CN118305101A - Automatic bearing defect detection system and method - Google Patents

Abstract

The present invention discloses an automated bearing defect detection system and method. The system includes a parts supply module, a parts conveying module, a lighting and image acquisition module, a visual inspection module, a control module and a safety module. The method is as follows: first, the parts supply module uses an automatic rotating material tray to automatically load the parts to be inspected, and the parts conveying module uses a conveyor belt and a loading robot to transport the bearings to the visual inspection area; then the lighting and image acquisition module shoots the bearings from multiple angles, sends them to the visual inspection module, uses the detection algorithm and edge computing equipment to detect whether the bearing parts have defects, and sends the detection results to the control module; finally, the control module transports the inspected bearings to the waste area and the unloading area respectively; the safety module protects the safety of personnel and equipment during the automated inspection process, and promptly detects and responds to potential dangerous situations. The present invention improves the automation level, detection efficiency and detection accuracy of bearing defect detection.

Description

An automated bearing defect detection system and method

Technical Field

The present invention relates to the technical field of computer vision detection, and in particular to an automatic bearing defect detection system and method.

Background technique

With the development of science and technology, automatic detection and classification of objects can be achieved through image processing and target detection algorithms. At present, computer vision technology is mostly used to achieve quality inspection of industrial production, and the research on bearing defect detection system includes the use of computer vision technology to capture and analyze bearing images in real time to detect surface defects, dimensional deviations and other problems. However, modern bearing defect detection systems do not rely solely on a single technical means, but combine multiple technologies such as computer vision, deep learning, and sensor technology to build an intelligent detection system. These systems are usually characterized by automation, high efficiency, and high accuracy, and can meet the needs of industrial production lines.

A visual inspection method and a visual inspection device disclosed in the application publication number CN113552123A, the method relates to the field of visual inspection, and mainly includes the following steps: acquiring image detection data of at least one image acquisition device from a remote end, and then generating a detection result using a standard image database and the image detection data, the detection result including a judgment on whether the image detection data is abnormal, and the corresponding abnormal category under abnormal circumstances. This method separates image acquisition from processing, and improves the flexibility of the layout. In addition, at least one image acquisition device can share a standard image database, computing resources, and processing resources, which not only reduces the cost of setting up image acquisition devices locally, but also utilizes better resources and richer data, thereby improving the quality of visual inspection. A visual inspection method and system with the application publication number CN112240888A, visually inspects products through a visual inspection system, and uses a sensor to detect the time and number of each product passing through. A synchronous speed measuring device records the initial position information of each product when it passes through the sensor, and tracks its moving distance. The server controls multiple cameras to detect multiple products at the same time, and obtains the moving distance of each product in real time. When the product moves a distance that enters the camera shooting range, the server triggers the camera to take a picture of the product. This method solves the problem that single vision cannot handle complex visual inspection content, and ensures that the overall inspection efficiency is not affected by the camera's shooting speed, thereby significantly improving the inspection efficiency.

Existing technology only uses computer vision for product defect detection. Material loading and sorting after detection require manual operation. There are delays and operational errors caused by human factors, low detection efficiency, insufficient monitoring data sampling and low detection accuracy.

Summary of the invention

The object of the present invention is to provide an automated bearing defect detection system and method with high automation, high detection efficiency and high detection accuracy.

The technical solution to achieve the purpose of the present invention is: an automated bearing defect detection system, including a parts supply module, a parts conveying module, a lighting and image acquisition module, a visual inspection module, a control module and a safety module;

The parts supply module uses an automatically rotating material tray to supply the parts to be inspected;

The parts conveying module includes a conveyor belt, a loading robot arm, a unloading robot arm and an operating unit, which are used to convey the parts to be inspected to the visual inspection area;

The lighting and image acquisition module includes a fill light source and an image acquisition module, which are used to capture the details of the parts to be inspected to meet the complex visual recognition requirements;

The visual inspection module includes a detection algorithm model and an edge computing device for detecting whether the bearing parts have defects;

The control module includes a main control unit and a control panel, which are used to monitor and control the coordinated operation between various modules, receive and process the detection data of the visual detection module, and make corresponding control decisions;

The safety module includes physical protection, emergency stop buttons and safety sensors to protect the safety of personnel and equipment during the detection process and to detect and respond to potential dangerous situations.

Furthermore, the automatic rotating tray of the parts supply module cooperates with the mechanical arm of the parts conveying module to clamp the parts. After clamping a part, the automatic rotating tray rotates once to complete the automatic loading process;

The automatic rotating material tray includes guide rods, a parts mounting plate, a stepper motor and a lifting mechanism arranged from top to bottom. The stepper motor is used to control the rotation angle of the material tray, and the loading task is completed in cooperation with the loading robot arm; the parts mounting plate is equipped with a detachable guide rod for fixing bearing parts to prevent the parts from sliding during rotation to complete the clamping task; the guide rods are adapted to bearing parts of different sizes, and the number of guide rods is selected according to actual conditions; the lifting mechanism adjusts the height of the material tray for docking with the work surface.

