Visual online detection method for grinding damage of side edge of end mill

Abstract

The invention discloses a visual online detection method for grinding damage of a side edge of an end mill. The method comprises the following steps: step 1, acquiring a sample image; step 2, self-compensating a shooting angle; step 3, establishing a VGG16-Unet side blade wear image recognition model; 4, collecting a detection image; 5, obtaining a milling cutter abrasion result; and 6, evaluating the milling cutter wear detection precision in the step 5. The shooting angle self-compensation method based on the least square is developed according to the situation that the side blade abrasion loss changes along with the shooting angle, and the problem that image collection is incomplete due to the side blade radian and the camera angle during abrasion contour extraction is solved. A deep learning model based on VGG16-UNet is established, image features of different scales are extracted for feature fusion, image splicing and matching are omitted, the problem of matching and positioning of wear regions between adjacent wear images is solved, the model has high robustness and high accuracy, the problems of wrong division and missing division of the wear regions are greatly improved, and the accuracy of the wear regions is improved. And the detection precision and efficiency are greatly improved.

Description

Technical Field

[0001] The invention relates to a visual online detection method for side grinding damage of an end mill, belonging to the technical field of image analysis and processing. Background Art

[0002] As a typical high-end CNC tool, the integral carbide spiral end mill plays an important role in the processing of parts in the fields of aerospace, automobile manufacturing, mold production, medical equipment, etc. According to statistics, more than 20% of machine tool downtime is caused by tool wear and breakage. Tool status detection is a complex system engineering. On the one hand, various faults will inevitably occur during the use of the tool. If these problems are not discovered and solved in time, the tool life will be shortened or even completely scrapped, which will bring great hidden dangers to the success rate of material processing; if the tool fails and cannot be eliminated in time, it will not only affect the processing quality and precision, but also cause major casualties, and even cause downtime for maintenance, which seriously restricts the production progress.

[0003] In traditional tool status detection, production personnel rely on the experience of observing vibration, noise and processing time during the processing to judge the degree of tool wear. This method is easily affected by subjective judgment, which is not only time-consuming and labor-intensive, but also not conducive to improving the quality of part processing, and may even lead to serious safety accidents. At present, the manufacturing industry is in a critical stage of transformation and upgrading, and intelligent development has become an irreversible trend. Machine vision technology can not only replace the human eye, but also complete tasks that the human eye cannot complete, realize information integration, and promote intelligent production. Due to the uncontrollability of many aspects of the processing process, the differences in tool types and workpiece materials, it is difficult to break through the online detection of tool wear. Some scholars have used indirect measurement methods such as monitoring force signals and spindle currents in actual production, but the effect is not ideal; some scholars have used direct measurement methods of machine vision to detect wear, but they mainly focus on the detection of bottom blade wear and rarely involve the side blade of the tool.

[0004] When visually inspecting the worn and damaged bottom blade of a tool, the machine needs to be stopped or the spindle speed needs to be reduced, and the camera can be placed perpendicular to the bottom blade of the tool. When collecting images of worn and damaged side blades of a tool, the machine tool reduces the spindle speed without stopping the machine, and the camera can be fixed in position for collection. Since the side blade is spiral along the axial direction, during the rotation of the tool, the camera is only perpendicular to the side blade in the center area of ​​the image. The different angles between the camera and the arc surface of the side blade will cause large distortion of the side blade image in other areas. In addition, due to the limited field of view of the camera, multiple images need to be collected for one side blade, which is prone to duplicate or missing images. After image recognition is completed, if the start and end order of each blade position is not matched when splicing the entire side blade, overlapping problems are very likely to occur, resulting in the inability to determine the maximum wear amount and its blade position. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art, the present invention provides a visual online detection method for side edge wear and damage of an end mill. The present invention solves the problem of incomplete image acquisition through a self-compensation method of shooting angle, and establishes a deep learning model based on VGG16-UNet to solve the problem of matching and positioning wear areas between adjacent wear images.

