Tool blade grinding condition detection system based on deep learning

Abstract

The invention discloses a tool blade grinding completion condition detection system based on deep learning, and relates to an intelligent detection technology, the tool blade grinding completion condition detection system comprises an image acquisition module which at least comprises a tool blade camera device, an automatic clamp device, an image preprocessing device and an image marking device; the database module is mainly used for storing the detection information, the grinding completion information and the abnormal information of the detected cutting edge of the cutter; the feature learning module is used for training and recognizing a grinding completion state and an abnormal condition; the detection module is used for detecting whether the cutting edge of the cutter is in a completely-ground state or abnormal conditions; a background module; a management device module; and a user terminal module. The method has the advantages that the complete grinding state of the cutting edge of the cutter can be monitored in real time, the problem of excessive complete grinding or insufficient complete grinding can be intelligently recognized, and through a novel network structure, an attention mechanism is added, so that the extraction capability of complete grinding features of the cutter is improved.

Description

Technical Field

[0001] The present invention relates to intelligent detection technology, and in particular to a tool blade grinding condition detection system based on deep learning. Background Art

[0002] The tool blade grinding condition detection system based on deep learning is a key technology in the field of modern intelligent manufacturing. With the rapid development of industrial automation and intelligence, the requirements for processing accuracy and efficiency are getting higher and higher. As a key tool in the manufacturing process, the grinding status of the tool directly affects the product quality and production efficiency. Traditional grinding detection methods have problems such as low detection accuracy and poor robustness, which are difficult to meet the needs of real-time online detection. The development of deep learning technology provides a new solution for tool grinding detection. By building an intelligent model, it can accurately identify the grinding status of the tool, realize real-time detection and prediction, thereby improving production efficiency, reducing costs, and ensuring product quality. It is of great significance to enhance the overall competitiveness of the manufacturing industry.

[0003] At present, the method for detecting the grinding condition of tool blades is generally an online detection method. Although it can provide detection capabilities in specific scenarios, it is limited in comprehensively detecting abnormal conditions that may occur after tool grinding. Existing systems often focus on a single grinding condition indicator and lack in-depth analysis of the overall performance of the tool. In order to solve this problem, it is necessary to develop more advanced detection technologies that can capture other abnormal conditions caused by tool blade processing in real time, such as possible chipping and fracture, and integrate this information to more comprehensively evaluate the grinding condition of the tool. Summary of the invention

[0004] In view of some deficiencies in the prior art, the present invention proposes an innovative tool blade grinding condition detection system based on deep learning. The system aims to accurately identify the over-grinding or under-grinding problems that may occur in the tool blade during the production process, and detect any potential blade abnormalities. By real-time monitoring of the blade grinding condition. This not only effectively solves the non-human factors that may occur before and after the tool blade is ground, but also has the ability to predict unknown situations, thereby preventing potential production risks in advance.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A tool blade grinding condition detection system based on deep learning, characterized by comprising:

[0007] Image acquisition module: The image acquisition module at least includes a tool blade camera device, an automated fixture device, an image preprocessing device, and an image annotation device. The tool blade camera device is used to capture images of the tool blade. The automated fixture device is used to clamp the tool in front of the camera for acquisition and screen out unqualified tools for processing. The image preprocessing device is used to process the image so that the blade has integrity and continuity. The image annotation device is used to mark the area that affects the system's judgment of whether the tool is fully ground.

[0008] Database module: The database module is mainly used to store the detection information, grinding information and abnormal information of the detected tool blade;

[0009] Feature learning module: The feature learning module is used to train and identify the grinding state and abnormal conditions;

[0010] Detection module: The detection module is electrically connected to the image acquisition module, the database module, and the feature learning module, and is used to detect whether the tool blade is fully ground or in an abnormal state;

[0011] Backend module: The backend module is electrically connected to the detection module, database module, and feature learning module. The backend module is used to obtain detection image data and record it in the database. At the same time, the collected data is regularly included in the feature learning module. When an abnormality is detected, an alarm is sent to the user terminal in a timely manner. ;

[0012] Management equipment module: The management equipment module is used to manage all the information of the tools in the database, and display the information of the tools in the form of web pages through the database, including the grinding status and abnormal conditions of the tools, and has the authority to update the database information.