Furthermore, the operating unit of the parts conveying module controls the loading robot arm to place the parts at a specified position on the conveyor belt after clamping them, and arranges a batch of parts in a layout of 3 rows and 4 columns; the operating unit is provided with a coordinate system to guide the loading robot arm to accurately place the parts at the specified position on the conveyor belt; the unloading robot arm is controlled by the signal of the operating unit to unload the defective parts; the operating unit controls the speed of the conveyor belt to ensure that the parts are stably transmitted to the inspection area.

Furthermore, the fill light source of the lighting and image acquisition module adopts a multi-angle light source to reduce shadows and reflections, and the lighting intensity is adjusted according to the material and color of the part to obtain the best image quality; the image acquisition module uses a high-resolution camera to capture the tiny details of the part surface, and adopts multi-angle shooting to comprehensively detect every surface of the part.

Furthermore, the detection algorithm model of the visual inspection module includes a model training unit, a feature extraction unit, a defect detection unit, a feature fusion unit and a detection head unit;

The model training unit includes a model selection submodule, a model initialization submodule, a loss function submodule, an optimizer submodule, a batch processing submodule, a regularization submodule, a model evaluation submodule, and a model saving submodule;

The model selection submodule is used to select a machine learning or deep learning model structure that is suitable for the problem; the model initialization submodule is used to initialize the weights and biases of the model; the loss function submodule is used to optimize the model's loss function, which is a function used to measure the difference between the model output and the actual label; the optimizer submodule selects and configures the optimization algorithm used to optimize the model; the batch processing submodule divides the data into small batches during training to calculate gradients and update model parameters; the regularization submodule adds regularization terms to prevent the model from overfitting; the model evaluation submodule is used to evaluate the performance of the model on the training set, validation set, or test set; the model saving submodule saves the trained model weights for future use;

The feature extraction unit includes a convolution layer submodule, a pooling layer submodule, a normalization layer submodule, an activation layer submodule, a feature selection submodule and a feature dimension reduction submodule;

The convolution layer submodule applies convolution operation to extract image features; the pooling layer submodule reduces the spatial dimension of the feature map through pooling operation; the normalization layer submodule normalizes the features; the activation layer submodule applies activation function to introduce nonlinearity and increase the expressive power of the model; the feature selection submodule selects the most representative features for subsequent defect detection; the feature dimensionality reduction submodule reduces the dimension of the features;

The defect detection unit includes an image preprocessing submodule, a feature extraction submodule, a defect detection submodule, a defect location submodule and a visualization submodule;

The image preprocessing submodule preprocesses the input image, including resizing, cropping or enhancing the contrast of the image; the feature extraction submodule extracts features from the preprocessed image; the defect detection submodule uses the extracted features to detect defects in the image, including using classifiers or regressors; the defect localization submodule locates and marks the locations of the detected defects in the image; the visualization submodule visualizes the detected defect information for users to understand and analyze the detection results;

The detection process of the detection algorithm model is as follows:

(1) The image acquisition module acquires image information of the bearing parts and transmits the image information of the bearing parts to the image preprocessing submodule in the visual inspection module. The image preprocessing submodule performs denoising on the acquired image information of the bearing parts through a filtering algorithm to reduce the error of feature extraction and model training;

(2) The convolutional layer submodule in the feature extraction unit uses a lightweight model based on the improved YOLOv8 to perform convolution operations on the image information of the preprocessed bearing parts; in the backbone network of YOLOv8, the convolution layer of the downsampling operation is replaced by a lightweight adaptive convolution module LACM. The structure of LACM is divided into two paths, one path obtains an adaptive weight, and the other path uses group convolution to reduce the number of parameters, and finally multiplies and adds the obtained adaptive weights to obtain focused feature parameters; in the backbone network of YOLOv8, the Bottleneck layer used by C2f is replaced by a multi-scale convolution module EMC. The structure of EMC divides the input features into two parts, one part passes through 3×3Conv and 5×5Conv, and the other part retains the original features without operation; finally, the two parts of the features are combined to perform 1×1Conv; the bearing contour and texture required for bearing part defect detection are extracted;

(3) The model selection submodule in the model training unit selects the specified lightweight model based on the YOLOv8 improvement, and uses the preprocessed image data and extracted features to train the deep learning model; finally, YOLOv8m is used as the teacher model to perform knowledge distillation on the lightweight model based on the YOLOv8 improvement;

(4) The model evaluation submodule in the model training unit evaluates the performance of the trained deep learning model. The evaluation indicators include calculation accuracy, precision, and recall rate.