[0006] Technical solution: A visual online detection method for side edge wear damage of an end mill, comprising the following steps:

[0007] Step 1: Sample image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect S>4 wear images as a sample set;

[0008] Step 2: Shooting angle self-compensation: Perform shooting angle self-compensation processing on the sample set in step 1 to obtain a processed sample set;

[0009] Step 3: Establish a VGG16-Unet side blade wear image recognition model: Use the VGG16-Unet neural network model to perform multi-scale feature recognition on the images in the sample set, extract image features of different scales for feature fusion, and finally train to obtain a VGG16-Unet side blade wear image recognition model;

[0010] Step 4: Detection image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect n wear images as a test set;

[0011] Step 5. Obtain the milling cutter wear result: Input the test set into the VGG16-Unet side edge wear image recognition model in step 3 for processing to obtain the final milling cutter wear result;

[0012] Step 6: Evaluate the accuracy of the milling cutter wear detection in step 5.

[0013] Preferably, in the step 2, in order to describe the change of the end mill wear amount with the shooting angle, the curve fitting method is used to fit the discrete data points when the wear image contour is extracted to obtain the fitting curve that is most suitable for characterizing the change of discrete points, that is, the fitting curve that best fits the change of discrete points. With the help of the least squares method, the most suitable fitting function is solved by minimizing the sum of squares of errors. The shooting angle x corresponds to the wear value y one by one (x, y). After the t-order polynomial fitting process, the fitting degree R is used. 2 Conduct an assessment.

[0014] Preferably, the shooting angle self-compensation process in step 2 is specifically as follows:

[0015] Let the quartic polynomial function

[0016] y(x)=a0+a1x+a2x 2 +a3x3 +a4x 4 (1)

[0017] In the formula, a0, a1, a2, a3, a4——polynomial coefficients;

[0018] The error sum of squares ε and its partial derivative can be obtained

[0019]

[0020] Among them, S is the number of wear images collected, y i is the wear value of the i-th image, x i is the shooting angle of the i-th image, a j are the polynomial coefficients, is the j-th power term of the shooting angle in the polynomial function of the wear value of the i-th image;

[0021] The matrix expression of the error square sum ε partial derivative is

[0022] X T XQ=X T Y (3)

[0023]

[0024] Where X, Q, and Y are the matrix forms of the partial derivatives of ε in equation (2); X T is the transposed matrix of matrix X;

[0025]

[0026] Since the independent variables x1, x2, …, x S They are different from each other, and the number of acquisitions is greater than the number of fittings. After a0, a1, a2, a3, and a4 are determined, the fitting degree is calculated to find the extreme value within the acquisition range, and the maximum value is the output result of the shooting angle self-compensation processing.

[0027] Preferably, the VGG16-Unet neural network model in step 3 includes trunk feature extraction, enhanced feature extraction, and output prediction results;

[0028] The backbone feature extraction adopts a pure convolutional network VGG16 with the same 3x3 convolution kernel parameters, that is, a pure convolutional network VGG16 with 13 convolutional layers and 3 fully connected layers. The deep features of the image are extracted through 13 convolutions and 4 poolings. The original image size is downsampled 4 times to obtain a feature map, and the milling cutter wear map feature extraction is completed.

[0029] The enhanced feature extraction is upsampling plus deconvolution stacking, and the resolution of the milling cutter wear feature map is restored by fusing low-level features and high-level features at the same resolution;

[0030] The output prediction result is to classify each feature point using the last valid feature layer finally obtained, which is equivalent to dividing each pixel point, and finally outputting a binary map of milling cutter wear.

[0031] Preferably, the training steps of the VGG16-Unet neural network model include forward propagation and back propagation;

[0032] The forward pass is the process of passing the input data from the input layer through each hidden layer of the neural network in sequence and finally outputting the prediction result;

[0033] The back propagation is a process of minimizing the loss function by calculating the gradient of the loss function with respect to the parameters of each layer, namely, weights and biases, and updating the parameters from the output layer to the input layer.

[0034] Preferably, the forward transfer is specifically:

[0035] For any side edge wear image (i, j), when undergoing the l-th layer of convolution operation, the forward input of the l-th layer without the activation function and the forward output after the activation function can be expressed as

[0036]

[0037] Where n l Indicates the number of convolution kernels in layer l; Represents the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer; represents the bias of the p-channel of the l-th layer; u is the three neighbors in the row direction of the wear image (i, j), and v is the three neighbors in the column direction of the wear image (i, j). is the forward output of the q channel of the l-1th layer after the activation function, is the forward output of the p-channel of the l-th layer after the activation function, is the forward input of the kth channel of the lth layer without the activation function.