[0013] User terminal module: The user terminal module includes mobile phone applications, tablet computers and computer applications, which are only granted the authority to browse the tool information in the database and receive abnormal alarm notifications from the background module.

[0014] Preferably, the image acquisition module specifically includes:

[0015] A tool blade camera device, wherein the tool blade camera device is three industrial cameras that shoot from the left, right and top directions of the tool to obtain the tool blade image;

[0016] An image preprocessing device, wherein the image preprocessing device eliminates image noise by using Gaussian filtering technology, uses Canny algorithm to detect edges to highlight the contour of the tool blade, enhances image contrast by histogram equalization, simplifies the image to present the tool shape by binarization operation, and uses morphological operations to eliminate small noise points and fill the gaps on the tool edge;

[0017] An image annotation device is used to identify and annotate key features of a blade, such as sharpness, notches, cracks, etc.

[0018] Preferably, the feature learning module specifically includes:

[0019] Through forward propagation, the input image is calculated layer by layer through the deep learning model to obtain the prediction result;

[0020] The gap between the predicted result and the true label is measured by calculating the loss;

[0021] The gradient of loss to each layer parameter is calculated through back propagation to guide the adjustment of model parameters;

[0022] Use the method of optimizing parameters to update weights and reduce the loss function value;

[0023] Perform performance verification on the validation set to evaluate the effect of the model and avoid overfitting;

[0024] The trained model is persisted through the model saving step.

[0025] Preferably, the detection module specifically includes:

[0026] Grinding state detection network module: The grinding state detection network module includes a multi-scale feature fusion unit, an attention mechanism unit, a deep feature fusion unit, and a grinding state classification unit;

[0027] Anomaly detection network module: The anomaly detection network module includes a multi-scale feature extraction unit, a feature pyramid unit, an attention mechanism unit, a deep feature fusion unit, and anomaly detection unit.

[0028] Preferably, the grinding completion status detection network module specifically includes:

[0029] Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps;

[0030] Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN;

[0031] By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear.

[0032] Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information;

[0033] Based on the feature map extracted, fused and optimized above, the global average pooling layer is input to obtain a vector containing all the key information of the image, and then input into the classification head to classify the tool according to its wear state, including (complete grinding / uncomplete grinding).

[0034] For the binary classification problem (completed / uncompleted), the loss function used is the binary cross entropy loss function, and the function formula is:

[0035]

[0036] Among them, y is the true label, and its value is usually 0 (not finished) or 1 (finished). is the predicted probability, which is the probability that the sample is normal. In practical applications, the output is usually generated by an activation function, because the Sigmoid function can map any real number to the (0,1) interval to represent the probability. The definition of the Sigmoid function is as follows:

[0037]

[0038] Preferably, the anomaly detection network module specifically includes:

[0039] Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps;

[0040] Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN;

[0041] By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear.

[0042] Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information;

[0043] The fused high-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer to obtain a higher-level nonlinear relationship. The output of the fully connected layer is compressed to between 0 and 1 through the Sigmoid function to determine the probability that the tool is in an abnormal state.

[0044] Based on the abnormal state probability of the tool blade, if it is greater than 0.5, it is abnormal and an abnormal alarm notification will be sent to the user terminal. If it is less than 0.5, it is normal.

[0045] Compared with the prior art, the advantages of the present invention are:

[0046] The present invention adopts deep learning technology, which can monitor the grinding status of the tool blade in real time and intelligently identify the problem of over-grinding or under-grinding, which is more accurate and efficient than the traditional method relying on visual observation. Through real-time monitoring and intelligent identification, the system can quickly discover and adjust problems in the grinding process, significantly reduce rework and scrap rates, and thus greatly improve production efficiency. The deep learning model of the present invention uses a novel network structure and adds an attention mechanism to improve the ability to extract tool grinding features, reduce dependence on prior knowledge, and enhance the model's adaptability to new data. The system has the ability to predict unknown situations. The present invention not only realizes real-time detection of the tool grinding status, but also further realizes the prediction of the future grinding trend of the tool, providing a scientific basis for arranging maintenance and replacement of tools in advance. The automated detection system reduces the reliance on manual detection, reduces labor costs, and also reduces detection errors caused by human factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is an overall system schematic diagram of the method and system proposed by the present invention;

[0048] Figure 2 It is a schematic diagram of the overall flow of the method and system proposed in the present invention;

