CN110405540B - Artificial intelligence broken cutter detection system and method - Google Patents

Abstract

The invention discloses an artificial intelligence broken knife detection system which comprises an image acquisition module, a knife library, a cloud server and an image processing module, wherein the cloud server is used for obtaining a broken knife detection model based on knife intersection point characteristic training; the image processing module is used for preprocessing the cutter image, loading a cutter breaking detection model obtained by training in the cloud server, and then calculating a cutter detection result based on the cutter breaking detection model. The invention discloses an artificial intelligence broken cutter detection method, which comprises the steps of shooting a processed cutter image, marking an auxiliary line, extracting intersection point characteristics between a cutter point of a cutter and the auxiliary line to train a broken cutter detection model, and carrying out incremental training on the cutter image with errors to continuously update the broken cutter detection model, so that the detection of various cutters of different types can be realized, the influence of external environment is reduced, and the generalization capability and the detection accuracy of the broken cutter detection model are greatly improved.

Description

A kind of artificial intelligence broken knife detection system and method

technical field

The invention belongs to the field of intelligent numerical control machine tools, and more particularly relates to an artificial intelligence tool breakage detection system and method.

Background technique

In order to make CNC machine tools more intelligent and automatic, ensuring the automatic operation of CNC machine tools with high precision, high speed and high efficiency has become one of the main research directions of CNC machine tools. During the machining process, the machining tools of CNC machine tools will be damaged to varying degrees along with the machining process, which will affect the machining efficiency. Therefore, timely detection of tool breakage can reduce the scrap rate of subsequent parts, reduce the loss of machine tools, and ensure the processing efficiency of machine tools.

The existing vision-based tool breakage detection methods are mainly divided into two types: indirect detection and direct detection. The indirect detection detects the processed parts. The photoelectric devices on the left and right sides and the fixing device detection device realize the detection of the left and right sides of the part, and then a photoelectric detection device is set under the worktable to detect the bottom of the part to determine whether there is a broken knife, which is an indirect In this way, not only the cost of the entire equipment is relatively high, but also different judgment methods are required for different types of processed parts, and the usable range of the entire equipment is relatively narrow. Another direct detection method directly photographs the tool to determine whether the tool is worn or not. In the patent numbered CN109500657A, the image of the corresponding machining tool is first obtained through the set threshold value, and the image is grayed to obtain The corresponding binary image, after morphological processing, extracts the contour of the binary image, calculates the area ratio of the extracted contour and the calibrated normal tool, determines the relationship between it and the preset threshold, and realizes the detection of the tool. This method needs to pre-calibrate the area of the normal tool as a reference value. When calibrating the entire tool area on the picture, the influence of manual calibration error is relatively large. In addition, when extracting the contour, it is greatly affected by illumination and background, and generalization The ability is weak, when there are many types of tools, the fault tolerance rate is small and the accuracy is low.

Therefore, it is an urgent problem to propose a broken tool detection system and method with strong generalization ability and high accuracy.

SUMMARY OF THE INVENTION

In view of the defects of the prior art, the purpose of the present invention is to propose an artificial intelligence tool breakage detection system and method, which aims to solve the problem that the prior art is subject to manual calibration when the pre-calibrated area of a normal tool is used as a reference value to judge a broken tool. And the problem of lower accuracy caused by the greater influence of light and Beijing.

In order to achieve the above object, one aspect of the present invention provides an artificial intelligence broken knife detection system, comprising:

Image acquisition module, tool magazine, cloud server, image processing module;

Wherein, the output end of the image acquisition module is connected with the input end of the image processing module, there is a distance between the image acquisition module and the tool magazine, and the image processing module and the cloud server communicate through Ethernet;

The image acquisition module is used to capture the tool to be processed in the tool magazine to obtain the calibration information of the tool tip position, and to capture the tool image after processing to obtain a tool image with a blurred background and a prominent tool object, and transmit it to the image processing module;

The tool magazine is used to store the tools to be processed and the tools to be tested after processing;

The cloud server is used to obtain a broken tool detection model based on the tool intersection feature training;

The image processing module is used to preprocess the tool image, load the broken tool detection model trained in the cloud server, and then calculate the tool detection result based on the broken tool detection model.

Further preferably, the image acquisition module includes an endoscope, and the endoscope is used to photograph the tools in the tool magazine. When the tool is not being processed, the tool image is photographed, and the position information of the tool tip is calibrated. After the tool is processed, the tool is reset. Switch back to the tool magazine, use an endoscope to photograph the tool to be inspected after processing, and obtain a tool image with a blurred background to highlight the tool object.

Further preferably, the image processing module may be an IPC equipped with an AI chip.