[0038] Preferably, the back propagation is specifically:

[0039] Using cross entropy loss as the loss function

[0040]

[0041] in, is the forward output of the lth layer after the activation function, y i is the label of the training sample, indicating the correct classification type of the sample;

[0042] The gradient of the loss function with respect to the forward input of layer l It is called the error δ of the lth layer, and the back propagation error of the jth channel of the lth layer is

[0043]

[0044] Where W l represents the weight of the fully connected network at layer l; r l Indicates the connectivity of the nodes in the lth layer, which obeys the Bernoulli distribution; is the forward input of the ith channel of the l+1th layer without the activation function, is the forward output of the j-th channel of the l-th layer after the activation function, is the forward input of the jth channel of the lth layer without the activation function;

[0045] So the loss function can be obtained and The gradient of , and then use the gradient descent method to update the parameters:

[0046]

[0047] in, is the back propagation error of the p-channel in the l-th layer, is the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer.

[0048] Preferably, the step six is ​​specifically as follows:

[0049] The IoU value, recall rate and precision rate are used to measure the model segmentation ability indicators, among which:

[0050] Recall: The accuracy of the model predicting positive samples:

[0051]

[0052] Precision: The ratio of positive samples to total samples in the prediction results:

[0053]

[0054] Among them, TP is the situation where the tool wear area can be accurately identified; TN is the situation where the non-wear area can be accurately identified; FP is the situation where the tool wear area is incorrectly identified; FN is the situation where the non-wear area is incorrectly identified.

[0055] Beneficial effects: According to the change of side blade wear with shooting angle, the present invention develops a shooting angle self-compensation method based on least squares to solve the problem of incomplete image acquisition caused by side blade curvature and camera angle when extracting wear contour. A deep learning model based on VGG16-UNet is established to extract image features of different scales for feature fusion, eliminating image splicing and matching, and solving the problem of matching and positioning of wear areas between adjacent wear images. The model has strong robustness and high accuracy, greatly improving the problem of misclassification and omission of wear areas, and greatly improving the detection accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0057] Figure 1 is a flow chart of the method of the present invention;

[0058] Figure 2 It is the VGG16-Unet neural network model diagram of the present invention;

[0059] Figure 3 There are 8 groups of comparison pictures for the detection and analysis of milling cutter wear results, among which (a) is the original image, (b) is the image detected by the proposed method, (c) is the image detected by threshold segmentation, and (d) is the image detected by Canny method. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] In the description of the present invention, it is necessary to understand that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0062] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0063] like Figure 1 As shown, a visual online detection method for side edge wear and damage of an end mill comprises the following steps:

[0064] Step 1: Sample image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect S>4 wear images as a sample set;

[0065] In this embodiment, a total of 1050 images are collected for model training and verification, and the ratio of sample set to verification set is 9:1. After preprocessing, feature extraction and other image processing methods, labels of corresponding images are produced, and the original images and labels are input into the model for model training and testing. After that, the milling cutter wear images are collected online and imported into the VGG16-Unet model to verify the model segmentation results.

[0066] Step 2: Shooting angle self-compensation: Perform shooting angle self-compensation processing on the sample set in step 1 to obtain a processed sample set;

[0067] In order to describe the change of end mill wear with shooting angle, the curve fitting method is used to fit discrete data points when extracting the wear image contour, and the fitting curve that best represents the change of discrete points is obtained, that is, the fitting curve that best fits the change of discrete points. With the help of the least squares method, the most suitable fitting function is solved by minimizing the sum of squares of errors. The shooting angle x corresponds to the wear value y one by one (x, y). After the t-order polynomial fitting process, the fitting degree R is used. 2 Conduct an assessment.