(5) The visualization submodule in the defect detection unit is electrically connected to the control panel of the control module to visualize the defect information and generate a bearing part defect information report, which includes the number of bearing part defects, defect locations and bearing defect types.

Furthermore, the loss function submodule of the visual detection module uses a composite loss function L _total to simultaneously optimize the position, size, object confidence, and category probability of the bounding box. The calculation formula of the loss function L _total is:

L _total = L _coord + L _conf + L _class

Where L _coord is the coordinate loss, using the squared error loss to measure the difference between the predicted bounding box and the true bounding box; L _conf is the confidence loss, using MSE or cross entropy loss to evaluate the accuracy of confidence prediction; L _class is the classification loss, using cross entropy loss to evaluate the prediction accuracy of class probability;

The cross entropy loss function L is:

Where M is the number of categories, yo _,c is the observed value, and _po,c is the predicted probability.

Furthermore, the control module adopts PLC as the main control unit of the system to control the rotation of the automatic rotating tray, the operation of the loading robot arm, the movement of the conveyor belt and the operation of the unloading robot arm, and uses an industrial PC to run a lightweight deep learning model to process the images taken by the image acquisition module, perform defect detection, and feed back the results to the PLC.

Furthermore, the physical protection monitors the environment around the loading robot arm and the unloading robot arm in real time through safety sensors, sets safety limits for the loading robot arm and the unloading robot arm, and controls the movement range of the loading robot arm and the unloading robot arm not to exceed a preset range;

The emergency stop button is used to stop the movement of the loading and unloading robotic arms in an emergency;

The safety sensor is used to monitor the status parameters of the loading robot arm and the unloading robot arm, and to issue an alarm and trigger an emergency stop button in case of abnormal conditions.

An automated bearing defect detection method based on the automated bearing defect detection system comprises the following steps:

Step 1: The parts supply module uses an automatic rotating tray to automatically load the parts to be inspected;

Step 2: The parts conveying module uses a conveyor belt and a loading robot to transport the bearing to the visual inspection area;

Step 3: The lighting and image acquisition module takes multi-angle photos of the bearing and sends them to the visual inspection module;

Step 4: The visual inspection module uses the inspection algorithm and edge computing device to detect whether the bearing parts are defective and sends the inspection results to the control module;

Step 5: The control module classifies the tested bearings, transports the defective bearings to the scrap area, and transports the non-defective bearings to the unloading area;

Step 6: The safety module protects the safety of personnel and equipment during the automated testing process, and promptly detects and responds to potential dangerous situations.

A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the automatic bearing defect detection method when executing the program.

Compared with the prior art, the present invention has the following significant advantages:

(1) By automating the loading, transporting, testing and unloading processes, the need for manual operation is reduced, and the speed and efficiency of the bearing defect detection process are improved;

(2) The automated system can operate continuously, reducing delays caused by human factors and achieving a highly efficient production process;

(3) By reducing reliance on manual operations, labor costs can be significantly reduced;

(4) Using deep learning algorithms to perform 360-degree inspections on bearings can efficiently and accurately identify various defects, such as cracks, scratches or other irregularities, ensuring that only qualified products can enter the next production link, making the inspection more accurate and consistent, and improving the overall quality of the product;

(5) By using the improved deep learning lightweight target detection algorithm, the number of parameters is greatly reduced while ensuring the detection accuracy, making it more suitable for application in industrial real-time detection scenarios.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a structural block diagram of an automatic bearing defect detection system according to the present invention.

FIG. 2 is a schematic diagram of the structure of an automated bearing defect detection system provided in an embodiment of the present invention.

FIG. 3 is a schematic diagram of the structure of the automatic rotating material tray in an embodiment of the present invention.

FIG. 4 is an algorithm principle diagram of a lightweight detection algorithm in an embodiment of the present invention.

Detailed ways

In conjunction with FIG1 , the present invention provides an automated bearing defect detection system, including a parts supply module, a parts conveying module, a lighting and image acquisition module, a visual inspection module, a control module and a safety module;

The parts supply module uses an automatically rotating material tray to supply the parts to be inspected;

The parts conveying module includes a conveyor belt, a loading robot arm, a unloading robot arm and an operating unit, which are used to convey the parts to be inspected to the visual inspection area;

The lighting and image acquisition module includes a fill light source and an image acquisition module, which are used to capture the details of the parts to be inspected to meet the complex visual recognition requirements;

The visual inspection module includes a detection algorithm model and an edge computing device for detecting whether the bearing parts have defects;

The control module includes a main control unit and a control panel, which are used to monitor and control the coordinated operation between various modules, receive and process the detection data of the visual detection module, and make corresponding control decisions;

The safety module includes physical protection, emergency stop buttons and safety sensors to protect the safety of personnel and equipment during the detection process and to detect and respond to potential dangerous situations.