[0049] Figure 3 It is a structural block diagram of a tool blade grinding detection system of the method and system proposed in the present invention;

[0050] Figure 4 A structural block diagram of a detection module of the method and system proposed in the present invention;

[0051] Figure 5 A network flow chart of tool grinding state detection of the method and system proposed by the present invention;

[0052] Figure 6 This is a network flow chart of tool anomaly detection of the method and system proposed in the present invention. DETAILED DESCRIPTION

[0053] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0054] Reference Figure 3 As shown, a tool blade grinding condition detection system based on deep learning includes:

[0055] Image acquisition module: The image acquisition module at least includes a tool blade camera device, an automated fixture device, an image preprocessing device, and an image annotation device. The tool blade camera device is used to capture images of the tool blade. The automated fixture device is used to clamp the tool in front of the camera for acquisition and screen out unqualified tools for processing. The image preprocessing device is used to process the image so that the blade has integrity and continuity. The image annotation device is used to mark the area that affects the system's judgment of whether the tool is fully ground.

[0056] Database module: The database module is mainly used to store the detection information, grinding information and abnormal information of the detected tool blade;

[0057] Feature learning module: The feature learning module is used to train and identify the grinding state and abnormal conditions;

[0058] Detection module: The detection module is electrically connected to the image acquisition module, the database module, and the feature learning module, and is used to detect whether the tool blade is fully ground or in an abnormal state;

[0059] Backend module: The backend module is electrically connected to the detection module, database module, and feature learning module. The backend module is used to obtain detection image data and record it in the database. At the same time, the collected data is regularly included in the feature learning module. When an abnormality is detected, an alarm is sent to the user terminal in a timely manner. ;

[0060] Management equipment module: The management equipment module is used to manage all the information of the tools in the database, and display the information of the tools in the form of web pages through the database, including the grinding status and abnormal conditions of the tools, and has the authority to update the database information.

[0061] User terminal module: The user terminal module includes mobile phone applications, tablet computers and computer applications, which are only granted the authority to browse the tool information in the database and receive abnormal alarm notifications from the background module.

[0062] Reference Figure 3 As shown, the image acquisition module specifically includes:

[0063] A tool blade camera device, wherein the tool blade camera device is three industrial cameras that shoot from the left, right and top directions of the tool to obtain the tool blade image;

[0064] An image preprocessing device, wherein the image preprocessing device eliminates image noise by using Gaussian filtering technology, uses Canny algorithm to detect edges to highlight the contour of the tool blade, enhances image contrast by histogram equalization, simplifies the image to present the tool shape by binarization operation, and uses morphological operations to eliminate small noise points and fill the gaps on the tool edge;

[0065] An image annotation device is used to identify and annotate key features of a blade, such as sharpness, notches, cracks, etc.

[0066] The automated fixture device specifically includes:

[0067] On the tool production line, when the tool to be inspected moves smoothly to the inspection area via the conveyor belt, the automated fixture system above responds immediately. First, the sensor on the support frame detects the position of the tool and then triggers the control signal. The drive motor is then started, driving the cam to rotate, and the rotation of the cam causes the connected moving rod to drop rapidly in a direction perpendicular to the conveyor belt. The lower end of the moving rod is connected to a special pressure block, which is designed according to the shape of the tool. It can be a flat shape with a groove, or a customized shape that adapts to the contour of the tool. When the pressure block contacts the tool, it clamps the tool firmly to ensure that it is completely motionless during the subsequent quality inspection process. At this time, three industrial cameras capture images of the tool blade in three directions: left, right, and top. Once the acquisition is completed, the cam mechanism rotates in the opposite direction, the moving rod rises, and the pressure block releases the tool. The tool is then smoothly transported to the next screening link or directly into the finished product area, and the entire clamping and release process does not require human intervention.

[0068] The feature learning module specifically includes:

[0069] Through forward propagation, the input image is calculated layer by layer through the deep learning model to obtain the prediction result;

[0070] The gap between the predicted result and the true label is measured by calculating the loss;

[0071] The gradient of loss to each layer parameter is calculated through back propagation to guide the adjustment of model parameters;

[0072] Use the method of optimizing parameters to update weights and reduce the loss function value;

[0073] Perform performance verification on the validation set to evaluate the effect of the model and avoid overfitting;

[0074] The trained model is persisted through the model saving step.