[0068] The shooting angle self-compensation process is specifically as follows:

[0069] Let the quartic polynomial function

[0070] y(x)=a0+a1x+a2x 2 +a3x 3 +a4x 4 (1)

[0071] In the formula, a0, a1, a2, a3, a4——polynomial coefficients;

[0072] The error sum of squares ε and its partial derivative can be obtained

[0073]

[0074] Among them, S is the number of wear images collected, y i is the wear value of the i-th image, x i is the shooting angle of the i-th image, a j are the polynomial coefficients, is the j-th power term of the shooting angle in the polynomial function of the wear value of the i-th image;

[0075] The matrix expression of the error square sum ε partial derivative is

[0076] X T XQ=X T Y (3)

[0077]

[0078] Where X, Q, and Y are the matrix forms of the partial derivatives of ε in equation (2); X T is the transposed matrix of matrix X;

[0079]

[0080] Since the independent variables x1, x2, …, x S They are different from each other, and the number of acquisitions is greater than the number of fittings. After a0, a1, a2, a3, and a4 are determined, the fitting degree is calculated to find the extreme value within the acquisition range, and the maximum value is the output result of the shooting angle self-compensation processing.

[0081] Step 3: Establish a VGG16-Unet side blade wear image recognition model: Use the VGG16-Unet neural network model to perform multi-scale feature recognition on the images in the sample set, extract image features of different scales for feature fusion, and finally train to obtain a VGG16-Unet side blade wear image recognition model;

[0082] like Figure 2 As shown, the VGG16-Unet neural network model in step 3 includes trunk feature extraction, enhanced feature extraction, and output prediction results;

[0083] The backbone feature extraction adopts a pure convolutional network VGG16 with the same 3x3 convolution kernel parameters, i.e., 13 convolutional layers and 3 fully connected layers. The deep features of the image are extracted through 13 convolutions and 4 poolings. The original image size is downsampled 4 times to obtain a feature map, and the milling cutter wear map feature extraction is completed.

[0084] The enhanced feature extraction is upsampling plus deconvolution stacking, and the resolution of the milling cutter wear feature map is restored by fusing low-level features and high-level features at the same resolution;

[0085] The output prediction result is to classify each feature point using the last valid feature layer finally obtained, which is equivalent to dividing each pixel point, and finally outputting a binary image of milling cutter wear;

[0086] The training steps of the VGG16-Unet neural network model include forward propagation and back propagation;

[0087] The forward pass is the process of passing the input data from the input layer through each hidden layer of the neural network in sequence and finally outputting the prediction result:

[0088] For any side edge wear image (i, j), when undergoing the l-th layer of convolution operation, the forward input of the l-th layer without the activation function and the forward output after the activation function can be expressed as

[0089]

[0090] Where n l Indicates the number of convolution kernels in the lth layer; Represents the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer; represents the bias of the p-channel of the l-th layer; u is the three neighbors in the row direction of the wear image (i, j), and v is the three neighbors in the column direction of the wear image (i, j). is the forward output of the q channel of the l-1th layer after the activation function, is the forward output of the p-channel of the l-th layer after the activation function, is the forward input of the kth channel of the lth layer without the activation function.

[0091] The back propagation is the process of minimizing the loss function by calculating the gradient of the loss function with respect to the parameters of each layer, namely the weights and biases, and updating the parameters from the output layer to the input layer:

[0092] Using cross entropy loss as the loss function

[0093]

[0094] in, is the forward output of the lth layer after the activation function, y i is the label of the training sample, indicating the correct classification type of the sample;

[0095] The gradient of the loss function with respect to the forward input of layer l It is called the error δ of the lth layer, and the back propagation error of the jth channel of the lth layer is

[0096]

[0097] Where W l represents the weight of the fully connected network at layer l; r l Indicates the connectivity of the nodes in the lth layer, which obeys the Bernoulli distribution; is the forward input of the ith channel of the l+1th layer without the activation function, is the forward output of the j-th channel of the l-th layer after the activation function, is the forward input of the jth channel of the lth layer without the activation function;

[0098] So the loss function can be obtained and The gradient of , and then use the gradient descent method to update the parameters:

[0099]

[0100] in, is the back propagation error of the p-channel in the l-th layer, is the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer.