As a specific example, the automatic rotating tray of the parts supply module cooperates with the mechanical arm of the parts conveying module to clamp the parts. After clamping a part, the automatic rotating tray rotates once to complete the automatic loading process;

The automatic rotating material tray includes guide rods, a parts mounting plate, a stepper motor and a lifting mechanism arranged from top to bottom. The stepper motor is used to control the rotation angle of the material tray, and the loading task is completed in cooperation with the loading robot arm; the parts mounting plate is equipped with a detachable guide rod for fixing bearing parts to prevent the parts from sliding during rotation to complete the clamping task; the guide rods are adapted to bearing parts of different sizes, and the number of guide rods is selected according to actual conditions; the lifting mechanism adjusts the height of the material tray for docking with the work surface.

As a specific example, the operating unit of the parts conveying module controls the loading robot arm to place the parts at a specified position on the conveyor belt after clamping them, and arranges a batch of parts in a layout of 3 rows and 4 columns; the operating unit is provided with a coordinate system to guide the loading robot arm to accurately place the parts at the specified position on the conveyor belt; the unloading robot arm is controlled by the signal of the operating unit to unload the defective parts; the operating unit controls the speed of the conveyor belt to ensure that the parts are stably transmitted to the inspection area.

As a specific example, the fill light source of the lighting and image acquisition module adopts a multi-angle light source to reduce shadows and reflections, and the lighting intensity is adjusted according to the material and color of the part to obtain the best image quality; the image acquisition module uses a high-resolution camera to capture the tiny details on the surface of the part, and adopts multi-angle shooting to comprehensively detect every surface of the part.

As a specific example, the detection algorithm model of the visual inspection module includes a model training unit, a feature extraction unit, a defect detection unit, a feature fusion unit and a detection head unit;

The model training unit includes a model selection submodule, a model initialization submodule, a loss function submodule, an optimizer submodule, a batch processing submodule, a regularization submodule, a model evaluation submodule, and a model saving submodule;

The model selection submodule is used to select a machine learning or deep learning model structure that is suitable for the problem; the model initialization submodule is used to initialize the weights and biases of the model; the loss function submodule is used to optimize the loss function of the model, which is a function used to measure the difference between the model output and the actual label; the optimizer submodule selects and configures the optimization algorithm used to optimize the model, such as stochastic gradient descent (SGD) or Adam; the batch processing submodule divides the data into small batches during training to calculate gradients and update model parameters; the regularization submodule adds regularization terms to prevent the model from overfitting; the model evaluation submodule is used to evaluate the performance of the model on the training set, validation set, or test set; the model saving submodule saves the trained model weights for future use;

The feature extraction unit includes a convolution layer submodule, a pooling layer submodule, a normalization layer submodule, an activation layer submodule, a feature selection submodule and a feature dimension reduction submodule;

The convolution layer submodule applies convolution operation to extract image features; the pooling layer submodule reduces the spatial dimension of the feature map through pooling operation, reducing the computational complexity; the normalization layer submodule normalizes the features to improve the stability and convergence speed of the model; the activation layer submodule applies activation function to introduce nonlinearity and increase the expressiveness of the model; the feature selection submodule selects the most representative features for subsequent defect detection; the feature dimensionality reduction submodule reduces the dimension of the features, reduces the computational cost and improves the generalization ability of the model;

The defect detection unit includes an image preprocessing submodule, a feature extraction submodule, a defect detection submodule, a defect location submodule and a visualization submodule;

The image preprocessing submodule preprocesses the input image, including resizing, cropping or enhancing the contrast of the image; the feature extraction submodule extracts features from the preprocessed image; the defect detection submodule uses the extracted features to detect defects in the image, including using classifiers or regressors; the defect localization submodule locates and marks the locations of the detected defects in the image; the visualization submodule visualizes the detected defect information for users to understand and analyze the detection results;

The detection process of the detection algorithm model is as follows:

(1) The image acquisition module acquires image information of the bearing parts and transmits the image information of the bearing parts to the image preprocessing submodule in the visual inspection module. The image preprocessing submodule performs denoising on the acquired image information of the bearing parts through a filtering algorithm to reduce the error of feature extraction and model training;

(2) The convolutional layer submodule in the feature extraction unit uses a lightweight model based on the improved YOLOv8 to perform convolution operations on the image information of the preprocessed bearing parts. In the backbone network of YOLOv8, the convolution layer of the downsampling operation is replaced by a lightweight adaptive convolution module LACM. The structure of LACM is divided into two paths. One path obtains an adaptive weight, and the other path uses group convolution to reduce the number of parameters. Finally, the adaptive weight is multiplied and added to obtain the focused feature parameters. This reduces the number of parameters while making the feature extraction more accurate. In the backbone network of YOLOv8, the Bottleneck layer used by C2f is replaced by a multi-scale convolution module EMC. The structure of EMC divides the input features into two parts. One part is processed by 3×3Conv and 5×5Conv, and the other part retains the original features without operation. Finally, the two parts of the features are combined to perform 1×1Conv to achieve the purpose of reducing parameters and improving feature representation. The bearing contour and texture required for bearing part defect detection are extracted.

