model parameters and computational complexity, thereby solving the problems of complex models, slow detection, and high hardware requirements; effectively improving the mean average precision (mAP) of armature defect detection of micro and special motors. Compared with the existing YOLO algorithm, the accuracy and detection effect of the improved YOLO algorithm have been greatly improved, and on this basis, the number of model parameters and the model size have been reduced.

[0009] The present invention provides the following technical solutions:

[0010] The present invention provides a method for detecting armature defects of micro and special motors based on improved YOLOv11n, which includes the following steps:

[0011] S1: Collect image data of armature defects of micro and special motors, preprocess the image data, label the images and perform data augmentation to obtain a data set, and divide it into a training set, a test set, and a validation set;

[0012] S2: Construct an armature defect network model. The improvement of the original YOLOv11n model structure is achieved by constructing a C3K2-MA module to improve the C3K2 in the Backbone module and the Head module of the original YOLOv11n model structure; by replacing the convolutional modules in the backbone feature extraction and downsampling processes with a lightweight feature extraction module to improve the downsampling module; by using DIoU to improve the CIoU of the original YOLOv11n model to improve the model accuracy and improve the loss function;

[0013] S3: Use the training set obtained in step S1 and the test set to train through the armature defect network model constructed in step S2, and adjust the hyperparameters to obtain a trained armature defect network model;

[0014] S4: Use the trained armature defect network model to conduct on-site detection of the armature defects of the special micro-motor.

[0015] According to some embodiments, step S1 specifically includes the following steps:

[0016] S1.1: Collect images of the armature defects of the special micro-motor through on-site collection to obtain image data of the armature defects of the special micro-motor;

[0017] S1.2: Preprocess the obtained image data to eliminate blurred, ghosted, and similar images;

[0018] S1.3: Use annotation software to perform YOLO format annotation on the image data to ensure that the bounding box fits the contour of the armature defect and there is no excessive invalid information;

[0019] S1.4: Perform enhancement processing on each image. The enhancement processing includes any one operation or a combination of operations such as cropping, translation, flipping, adding Gaussian noise, and randomly rotating by any angle to obtain processed image data;

[0020] S1.5: Construct a data set according to the processed image data and the annotation file;

[0021] S1.6: Divide the obtained data set into a training set, a test set, and a validation set according to the ratio of 7:2:1.

[0022] According to some embodiments, the annotation software in step S1.3 is selected from LabelImg annotation software.

[0023] According to some embodiments, step S2 specifically includes the following steps:

[0024] S2.1: Construct the C3K2-MA module, where C is the number of input channels; C / / 2 is evenly split into two parts along the channel dimension; Conv1 is the first convolutional layer with equal input and output channel numbers and a convolutional kernel size of 3; Conv2 is the second convolutional layer with equal input and output channel numbers and a convolutional kernel size of 5; Conv3 is the third convolutional layer with equal input and output channel numbers and a convolutional kernel size of 7; Conv4 is the fourth convolutional layer with equal input and output channel numbers and a convolutional kernel size of 1; Concat is the feature map concatenation operation; the '+' sign represents the addition operation; significant feature information is extracted through convolutional kernels of different sizes to capture multi-scale feature information from the input; a complete convolutional operation is performed on some channels, that is, the Conv1 and Conv2 convolutional layers are evenly split into two parts along the channel dimension using the chunk function to improve computational efficiency; subsequently, the C3K2-MA module fuses the multi-scale features through a concatenation operation and a 1×1 convolutional layer;

[0025] S2.2: Use the C3K2-MA module to improve the C3K2 in the Backbone and head of the original YOLOv11n model;

[0026] S2.3: Replace the convolutional modules in the backbone feature extraction and downsampling processes with a lightweight feature extraction module to improve the downsampling module. In the lightweight feature extraction module, Conv2d is the convolutional 2d; Batch Norm2d is the BN normalization; SiLU is the activation function; AvgPool2d is the average pooling 2d; Split is the block division; Conv is the convolutional 2d; Maxpool2d is the max pooling 2d; Concat is the concatenation operation; the average pooling strategy is adopted to reduce the size of the feature map to reduce the computational amount; subsequently, a two-stream processing architecture is formed through channel dimension splitting: one part strengthens the local feature extraction through the max pooling layer, and the other part directly enters the convolutional layer; the parameter sharing mechanism is adopted to make the two-branch convolutional layers share the weight matrix, thereby effectively reducing the parameter scale while ensuring the feature expression ability; finally, the results of the two branches are spatially concatenated to generate the output feature map;