[0101] Step 4: Detection image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect n wear images as a test set;

[0102] Step 5. Obtain the milling cutter wear result: Input the test set into the VGG16-Unet side edge wear image recognition model in step 3 for processing to obtain the final milling cutter wear result;

[0103] Step 6. Evaluate the accuracy of the milling cutter wear detection in step 5:

[0104] The IoU value, recall rate and precision rate are used to measure the model segmentation ability indicators, among which:

[0105] Recall: The accuracy of the model predicting positive samples:

[0106]

[0107] Precision: The ratio of positive samples to total samples in the prediction results:

[0108]

[0109] Among them, TP is the situation where the tool wear area can be accurately identified; TN is the situation where the non-wear area can be accurately identified; FP is the situation where the tool wear area is incorrectly identified; FN is the situation where the non-wear area is incorrectly identified.

[0110] In order to verify the segmentation effect of the VGG16-UNet model, it is compared with the traditional threshold segmentation method and Canny edge detection method. The threshold segmentation method is a simple image segmentation technology that divides pixels into foreground and background by setting a threshold; Canny edge detection is an accurate edge detection method that can detect edge information in an image; both methods have important application value in the fields of image processing and computer vision.

[0111] The collected photos of the side wear of the milling cutter are input into VGG16-UNet, threshold segmentation method and edge detection method respectively. The segmentation effects of the three methods are shown in the following figure. Figure 3 As shown in Table 1, the IoU value of each image is obtained:

[0112] Table 1 IoU value of each image

[0113]

[0114] from Figure 3 From the detection results in Table 1, it can be seen that the maximum IoU value of threshold segmentation is 0.6169, and the maximum IoU of Canny operator is 0.7000. The methods used in this paper are all above 0.8, and the segmentation effect of VGG16-UNet model is significantly better than the other two methods. The threshold segmentation method also recognizes and segments the background that does not belong to the tool, and mistakenly regards the bright area in the background as the wear area, which is an incorrect segmentation. The segmentation of the wear area by the Canny method is fuzzy and incomplete, and basically cannot complete the recognition and segmentation tasks. The VGG16-UNet model in this paper has a significant recognition effect on the wear image of the milling cutter, can remove background noise, and can completely identify and segment the wear area, greatly improving the problem of misclassification and omission of the wear area. It verifies that the VGG16-UNet segmentation model proposed in this paper has strong robustness and high accuracy in the recognition and segmentation of the wear area, which has guiding significance for the online visual detection and prediction of tool wear, and the effect is significantly better than the other two methods.

[0115] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0116] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A visual online detection method for side edge wear damage of an end mill, characterized in that: The following steps are involved: Step 1: Sample image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect S>4 wear images as a sample set; Step 2: Shooting angle self-compensation: Perform shooting angle self-compensation processing on the sample set in step 1 to obtain a processed sample set; Step 3: Establish a VGG16-Unet side blade wear image recognition model: Use the VGG16-Unet neural network model to perform multi-scale feature recognition on the images in the sample set, extract image features of different scales for feature fusion, and finally train to obtain a VGG16-Unet side blade wear image recognition model; Step 4: Detection image acquisition: Change the shooting angle of the same side edge of the milling cutter and collect n wear images as a test set; Step 5. Obtain the milling cutter wear result: Input the test set into the VGG16-Unet side edge wear image recognition model in step 3 for processing to obtain the final milling cutter wear result; Step 6: Evaluate the accuracy of the milling cutter wear detection in step 5.

2. The visual online detection method for side grinding damage of end mill according to claim 1, characterized in that: In the step 2, in order to describe the change of the end mill wear with the shooting angle, the curve fitting method is used to fit the discrete data points when extracting the wear image contour, and the fitting curve that is most suitable for characterizing the change of discrete points is obtained, that is, the fitting curve that best fits the change of discrete points. With the help of the least squares method, the most suitable fitting function is solved by minimizing the sum of squares of errors. The shooting angle x corresponds to the wear value y one by one (x, y). After the t-order polynomial fitting process, the fitting degree R is used. 2 Conduct an assessment.