[0075] Reference Figure 4 As shown, the detection module specifically includes:

[0076] Grinding state detection network module: The grinding state detection network module includes a multi-scale feature fusion unit, an attention mechanism unit, a deep feature fusion unit, and a grinding state classification unit;

[0077] Anomaly detection network module: The anomaly detection network module includes a multi-scale feature extraction unit, a feature pyramid unit, an attention mechanism unit, a deep feature fusion unit, and anomaly detection unit.

[0078] Reference Figure 5 As shown, the grinding state detection network module specifically includes:

[0079] Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps;

[0080] Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN;

[0081] By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear.

[0082] Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information;

[0083] Based on the feature map extracted, fused and optimized above, the global average pooling layer is input to obtain a vector containing all the key information of the image, and then input into the classification head to classify the tool according to its wear state, including (complete grinding / uncomplete grinding).

[0084] For the binary classification problem (completed / uncompleted), the loss function used is the binary cross entropy loss function, and the function formula is:

[0085]

[0086] Among them, y is the true label, and its value is usually 0 (not finished) or 1 (finished). is the predicted probability, which is the probability that the sample is normal. In practical applications, the output is usually generated by an activation function, because the Sigmoid function can map any real number to the (0,1) interval to represent the probability. The definition of the Sigmoid function is as follows:

[0087]

[0088] Reference Figure 6 As shown in the figure, the anomaly detection network module specifically includes:

[0089] Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps;

[0090] Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN;

[0091] By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear.

[0092] Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information;

[0093] The fused high-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer to obtain a higher-level nonlinear relationship. The output of the fully connected layer is compressed to between 0 and 1 through the Sigmoid function to determine the probability that the tool is in an abnormal state.

[0094] Based on the abnormal state probability of the tool blade, if it is greater than 0.5, it is abnormal and an abnormal alarm notification will be sent to the user terminal. If it is less than 0.5, it is normal.

[0095] The feature learning module specifically includes:

[0096] After completing the image analysis of the tool blade, a repeated training process is performed to enhance the recognition and prediction capabilities of the model by integrating the analyzed image data into the existing training set. First, the images of the tool blade complete grinding status and anomaly detection are added to the training data set. Since they have been labeled before analysis, they can be directly used for grinding detection and anomaly detection model learning. Subsequently, a new round of training process is started using the updated data set. The model adjusts internal parameters through back propagation and gradient descent algorithms to better adapt to the new data features. After the training is completed, verification and testing are carried out to evaluate the improvement of model performance. If the performance meets expectations, the updated model will be deployed for actual blade image analysis tasks, completing the entire closed-loop process of repeated training.

[0097] In summary, the advantages of the present invention are: the present invention adopts deep learning technology, which can monitor the grinding status of the tool blade in real time and intelligently identify the problem of over-grinding or under-grinding, which is more accurate and efficient than the traditional method relying on naked eye observation. Through real-time monitoring and intelligent identification, the system can quickly discover and adjust problems in the grinding process, significantly reduce rework and scrap rates, and thus greatly improve production efficiency. The deep learning model of the present invention uses a novel network structure and adds an attention mechanism to improve the ability to extract tool grinding features, reduce dependence on prior knowledge, and enhance the model's adaptability to new data. The system has the ability to predict unknown situations. The present invention not only realizes real-time detection of the tool grinding status, but also further realizes the prediction of the future grinding trend of the tool, providing a scientific basis for arranging maintenance and replacement of tools in advance. The automated detection system reduces the reliance on manual detection, reduces labor costs, and also reduces detection errors caused by human factors.