[0027] S2.4: Use the DIoU to improve the CIoU of the original YOLOv11n model to improve the model accuracy and accelerate the bounding box regression speed. The loss function expression of the DIoU is as follows:

[0028]

[0029] In the above formula, ρ represents the straight-line distance between the center of the detection box and the center of the annotation box, C is the spatial diagonal length of the smallest enclosing rectangle of the two boxes, and b and b gt represent the center points of the anchor box and the target box respectively.

[0030] According to some embodiments, step S3 specifically includes the following steps:

[0031] S3.1: Use the training set as input, output a fused feature map through the constructed armature defect network model, and obtain a weight file;

[0032] S3.2: The armature defect network model predicts the validation set through weight parameters to initially obtain performance indicators;

[0033] S3.3: Adjust the hyperparameters of the armature defect network model according to the performance indicators, and return to step S3.1 for execution. After obtaining the optimal weights, jump to S3.4;

[0034] S3.4: Use the optimal weights to test the model generalization ability of the test set, verify the effectiveness and feasibility of the network model, and obtain a trained armature defect network model.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The method for detecting armature defects of a special micro-motor based on improved YOLOv11n provided by the present invention first adopts the ideas of efficient partial convolution and residual connection, designs a partial multi-scale feature aggregation module C3K2-MA, significantly improves the detection ability of targets with different scales, and improves the detection accuracy; secondly, introduces a lightweight feature extraction module (ADown) to replace the convolution module in the backbone feature extraction and downsampling process, reducing the number of parameters and computational complexity of the C3K2-MA module; finally, in order to make up for the problems of weak generalization and slow convergence speed of the CIoU loss function in the detection task, uses the DIoU loss function to improve the model accuracy and accelerate the bounding box regression speed; through the present invention, the number of parameters and computational complexity of the model can be reduced while identifying armature defects, providing effective technical support for meeting the deployment requirements of mobile and embedded devices.

[0037] The present invention can greatly improve the detection speed, meet the requirements of industrial real-time detection; significantly improve the detection accuracy of tiny defects and multiple defects; reduce data annotation and hardware costs; flexibly adapt to different production scenarios, and provide a more efficient, accurate and low-cost solution for the detection of armature defects of special micro-motors. Description of the Drawings

[0038] Figure 1 It is a schematic diagram of the C3K2-MA module provided by an embodiment of the present invention.

[0039] Figure 2 It is an overall structure diagram of the improved model provided by an embodiment of the present invention.

[0040] Figure 3This is a comparison diagram of the normal downsampling and ADown downsampling structures provided by an embodiment of the present invention.

[0041] Figure 4 Schematic diagram of the DIoU loss function provided by an embodiment of the present invention.

[0042] Figure 5 This is a flowchart provided for an embodiment of the present invention.

[0043] Figure 6 This is a visualization analysis diagram of the training results of the improved algorithm for the micro motor armature defect detection method based on the improved YOLOv11n provided in an example of the present invention.

[0044] Figure 7 A comparison chart of the micro motor armature defect detection results based on the improved YOLOv11n algorithm and the original YOLOv11n algorithm is provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention is described in detail below with reference to the embodiments and accompanying drawings. However, it should be understood that the embodiments and accompanying drawings are merely exemplary descriptions of the present invention and do not constitute any limitation on the scope of protection of the present invention. All reasonable variations and combinations within the scope of the inventive concept of the present invention fall within the scope of protection of the present invention.

[0046] The present invention will be further described below with reference to the accompanying drawings.