(3) The model selection submodule in the model training unit selects the specified lightweight model based on the YOLOv8 improvement, and uses the preprocessed image data and extracted features to train the deep learning model; finally, YOLOv8m is used as the teacher model to perform knowledge distillation on the lightweight model based on the YOLOv8 improvement, thereby improving its feature extraction and fusion capabilities and improving detection accuracy;

(4) The model evaluation submodule in the model training unit evaluates the performance of the trained deep learning model. The evaluation indicators include calculation accuracy, precision, and recall rate.

(5) The visualization submodule in the defect detection unit is electrically connected to the control panel of the control module to visualize the defect information and generate a bearing part defect information report, which includes the number of bearing part defects, defect locations and bearing defect types.

As a specific example, the loss function submodule of the visual detection module uses a composite loss function L _total to simultaneously optimize the position, size, object confidence, and category probability of the bounding box. The calculation formula of the loss function L _total is:

L _total = L _coord + L _conf + L _class

Where L _coord is the coordinate loss, using the squared error loss to measure the difference between the predicted bounding box and the true bounding box; L _conf is the confidence loss, using MSE or cross entropy loss to evaluate the accuracy of confidence prediction; L _class is the classification loss, using cross entropy loss to evaluate the prediction accuracy of class probability;

The cross entropy loss function L is:

Where M is the number of categories, yo _,c is the observed value, and _po,c is the predicted probability.

As a specific example, the control module uses PLC as the main control unit of the system to control the rotation of the automatic rotating material tray, the operation of the loading robot arm, the movement of the conveyor belt and the operation of the unloading robot arm, and uses an industrial PC to run a lightweight deep learning model to process images taken by the image acquisition module, perform defect detection, and feed back the results to the PLC.

As a specific example, the physical protection monitors the environment around the loading robot arm and the unloading robot arm in real time through safety sensors, sets safety limits for the loading robot arm and the unloading robot arm, and controls the movement range of the loading robot arm and the unloading robot arm not to exceed a preset range;

The emergency stop button is used to stop the movement of the loading and unloading robotic arms in an emergency;

The safety sensor is used to monitor the status parameters of the loading robot arm and the unloading robot arm, and to issue an alarm and trigger an emergency stop button in case of abnormal conditions.

The present invention also provides an automated bearing defect detection method based on the automated bearing defect detection system, comprising the following steps:

Step 1: The parts supply module uses an automatic rotating tray to automatically load the parts to be inspected;

Step 2: The parts conveying module uses a conveyor belt and a loading robot to transport the bearing to the visual inspection area;

Step 3: The lighting and image acquisition module takes multi-angle photos of the bearing and sends them to the visual inspection module;

Step 4: The visual inspection module uses the inspection algorithm and edge computing device to detect whether the bearing parts are defective and sends the inspection results to the control module;

Step 5: The control module classifies the tested bearings, transports the defective bearings to the scrap area, and transports the non-defective bearings to the unloading area;

Step 6: The safety module protects the safety of personnel and equipment during the automated testing process, and promptly detects and responds to potential dangerous situations.

A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the automatic bearing defect detection method when executing the program.

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Example

In this embodiment, a mechanical structure diagram of an automated bearing defect detection system is provided. As shown in FIG2 , the mechanical structure of the automated bearing defect detection system includes a feeding robot arm 1, an automatic assembly line 2, a visual tray 3, a control panel 4, a bottom camera device 5, a gantry bracket 6, a top camera 7, a bottom camera 8, a side camera 9, and a feeding robot arm 10, wherein:

The unloading robot arm 1 is responsible for sorting and clamping the defective bearing parts. After completion, it sends a signal to the control module to carry out the inspection of the next batch.

Automatic assembly line 2 is responsible for the transportation of parts. It is controlled by the control module and cooperates with two robot arms and the visual inspection module to complete the entire inspection system.

The number and size of the visual tray 3 are set according to the number of specific parts. It is made of polycarbonate PC with high transparency. The PC material not only has high transparency, but also has good impact resistance and temperature resistance. It is suitable for use in different environments, and has high transparency so that the camera below the detection area can collect images.