Further preferably, the image acquisition module, the tool magazine, and the image processing module can all be embedded in the machine tool;

Further preferably, another aspect of the present invention provides a kind of artificial intelligence broken knife detection method, comprises the following steps:

S1. The image processing module is connected to the cloud server, and the broken knife detection model is downloaded from the cloud server;

S2. Take pictures of all the tools to be processed in the tool magazine and calibrate the position of the tool tip;

S3. The tool to be processed is changed back to the corresponding position of the tool magazine after the processing is completed, and a photo is taken to obtain the image of the tool to be detected;

S4. Preprocess the image of the tool to be detected based on the calibrated tool tip position information to obtain a tool image with auxiliary marks;

S5, extracting the intersection feature between the tool tip of the tool image to be detected and the auxiliary line;

S6, input the intersection point feature into the pre-trained broken knife detection model to obtain the broken knife detection result;

S7. Send the tool image samples that have detected errors to the cloud server to obtain an error sample set.

Further preferably, the method for preprocessing the tool image in the above steps includes the following steps:

S41. Based on the calibrated tool tip position information, the tool image is trimmed to obtain a fixed-size tool image;

S42, performing grayscale processing on the cropped image to obtain a grayscale image;

S43, draw an auxiliary line at a fixed position in the obtained grayscale image to obtain a tool image with auxiliary marks.

Further preferably, the auxiliary line positions of all preprocessed images are the same.

Further preferably, the number of auxiliary lines is greater than or equal to 1, and there is a certain inclination angle and are parallel to each other; further preferably, the number of auxiliary lines is 3.

Further preferably, the more intersections between the tool tip and the auxiliary line, the less likely the tool is to be broken.

Further preferably, GoogleNet can be used to extract the feature of the intersection between the tool tip and the auxiliary line of the tool image.

Further preferably, the method for obtaining the broken knife detection model in step S6 comprises the following steps:

S61, determine whether there is a trained broken tool detection model on the cloud server, if not, collect the image data set of the broken tool and the normal tool as a training set, and go to step S62; if there is, go to step S63;

S62, train the broken knife detection model on the cloud server based on the obtained training set, and go to S63;

S63. When the number of samples in the error sample set on the cloud server is greater than the trainable threshold C, perform incremental training based on the error sample set on the basis of the current broken knife detection model, and update the broken knife detection model.

Further preferably, the method for training the broken knife detection model in step S62 comprises the following steps:

S621. Preprocess the tool images in the training set to construct a training set with auxiliary marks;

S622, extracting the intersection point feature between the tool tip and the auxiliary line of each tool image in the training set with auxiliary marks;

S623 , taking the intersection feature of the normal tool in the training set as a positive sample, and taking the intersection feature of the broken tool in the training set as a negative sample, train the classifier to obtain a broken tool detection model.

Further preferably, SENet can be used as the classifier.

Through the above technical scheme conceived by the present invention, compared with the prior art, the following beneficial effects can be achieved:

1. The present invention provides an artificial intelligence tool breakage detection method, which trains a broken tool detection model by extracting the intersection point feature between the tool tip and the auxiliary line, and performs incremental training on the wrong tool image to continuously detect the broken tool. The detection model can be updated to detect a variety of different types of tools, which greatly improves the generalization ability of the broken tool detection model and the detection accuracy.

2. The present invention provides a broken knife detection system, which uses an endoscope to capture a knife image, which can blur the background part of the knife image, thereby highlighting the tool object. The endoscope has low cost, simple installation and easy maintenance. In addition, the collected tool image is the image after the tool processing is completed, which reduces the interference of environmental factors, avoids the problem of coolant and workpiece occlusion when shooting the tool during processing, and effectively reduces the light exposure during the entire processing process. Influence, in the case of relatively harsh processing environment can also achieve better detection results.

3. The present invention provides an artificial intelligence tool breakage detection method. Before the detection is started, the tool tip position information of the tool to be processed is pre-calibrated, and the entire tool range is determined by the tool tip position information. Some methods that use threshold to determine the tool range get more accurate tool information and are less affected by the shooting background.

4. The present invention provides an artificial intelligence broken tool detection system. The broken tool detection model is trained on the cloud server. When the model is updated, it can be synchronized to different machine tools in time, which is more flexible and can realize the aggregation of the data of the entire system. , which is conducive to the processing and analysis of the data of the entire machine tool, and also has better maintainability.