[0047] Example 1

[0048] The flow chart provided in this embodiment is as follows Figure 5 shown

[0049] S1: Collect image data of micro motor armature defects, preprocess the collected data, annotate the images and perform data enhancement to obtain a data set, which specifically includes the following steps:

[0050] S1.1 Collect representative images of micro motor armature defects;

[0051] S1.2 preprocess the acquired image data to remove blurry, ghosting, and similar images;

[0052] S1.3 Use LabelImg annotation software to annotate the image data in YOLO format to ensure that the bounding box fits the armature defect outline and contains no excessive invalid information;

[0053] S1.4 performs enhancement processing on each image, such as cropping, translating, flipping, adding Gaussian noise, and randomly rotating at any angle, to obtain image data;

[0054] S1.5 Construct a dataset based on the image data and annotation files;

[0055] S1.6 Divide the obtained dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1.

[0056] S2: Construct a network model for armature defects, improve the original model structure of YOLOv11n (the original YOLOv11 model is open-source code provided by Ultralytics, but there is already a paper using YOLOv11n to complete the object detection task. The source of this paper is: [1] He Zhixuan, Chen Lili, Wang Xiang, et al. DMF-YOLOv11: An Object Detection Algorithm for UAV Aerial Images Based on Improved YOLOv11n [J / OL]. Computer Engineering and Applications, 1-14 [2025-04-16]. http: / / kns.cnki.net / kcms / detail / 11.2127.tp.20250403.2136.016.html.), and use the C3K2-MA module to improve the C3K2 in the Backbone module and the Head module, improve the downsampling module, and improve the model loss function. The specific steps are as follows:

[0057] S2.1 Construct a partial multi-scale feature aggregation module C3K2-MA (C3K2-Multi ScalePartialFeature Aggregation, MPFA), as Figure 1 shown. Figure 1 In it, C is the number of input channels; C / / 2 is evenly split into two parts along the channel dimension; Conv1 is the first convolutional layer, with the same number of input and output channels and a convolutional kernel size of 3; Conv2 is the second convolutional layer, with the same number of input and output channels and a convolutional kernel size of 5; Conv3 is the third convolutional layer, with the same number of input and output channels and a convolutional kernel size of 7; Conv4 is the fourth convolutional layer, with the same number of input and output channels and a convolutional kernel size of 1; Concat is the feature map concatenation operation; the "+" sign represents the addition operation.

[0058] S2.2 Improve the original model structure of YOLOv11n. After improvement, as Figure 2 shown, use the C3K2-MA module to improve the C3K2 in the original model's Backbone and head. Figure 2Among them, note that Conv is the most basic convolutional block in the architecture, used to extract local features of the input feature map; C3K2-MA is a partial multi-scale feature aggregation module; ADown is a downsampling module; Upsample is an upsampling operation that enlarges the spatial size of the feature map; Concat is a feature map concatenation operation that concatenates feature maps along the channel dimension, and Detect is a detection head.

[0059] The design of this convolutional module and the improvement of the model structure effectively retain the original information and introduce new multi-scale information, improving the recognition accuracy of the model.

[0060] S2.3 Improve the downsampling module, and use a lightweight feature extraction module (ADown) to replace the convolutional module in the backbone feature extraction and downsampling process, as Figure 3 shown. Figure 3 Among them, Conv2d is convolutional 2d; Batch Norm2d is BN normalization; SiLU is an activation function; AvgPool2d is average pooling 2d; Split is block splitting; Conv is convolutional 2d; Maxpool2d is max pooling 2d; Concat is a concatenation operation.

[0061] This improvement method reduces the number of parameters and computational amount of the C3K2-MA module, and there is also a significant reduction in the number of parameters and model size compared with the original YOLOv11n model. In this embodiment, by replacing the traditional downsampling Conv layer, the detection performance of the model for complex targets is significantly improved, especially showing stronger robustness in the armature defect detection scenario;

[0062] S2.4 Improve the model loss function, use DIoU to improve the CIoU of the original model to improve the model accuracy and speed up the bounding box regression speed. DIoU adds a penalty on the basis of the IoU loss function, which can minimize and normalize the distance between the center points, and speeds up the convergence process, as Figure 4 shown. The expression of the DIoU loss function is as follows:

[0063]

[0064] In the formula, ρ represents the straight-line distance between the center of the detection box and the center of the labeled box, C is the spatial diagonal length of the smallest enclosing rectangle of the two boxes, and b and b gt represent the center points of the anchor box and the target box respectively. By fusing the two-dimensional information of the region overlap degree and the geometric center offset, when the target box completely contains and there is overlap, the constraint mechanism based on the center point spacing can significantly improve the parameter convergence efficiency.