The control panel 4 is an interface for the user to interact with the part defect detection system, allowing the operator to set detection parameters, start and stop the detection process, monitor the system status, and view alarms and notifications.

The bottom camera device 5 supports the bottom camera and includes an intelligent searchlight to adjust the brightness according to the indoor light.

The gantry bracket 6 is a supporting component, used to support the camera and the control panel, etc.

The top camera 7 is a key hardware component responsible for capturing image data of parts on the production line, which are used to identify any possible defects through visual algorithms. The choice of camera is crucial to ensure the accuracy and efficiency of the inspection system; the camera needs to be compatible with the defect detection algorithm so that the model can accurately analyze the captured images and contain intelligent searchlights that adjust the brightness according to the room light and shoot the parts from above.

The bottom camera 8 is an upward shooting part, which contains an intelligent searchlight and adjusts the brightness according to the indoor light.

There are two side cameras 9, which are responsible for shooting from the left and right perspectives, and include intelligent searchlights to adjust the brightness according to the indoor light.

The loading robot arm 10 is responsible for grabbing parts from the material tray and placing them neatly on the visible tray of the conveyor belt in a certain arrangement. After a batch of parts is clamped, the conveyor belt starts to operate.

As shown in FIG3 , the mechanical structure of the automatic rotating material tray includes a guide rod 11, a part mounting plate 12, a stepper motor 13, and a lifting mechanism 14. The stepper motor 13 is used to accurately control the stepping of the motor, and the rotation angle of the material tray can be accurately controlled. It is suitable for scenes that require precise position control and cooperates with the feeding robot arm to complete the feeding task; the upper part mounting plate 12 is equipped with three detachable guide rods 11, which are used to fix the bearing parts to prevent the parts from sliding during the rotation process to complete the precise clamping task; three guide rods 11 are set to adapt to bearing parts of different sizes, and the number of guide rods can be selected according to actual conditions; a lifting mechanism 14 is set below, and the lifting mechanism 14 can lift the material tray from a certain height to another certain height, so as to facilitate docking with other equipment or work surfaces, so that the parts can smoothly transition between different processing stages. By using the lifting mechanism 14, the need for manual handling of parts or products can be reduced, thereby reducing labor intensity and improving production safety and efficiency.

FIG4 is a detection algorithm structure of a visual detection module, including a model training module, a feature extraction module, a defect detection module, a feature fusion module, and a detection head module;

The model training module includes a model selection module, a model initialization module, a loss function module, an optimizer module, a batch processing module, a regularization module, a model evaluation module, and a model saving module;

The feature extraction module includes a convolution layer module, a pooling layer module, a normalization layer module, an activation layer module, a feature selection module and a feature dimension reduction module;

The defect detection module includes an image preprocessing module, a feature extraction module, a defect detection module, a defect location module, and a visualization module;

The detection algorithm flow of the visual inspection module is:

(1) The image information of the bearing parts collected by the high-definition camera device of the image acquisition module is stored in the storage module, and the image of the bearing parts is transmitted to the image preprocessing module in the defect detection module. The image denoising module in the image preprocessing module denoises the collected bearing images through a filtering algorithm to reduce the errors of feature extraction and model training;

(2) The convolution layer module in the feature extraction module uses a convolutional neural network deep learning model to perform a convolution operation on the preprocessed bearing part image information to extract the bearing contour and texture required for bearing part defect detection;

(3) The model selection module in the model training module selects a specified convolutional neural network model and uses the preprocessed image data and extracted features to train a deep learning model with good performance;

(4) The model evaluation module in the model training module evaluates the performance of the trained deep learning model. The evaluation indicators include calculation accuracy, precision and recall rate;

(5) The visualization module in the defect detection module is electrically connected to the external control panel device to visualize the defect information and generate a bearing part defect information report, which includes the number of bearing part defects, defect locations and bearing defect types.

Furthermore, the deep learning model is lightweight and improved based on the YOLOv8 algorithm, uses the self-developed LACM module to complete the downsampling operation in the backbone network, and designs adaptive selection weights and group convolutions to improve the accuracy while reducing the parameters of the model.

The self-developed EMC module is used as the bottleneck layer in the backbone network, and the bottleneck structure is rebuilt using the idea of group convolution, which further improves the detection accuracy and reduces the number of model parameters.

Knowledge distillation is performed on the improved model, and the detection accuracy is further improved using the teacher model.

In summary, the lightweight network improved based on the deep learning model YOLOv8 not only reduces the number of model parameters, but also improves the detection accuracy, making it easier to use in edge devices to complete real-time detection tasks; the detection results can be visualized for staff, which is more intuitive, reduces manual intervention, and improves detection efficiency, thereby completing automated bearing defect detection.