[0098] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A tool blade grinding condition detection system based on deep learning, characterized in that: include: Image acquisition module: The image acquisition module at least includes a tool blade camera device, an automated fixture device, an image preprocessing device, and an image annotation device. The tool blade camera device is used to capture images of the tool blade. The automated fixture device is used to clamp the tool in front of the camera for acquisition and screen out unqualified tools for processing. The image preprocessing device is used to process the image so that the blade has integrity and continuity. The image annotation device is used to mark the area that affects the system's judgment of whether the tool is fully ground. Database module: The database module is mainly used to store the detection information, grinding information and abnormal information of the detected tool blade; Feature learning module: The feature learning module is used to train and identify the grinding state and abnormal conditions; Detection module: The detection module is electrically connected to the image acquisition module, the database module, and the feature learning module, and is used to detect whether the tool blade is fully ground or in an abnormal state; Backend module: The backend module is electrically connected to the detection module, database module, and feature learning module. The backend module is used to obtain detection image data and record it in the database. At the same time, the collected data is regularly included in the feature learning module. When an abnormality is detected, an alarm is sent to the user terminal in a timely manner. ; Management equipment module: The management equipment module is used to manage all the information of the tools in the database, and display the information of the tools in the form of web pages through the database, including the grinding status and abnormal conditions of the tools, and has the authority to update the database information. User terminal module: The user terminal module includes mobile phone applications, tablet computers and computer applications, which are only granted the authority to browse the tool information in the database and receive abnormal alarm notifications from the background module.

2. The tool blade grinding condition detection system based on deep learning according to claim 1, characterized in that: The image acquisition module specifically includes: A tool blade camera device, wherein the tool blade camera device is three industrial cameras that shoot from the left, right and top directions of the tool to obtain the tool blade image; An image preprocessing device, wherein the image preprocessing device eliminates image noise by using Gaussian filtering technology, uses Canny algorithm to detect edges to highlight the contour of the tool blade, enhances image contrast by histogram equalization, simplifies the image to present the tool shape by binarization operation, and uses morphological operations to eliminate small noise points and fill the gaps on the tool edge; An image annotation device is used to identify and annotate key features of a blade, such as sharpness, notches, cracks, etc.

3. The tool blade grinding condition detection system based on deep learning according to claim 1, characterized in that: The feature learning module specifically includes: Through forward propagation, the input image is calculated layer by layer through the deep learning model to obtain the prediction result; The gap between the predicted result and the true label is measured by calculating the loss; The gradient of loss to each layer parameter is calculated through back propagation to guide the adjustment of model parameters; Use the method of optimizing parameters to update weights and reduce the loss function value; Perform performance verification on the validation set to evaluate the effect of the model and avoid overfitting; The trained model is persisted through the model saving step.

4. The tool blade grinding condition detection system based on deep learning according to claim 1, characterized in that: The detection module specifically includes: Grinding state detection network module: The grinding state detection network module includes a multi-scale feature fusion unit, an attention mechanism unit, a deep feature fusion unit, and a grinding state classification unit; Anomaly detection network module: The anomaly detection network module includes a multi-scale feature extraction unit, a feature pyramid unit, an attention mechanism unit, a deep feature fusion unit, and anomaly detection unit.

5. The tool blade grinding condition detection system based on deep learning according to claim 3 is characterized in that: The grinding state detection network module specifically includes: Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps; Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN; By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear. Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information; Based on the feature map extracted, fused and optimized above, the global average pooling layer is input to obtain a vector containing all the key information of the image, and then input into the classification head to classify the tool according to its wear state, including (complete grinding / uncomplete grinding). For the binary classification problem (completed / uncompleted), the loss function used is the binary cross entropy loss function, and the function formula is: Among them, y is the true label, and its value is usually 0 (not finished) or 1 (finished). is the predicted probability, which is the probability that the sample is normal. In practical applications, the output is usually generated by an activation function, because the Sigmoid function can map any real number to the (0,1) interval to represent the probability. The definition of the Sigmoid function is as follows:

6. The tool blade grinding condition detection system based on deep learning according to claim 3 is characterized in that: The anomaly detection network module specifically includes: Preprocess and extract features of the acquired image data based on the pre-trained ResNet-50, including normalization, cropping, rotation and other steps; Based on the different characteristics of tool wear at different scales, multi-scale feature fusion is performed through multi-scale feature extraction of FPN; By introducing the self-attention mechanism to enhance the expression of key features, the model can further focus on the key areas of blade wear. Use feature fusion based on ResNeXt module to further improve the representation ability of the network and integrate multi-scale and multi-level feature information; The fused high-dimensional feature map is flattened into a one-dimensional vector and input into the fully connected layer to obtain a higher-level nonlinear relationship. The output of the fully connected layer is compressed to between 0 and 1 through the Sigmoid function to determine the probability that the tool is in an abnormal state. Based on the abnormal state probability of the tool blade, if it is greater than 0.5, it is abnormal and an abnormal alarm notification will be sent to the user terminal. If it is less than 0.5, it is normal.

Images