[0065] The improvement of the loss function provided in this step makes up for the weak generalization and slow convergence of the CIoU loss function in the detection task, which can make the prediction of subsequent detection anchor boxes more accurate and improve the accuracy of the network model to a certain extent.

[0066] In step S2, the Backbone module converts the raw data of the input training set into feature maps of different semantic levels by downsampling. The Backbone module includes the Conv module, C3K2-MA module, ADown module, SPPF module and C2PSA. The specific structure is as follows: Figure 2 As shown;

[0067] The Head module comprehensively utilizes the feature information of different scales output by the Backbone module through upsampling and feature fusion of each layer, and outputs feature information. Then, the detection head with a decoupled head structure is used to perform target detection on the output feature information, predict the bounding box and its category and confidence.

[0068] S3: Use the training set and test set obtained in step S1 to train the network model constructed in step S2 and adjust the hyperparameters. Specifically, the following steps are included:

[0069] S3.1 takes the training set as input, outputs the fused feature map through the network model and obtains the weight file;

[0070] S3.2 The network model predicts the validation set using weight parameters and obtains preliminary performance indicators;

[0071] S3.3 adjusts the model hyperparameters based on the indicators and returns to step S3.1 until the optimal weight is obtained;

[0072] S3.4 Use the optimal weights to test the model generalization ability on the test set to verify the effectiveness and feasibility of the network model and obtain the trained network model;

[0073] S4: Use the trained network model to conduct field detection of micro motor armature defects.

[0074] Example 2

[0075] This example adopts a modular improvement strategy, screening the optimal architecture design through C3K2-MA replacement experiments. The specific implementation process includes three comparison groups: the first group replaces the original C3K2 modules in the backbone network; the second group replaces the corresponding components in the head network; and the third group replaces the C3K2 modules in both the backbone and head networks simultaneously.

[0076] In this embodiment, to verify the effectiveness of the adopted ADown module in reducing the number of parameters of the original network model, the backbone network, the head network, and the overall backbone and head network were verified respectively;

[0077] In this embodiment, to address the problems of weak generalization and slow convergence rate of the CIoU loss function in detection tasks, the impacts of different loss functions such as EIOU, Inner-MDPIoU, ShapeIoU, and DIoU on the model were verified;

[0078] In this embodiment, to verify the effectiveness of the algorithm for detecting armature defects of special micro-motors, the algorithm proposed in Embodiment 1 was trained and tested on a self-built dataset of armature defects of special micro-motors together with high-performance models such as Faster R-CNN, YOLOv5s, YOLOv8n, and YOLOv11n; As Figure 6 , is the analytical diagram of the visualized training results of the improved algorithm in this example; As Figure 7 , is the comparison diagram of the detection results of the improved algorithm and the original YOLOv11n algorithm for armature defects of special micro-motors in this example.

[0079] The above embodiments are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. A method for detecting defects in the armature of a special micro-motor based on improved YOLOv11n, characterized in that: The following steps are involved: S1: Collect image data of micro motor armature defects, preprocess the image data, annotate the images and perform data enhancement to obtain a dataset, which is divided into training set, test set and validation set; S2: Constructing an armature defect network model and improving the original YOLOv11n model structure by constructing a C3K2-MA module to improve the C3K2 in the Backbone module and Head module of the original YOLOv11n model structure; replacing the convolution module in the backbone feature extraction and downsampling process with a lightweight feature extraction module to improve the downsampling module; improving the CIoU of the original YOLOv11n model by using DIoU to improve the model accuracy and improve the loss function; S3: using the training set and test set obtained in step S1 to train the armature defect network model constructed in step S2, and adjusting the hyperparameters to obtain a trained armature defect network model; S4: Use the trained armature defect network model to conduct field detection of micro motor armature defects.