Description of drawings

Fig. 1 is a kind of artificial intelligence broken knife detection system provided by the embodiment of the present invention;

Fig. 2 is a kind of artificial intelligence broken knife detection method provided by the embodiment of the present invention;

3 is a training method of a broken knife detection model provided by an embodiment of the present invention;

4 is a normal tool image with auxiliary marks after preprocessing provided by an embodiment of the present invention;

FIG. 5 is an image of a broken tool with auxiliary marks after preprocessing provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The present invention proposes an artificial intelligence broken knife detection system, as shown in FIG. 1 , including an image acquisition module 1, a tool magazine 2, a cloud server 3, and an image processing module 4;

Wherein, the output end of the image acquisition module 1 is connected with the input end of the image processing module 4, there is a distance between the image acquisition module 1 and the tool magazine 2, and the image processing module 4 and the cloud server 3 communicate through Ethernet;

The image acquisition module 1 is used to take pictures of the tools to be processed in the tool magazine 2 and calibrate to obtain the calibration information of the position of the tool tip, and to take pictures of the finished tools to obtain a tool image with a blurred background and a prominent tool object, and transmit it to the image processing module 4 in;

The tool magazine 2 is used to store the tools to be processed and the tools to be tested after the processing is completed;

The cloud server 3 is used to obtain a broken tool detection model based on the tool intersection feature training;

The image processing module 4 is used to preprocess the tool image, load the broken tool detection model trained in the cloud server, and then calculate the tool detection result based on the broken tool detection model.

Specifically, the image acquisition module 1 includes an endoscope, a lighting lamp, and a USB data cable. The endoscope is fixed in the CNC machine tool with a bracket, and the distance from the tool in the tool magazine is 10 cm. The USB data cable is used to connect the endoscope and the CNC machine. The upper computer in the machine tool is connected. Lights are used to illuminate the tool magazine area. Since the tool magazine location is dimly lit, the use of lights can make the tools in the photo clearer. The endoscope is used to photograph the tools in the tool magazine. When the tool is not being processed, the tool image is taken and the position information of the tool tip is calibrated. After the tool is processed, the tool is returned to the tool magazine. The tool to be detected is obtained, the tool image with blurred background is obtained to highlight the tool object, and the image data is transmitted to the image processing module through the USB data line. There is a problem of coolant and workpiece occlusion when the tool is photographed during the machining process. The tool needs to be replaced from the machining area to the tool magazine to avoid the occlusion problem.

Specifically, the broken tool detection model is trained on the cloud server and loaded into the image processing module, and the tool image captured in the image acquisition module is transmitted to the image processing module. After the preprocessing is completed, the broken tool in the image processing module is used The knife detection model performs knife break detection on the processed image.

Specifically, the image processing module may be an IPC equipped with an AI chip.

Specifically, another aspect of the present invention provides an artificial intelligence detection method for broken knife, as shown in Figure 2, comprising the following steps:

S1. The image processing module is connected to the cloud server to determine whether the broken knife detection model in the image processing module exists. If it does not exist, download the broken knife detection model from the cloud server. If the download of the model fails, an error message will be returned, and the algorithm will end; , go to step 2;

S2. Use an endoscope to take pictures of all the tools to be processed in the tool magazine and calibrate the position of the tool tip;

S3. When the machine tool is processing, poll the tool change signal of the machine tool, after capturing the tool change signal, the tool to be detected after the processing is returned to the corresponding position of the tool magazine, and the endoscope is used to take a picture of the tool to obtain the image of the tool to be detected. ;

S4. Preprocess the image of the tool to be detected based on the calibrated tool tip position information to obtain a tool image with auxiliary marks;

S5. Use GoogleNet to extract the intersection feature between the tool tip and the auxiliary line of the tool image to be detected; specifically, through the intersection of the auxiliary line and the tool tip, the gradient change of the pixels near the intersection is enhanced, making it easier to extract image features when extracting image features. At the intersection, the number and position of the intersection can reflect the length of the tool, so that the qualitative feature of length can be converted into the quantitative feature of the position and quantity of the intersection, so that the training is faster and more accurate.

S6. Input the intersection point feature into the pre-trained tool break detection model to obtain the tool break detection result. If the tool is predicted to be broken, send an alarm signal to the machine tool, otherwise continue to poll the tool change information.

S7 , judging whether the detected result of the broken tool is correct, and transmitting the tool image sample of the detected error to the cloud server to obtain the error sample set, and the algorithm ends.

Specifically, the method for training a broken knife detection model on the cloud server, as shown in Figure 3, includes the following steps:

S61, determine whether there is a trained broken tool detection model on the cloud server, if not, collect image data sets of broken tools and normal tools as a training set, and go to step S62; if so, go to step S63; , the images in the training set are tool images that contain the tool tip;

S62, train the broken knife detection model on the cloud server based on the obtained training set, and go to S63;

S63. Determine whether the number of samples in the error sample set on the cloud server is greater than the trainable threshold C, if it is greater than the trainable threshold C, perform incremental training based on the error sample set on the basis of the current broken knife detection model, and update the broken knife Detect the model, the algorithm ends; otherwise, the current broken knife detection model is the result, and the algorithm ends. Specifically, the value of C is 4000. Specifically, the number of normal tool samples is 2000, and the number of broken tool samples is 2000. By incrementally training the model on the basis of the existing parameters of the current broken tool detection model, it is possible to Improve the fault tolerance and accuracy of the model.