It should be emphasized that the above examples are only for explaining the technical content of the present invention, not for limiting the same. Even though the present invention has been specifically explained in this article through preferred examples, professionals in the technical field should recognize that the technical solution of the present invention can be adjusted or replaced by equivalents, and these changes should not deviate from the core concept of the present invention and its scope of implementation, and all these adjustments and replacements should be regarded as the scope of protection of the present invention.

Claims (10)

1. An automated bearing defect detection system, characterized in that it includes a parts supply module, a parts conveying module, a lighting and image acquisition module, a visual inspection module, a control module and a safety module; The parts supply module uses an automatically rotating material tray to supply the parts to be inspected; The parts conveying module includes a conveyor belt, a loading robot arm, a unloading robot arm and an operating unit, which are used to convey the parts to be inspected to the visual inspection area; The lighting and image acquisition module includes a fill light source and an image acquisition module, which are used to capture the details of the parts to be inspected to meet the complex visual recognition requirements; The visual inspection module includes a detection algorithm model and an edge computing device for detecting whether the bearing parts have defects; The control module includes a main control unit and a control panel, which are used to monitor and control the coordinated operation between various modules, receive and process the detection data of the visual detection module, and make corresponding control decisions; The safety module includes physical protection, emergency stop buttons and safety sensors to protect the safety of personnel and equipment during the detection process and to detect and respond to potential dangerous situations. 2. The automated bearing defect detection system according to claim 1 is characterized in that the automatic rotating tray of the parts supply module cooperates with the mechanical arm of the parts conveying module to clamp the parts, and after clamping a part, the automatic rotating tray rotates once to complete the automated loading process; The automatic rotating material tray includes guide rods, a parts mounting plate, a stepper motor and a lifting mechanism arranged from top to bottom. The stepper motor is used to control the rotation angle of the material tray, and the loading task is completed in cooperation with the loading robot arm; the parts mounting plate is equipped with a detachable guide rod for fixing bearing parts to prevent the parts from sliding during rotation to complete the clamping task; the guide rods are adapted to bearing parts of different sizes, and the number of guide rods is selected according to actual conditions; the lifting mechanism adjusts the height of the material tray for docking with the work surface. 3. The automated bearing defect detection system according to claim 1 is characterized in that the operating unit of the parts conveying module controls the loading robot arm to place the parts at the specified position of the conveyor belt after clamping them, and arranges a batch of parts in a layout of 3 rows and 4 columns; the operating unit is provided with a coordinate system to guide the loading robot arm to accurately place the parts at the specified position on the conveyor belt; the unloading robot arm is controlled by the signal of the operating unit to unload the defective parts; the operating unit controls the speed of the conveyor belt to ensure that the parts are stably transmitted to the detection area. 4. The automated bearing defect detection system according to claim 1 is characterized in that the fill light source of the lighting and image acquisition module adopts a multi-angle light source to reduce shadows and reflections, and the lighting intensity is adjusted according to the material and color of the part to obtain the best image quality; the image acquisition module uses a high-resolution camera to capture the tiny details of the part surface, adopts multi-angle shooting, and comprehensively detects every surface of the part. 5. The automated bearing defect detection system according to claim 1, characterized in that the detection algorithm model of the visual inspection module comprises a model training unit, a feature extraction unit, a defect detection unit, a feature fusion unit and a detection head unit; The model training unit includes a model selection submodule, a model initialization submodule, a loss function submodule, an optimizer submodule, a batch processing submodule, a regularization submodule, a model evaluation submodule, and a model saving submodule; The model selection submodule is used to select a machine learning or deep learning model structure that is suitable for the problem; the model initialization submodule is used to initialize the weights and biases of the model; the loss function submodule is used to optimize the model's loss function, which is a function used to measure the difference between the model output and the actual label; the optimizer submodule selects and configures the optimization algorithm used to optimize the model; the batch processing submodule divides the data into small batches during training to calculate gradients and update model parameters; the regularization submodule adds regularization terms to prevent the model from overfitting; the model evaluation submodule is used to evaluate the performance of the model on the training set, validation set, or test set; the model saving submodule saves the trained model weights for future use; The feature extraction unit includes a convolution layer submodule, a pooling layer submodule, a normalization layer submodule, an activation layer submodule, a feature selection submodule and a feature dimension reduction submodule; The convolution layer submodule applies convolution operation to extract image features; the pooling layer submodule reduces the spatial dimension of the feature map through pooling operation; the normalization layer submodule normalizes the features; the activation layer submodule applies activation function to introduce nonlinearity and increase the expressive power of the model; the feature selection submodule selects the most representative features for subsequent defect detection; the feature dimensionality reduction submodule reduces the dimension of the features; The defect detection