2. The method for detecting defects of the armature of a special micro-motor based on the improved YOLOv11n according to claim 1, wherein: Step S1 specifically includes the following steps: S1.1: Collect images of micro motor armature defects through field collection to obtain image data of micro motor armature defects; S1.2: Preprocess the acquired image data to remove blur, ghosting, and similar images; S1.3: Use annotation software to annotate the image data in YOLO format to ensure that the bounding box fits the armature defect outline and does not contain excessive invalid information; S1.4: performing enhancement processing on each image, wherein the enhancement processing includes any operation or any combination of cropping, translation, flipping, adding Gaussian noise, and random rotation of any angle to obtain processed image data; S1.5: Construct a dataset based on the processed image data and annotation files; S1.6: Divide the obtained dataset into training set, test set and validation set in a ratio of 7:2:

1.

3. The method for detecting defects of the armature of a special micro-motor based on the improved YOLOv11n according to claim 2, wherein: The labeling software in step S1.3 is selected from LabelImg labeling software.

4. The method for detecting defects of the armature of a special micro-motor based on the improved YOLOv11n according to claim 2, characterized in that: Step S2 specifically includes the following steps: S2.1: Construct the C3K2-MA module, where C is the number of input channels; C / / 2 is evenly split into two parts along the channel dimension; Conv1 is the first convolution layer, the number of input and output channels is equal, and the convolution kernel size is 3; Conv2 is the second convolution layer, the number of input and output channels is equal, and the convolution kernel size is 5; Conv3 is the third convolution layer, the number of input and output channels is equal, and the convolution kernel size is 7; Conv4 is the fourth convolution layer, the number of input and output channels is equal, and the convolution kernel size is 1; Concat is the feature map splicing operation; the "+" sign represents the addition operation; significant feature information is extracted by convolution kernels of different sizes to capture multi-scale feature information from the input; full convolution operation is performed on some channels, that is, the chunk function is used to split the Conv1 and Conv2 convolution layers into two parts along the channel dimension to improve computational efficiency; then, the C3K2-MA module fuses multi-scale features through splicing operation and 1×1 convolution layer; S2.2: Improve C3K2 in the Backbone and head of the original YOLOv11n model using the C3K2-MA module; S2.3: Replace the convolution modules in the backbone feature extraction and downsampling processes with a lightweight feature extraction module, and improve the downsampling module. In the lightweight feature extraction module, Conv2d is a 2D convolution; Batch Norm2d is BN normalization; SiLU is an activation function; AvgPool2d is average pooling 2D; Split is block splitting; Conv is a 2D convolution; Maxpool2d is max pooling 2D; Concat is a concatenation operation; adopt an average pooling strategy to reduce the size of the feature map to reduce the computational cost; then form a dual-stream processing architecture through channel dimension splitting: one part enhances local feature extraction through a max pooling layer, and the other part directly enters the convolution layer; adopt a parameter sharing mechanism to make the two-branch convolution layers share the weight matrix, so as to effectively reduce the parameter scale while ensuring the feature expression ability; finally, perform spatial concatenation on the results of the two branches to generate the output feature map; S2.4: Use the DIoU to improve the CIoU of the original YOLOv11n model to improve the model accuracy and accelerate the bounding box regression speed. The loss function expression of the DIoU is as follows: In the above formula, ρ represents the straight-line distance between the center of the detection box and the center of the annotation box, C is the length of the space diagonal of the smallest enclosing rectangle of the two boxes, and b and b gt represent the center points of the anchor box and the target box respectively.

5. The method for detecting the defects of the armature of a special micro-motor based on the improved YOLOv11n according to claim 4, wherein: Step S3 specifically includes the following steps: S3.1: Use the training set as the input, output a fused feature map through the constructed armature defect network model, and obtain a weight file; S3.2: The armature defect network model predicts the validation set through the weight parameters to initially obtain performance indicators; S3.3: Adjust the hyperparameters of the armature defect network model according to the performance indicators, and return to step S3.1 for execution. After obtaining the optimal weights, jump to S3.4; S3.4: Use the optimal weights to test the model generalization ability of the test set, verify the effectiveness and feasibility of the network model, and obtain a trained armature defect network model.

Images