3. The visual online detection method for side grinding damage of end mill according to claim 2, characterized in that: The shooting angle self-compensation process in step 2 is specifically as follows: Let the quartic polynomial function y(x)=a0+a1x+a2x 2 +a3x 3 +a4x 4 (1) In the formula, a0, a1, a2, a3, a4——polynomial coefficients; The error sum of squares ε and its partial derivative can be obtained in, S is the number of wear images collected, y i is the wear value of the i-th image, x i is the shooting angle of the i-th image, a j are the polynomial coefficients, is the j-th power term of the shooting angle in the polynomial function of the wear value of the i-th image; The matrix expression of the error square sum ε partial derivative is X T XQ=X T Y (3) Where X, Q, and Y are the matrix forms of the partial derivatives of ε in equation (2); X T is the transposed matrix of matrix X; Since the independent variables x1, x2, …, x S They are different from each other, and the number of acquisitions is greater than the number of fittings. After a0, a1, a2, a3, and a4 are determined, the fitting degree is calculated to find the extreme value within the acquisition range, and the maximum value is the output result of the shooting angle self-compensation processing.

4. The visual online detection method for side grinding damage of end mill according to claim 1, characterized in that: The VGG16-Unet neural network model in step 3 includes trunk feature extraction, enhanced feature extraction, and output prediction results; The backbone feature extraction adopts a pure convolutional network VGG16 with the same 3x3 convolution kernel parameters, that is, a pure convolutional network VGG16 with 13 convolutional layers and 3 fully connected layers. The deep features of the image are extracted through 13 convolutions and 4 poolings. The original image size is downsampled 4 times to obtain a feature map, and the milling cutter wear map feature extraction is completed. The enhanced feature extraction is upsampling plus deconvolution stacking, and the resolution of the milling cutter wear feature map is restored by fusing low-level features and high-level features at the same resolution; The output prediction result is to classify each feature point using the last valid feature layer finally obtained, which is equivalent to dividing each pixel point, and finally outputting a binary map of milling cutter wear.

5. The visual online detection method for side edge wear damage of end mill according to claim 4, characterized in that: The training steps of the VGG16-Unet neural network model include forward propagation and back propagation; The forward pass is the process of passing the input data from the input layer through each hidden layer of the neural network in sequence and finally outputting the prediction result; The back propagation is a process of minimizing the loss function by calculating the gradient of the loss function with respect to the parameters of each layer, namely, weights and biases, and updating the parameters from the output layer to the input layer.

6. The visual online detection method for side grinding damage of end mill according to claim 5, characterized in that: The forward transfer is specifically as follows: For any side edge wear image (i, j), when undergoing the l-th layer of convolution operation, the forward input of the l-th layer without the activation function and the forward output after the activation function can be expressed as Where n l Indicates the number of convolution kernels in layer l; Represents the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer; represents the bias of the p-channel of the l-th layer; u is the three neighbors in the row direction of the wear image (i, j), and v is the three neighbors in the column direction of the wear image (i, j). is the forward output of the q channel of the l-1th layer after the activation function, is the forward output of the p-channel of the l-th layer after the activation function, is the forward input of the kth channel of the lth layer without the activation function.

7. The visual online detection method for side grinding damage of end mill according to claim 5, characterized in that: The back propagation is specifically as follows: Using cross entropy loss as the loss function in, is the forward output of the lth layer after the activation function, y i is the label of the training sample, indicating the correct classification type of the sample; The gradient of the loss function with respect to the forward input of layer l It is called the error δ of the lth layer, and the back propagation error of the jth channel of the lth layer is Where W l represents the weight of the fully connected network at layer l; r l Indicates the connectivity of the nodes in the lth layer, which obeys the Bernoulli distribution; is the forward input of the ith channel of the l+1th layer without the activation function, is the forward output of the jth channel of the lth layer after the activation function, is the forward input of the jth channel of the lth layer without the activation function; So the loss function can be obtained and The gradient of , and then use the gradient descent algorithm to update the parameters: in, is the back propagation error of the p-channel in the l-th layer, is the convolution kernel corresponding to the p channel of the lth layer and the q channel of the l-1th layer.

8. The visual online detection method for side grinding damage of end mill according to claim 1, characterized in that: The step six is ​​specifically as follows: The IoU value, recall rate and precision rate are used to measure the model segmentation ability indicators, among which: Recall: The accuracy of the model predicting positive samples: Precision: The ratio of positive samples to total samples in the prediction results: Among them, TP is the situation where the tool wear area can be accurately identified; TN is the situation where the non-wear area can be accurately identified; FP is the situation where the tool wear area is incorrectly identified; FN is the situation where the non-wear area is incorrectly identified.

Images