unit includes an image preprocessing submodule, a feature extraction submodule, a defect detection submodule, a defect location submodule and a visualization submodule; The image preprocessing submodule preprocesses the input image, including resizing, cropping or enhancing the contrast of the image; the feature extraction submodule extracts features from the preprocessed image; the defect detection submodule uses the extracted features to detect defects in the image, including using classifiers or regressors; the defect localization submodule locates and marks the locations of the detected defects in the image; the visualization submodule visualizes the detected defect information for users to understand and analyze the detection results; The detection process of the detection algorithm model is as follows: (1) The image acquisition module acquires image information of the bearing parts and transmits the image information of the bearing parts to the image preprocessing submodule in the visual inspection module. The image preprocessing submodule performs denoising on the acquired image information of the bearing parts through a filtering algorithm to reduce the error of feature extraction and model training; (2) The convolutional layer submodule in the feature extraction unit uses a lightweight model based on the improved YOLOv8 to perform convolution operations on the image information of the preprocessed bearing parts; in the backbone network of YOLOv8, the convolution layer of the downsampling operation is replaced by a lightweight adaptive convolution module LACM. The structure of LACM is divided into two paths, one path obtains an adaptive weight, and the other path uses group convolution to reduce the number of parameters, and finally multiplies and adds the obtained adaptive weights to obtain focused feature parameters; in the backbone network of YOLOv8, the Bottleneck layer used by C2f is replaced by a multi-scale convolution module EMC. The structure of EMC divides the input features into two parts, one part passes through 3×3Conv and 5×5Conv, and the other part retains the original features without operation; finally, the two parts of the features are combined to perform 1×1Conv; the bearing contour and texture required for bearing part defect detection are extracted; (3) The model selection submodule in the model training unit selects the specified lightweight model based on the YOLOv8 improvement, and uses the preprocessed image data and extracted features to train the deep learning model; finally, YOLOv8m is used as the teacher model to perform knowledge distillation on the lightweight model based on the YOLOv8 improvement; (4) The model evaluation submodule in the model training unit evaluates the performance of the trained deep learning model. The evaluation indicators include calculation accuracy, precision, and recall rate. (5) The visualization submodule in the defect detection unit is electrically connected to the control panel of the control module to visualize the defect information and generate a bearing part defect information report, which includes the number of bearing part defects, defect locations and bearing defect types. 6. The automated bearing defect detection system according to claim 1, characterized in that the loss function submodule of the visual inspection module uses a composite loss function L _total to simultaneously optimize the position, size, object confidence and category probability of the bounding box, and the calculation formula of the loss function L _total is: L _total = L _coord + L _conf + L _class Where L _coord is the coordinate loss, using the squared error loss to measure the difference between the predicted bounding box and the true bounding box; L _conf is the confidence loss, using MSE or cross entropy loss to evaluate the accuracy of confidence prediction; L _class is the classification loss, using cross entropy loss to evaluate the prediction accuracy of class probability; The cross entropy loss function L is: Where M is the number of categories, yo _,c is the observed value, and _po,c is the predicted probability. 7. The automated bearing defect detection system according to claim 1 is characterized in that the control module uses PLC as the main control unit of the system to control the rotation of the automatic rotating material tray, the operation of the loading robot arm, the movement of the conveyor belt and the operation of the unloading robot arm, and uses an industrial PC to run a lightweight deep learning model to process images taken by the image acquisition module, perform defect detection, and feed back the results to the PLC. 8. The automated bearing defect detection system according to claim 1 is characterized in that the physical protection monitors the environment around the loading robot arm and the unloading robot arm in real time through safety sensors, sets safety limits for the loading robot arm and the unloading robot arm, and controls the movement range of the loading robot arm and the unloading robot arm not to exceed a preset range; The emergency stop button is used to stop the movement of the loading and unloading robotic arms in an emergency; The safety sensor is used to monitor the status parameters of the loading robot arm and the unloading robot arm, and to issue an alarm and trigger an emergency stop button in case of abnormal conditions. 9. An automated bearing defect detection method based on the automated bearing defect detection system according to any one of claims 1 to 8, characterized in that it comprises the following steps: Step 1: The parts supply module uses an automatic rotating tray to automatically load the parts to be inspected; Step 2: The parts conveying module uses a conveyor belt and a loading robot to transport the bearing to the visual inspection area; Step 3: The lighting and image acquisition module takes multi-angle photos of the bearing and sends them to the visual inspection module; Step 4: The visual inspection module uses the inspection algorithm and edge computing device to detect whether the bearing parts are defective and sends the inspection results to the control module; Step 5: The control module classifies the tested bearings, transports the defective bearings to the scrap area, and transports the non-defective bearings to the unloading area; Step 6: The safety module protects the safety of personnel and equipment during the automated testing process, and promptly detects and responds to potential dangerous situations. 10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the automated bearing defect detection method according to claim 9 when executing the program.