Specifically, the step of training the broken knife detection model in step S62 includes:

S621. Preprocess the tool images in the training set to construct a training set with auxiliary marks;

S622, using the GoogleNet model to extract the intersection feature between the tool tip and the auxiliary line of each tool image in the training set with auxiliary marks;

S623 , taking the intersection feature of the normal tool in the training set as a positive sample, and taking the intersection feature of the broken tool in the training set as a negative sample, train the SENet classifier to obtain a broken tool detection model.

Specifically, the method for preprocessing the tool image in the above steps includes the following steps:

S41. Based on the position (x, y) of the tool tip in the tool image calibrated before tool processing, take 50 pixels on the left and right sides of x and 100 pixels above y to cut the processed tool image , get a 100×100 tool image;

S42, performing grayscale processing on the cropped image to obtain a grayscale image;

S43, draw three parallel auxiliary lines with a certain inclination angle at a fixed position in the lower half of the obtained grayscale image to obtain a tool image with auxiliary marks. Specifically, all the preprocessed images have the same position of the auxiliary lines. Specifically, FIG. 4 is a preprocessed image of a normal tool with auxiliary marks, and FIG. 5 is a preprocessed image of a broken tool with auxiliary marks. As can be seen from the figure, through the intersection of the auxiliary line and the tool tip, the longer the tool, the more the number of intersection points. The intersection of the normal tool and the auxiliary line is obviously more than that of the broken tool. In addition, the auxiliary line has a certain inclination angle to increase the intersection point. The number makes the detection more accurate.

Through the artificial intelligence broken tool detection system and method provided by the present invention, the broken tool detection model is trained by extracting the intersection point between the tool tip and the auxiliary line, and the wrong tool images are collected for incremental training, and the broken tool is continuously trained. The tool detection model is updated so that it can detect a variety of different types of tools, which greatly improves the generalization ability of the broken tool detection model and the detection accuracy. The model is trained by adding auxiliary line features to the 100×100 tool image containing the tool tip, which increases the accuracy of the model by 3%. The size of the training set used here is 4000, with 2000 for each type of interrupted knives and normal knives, and 4000 pieces of test set data, 2000 for each type of interrupted knives and normal knives, and the final accuracy rate of the model on the test set is 98.6 %, the accuracy rate is high.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (8)

1. An artificial intelligence broken knife detection system is characterized by comprising an image acquisition module, a knife library, a cloud server and an image processing model;

the output end of the image acquisition module is connected with the input end of the image processing module, the image acquisition module and the tool magazine are separated by a certain distance, and the image processing module and the cloud server are communicated through Ethernet;

the image acquisition module is used for shooting a tool to be machined in the tool magazine to obtain the calibration information of the position of the tool tip, shooting the machined tool to obtain a tool image with a fuzzy background and a highlighted tool object, and transmitting the tool image to the image processing module;

the tool magazine is used for storing the tool to be machined and the tool to be machined after machining is finished;

the image processing module is used for preprocessing the cutter image: cutting the cutter image by taking the marked cutter point position information as a reference to obtain a cutter image with a fixed size, and performing gray processing on the cut image to obtain a gray image; drawing an auxiliary line at a fixed position in the gray level image to obtain a cutter image with an auxiliary mark, and extracting intersection point characteristics between the cutter point of the cutter image and the auxiliary line to obtain intersection point characteristics of the cutter; loading a broken cutter detection model obtained by training in a cloud server, and calculating a cutter detection result based on the broken cutter detection model;

the cloud server is used for obtaining a broken cutter detection model based on cutter intersection point characteristic training.

2. The system for detecting the broken cutter according to claim 1, wherein the image acquisition module comprises an endoscope, the endoscope is used for shooting a cutter in the tool magazine, when the cutter is not machined, an image of the cutter is shot, the position information of the point of the cutter is calibrated, the cutter is replaced to the tool magazine after the cutter is machined, the endoscope is used for shooting the cutter to be detected after the machining is finished, and a cutter image with a fuzzy background is obtained so as to highlight a cutter object.

3. The system of claim 1, wherein the image processing module is an IPC equipped with an AI chip.

4. An artificial intelligence knife-breaking detection method is characterized by comprising the following steps:

s1, connecting the image processing module with a cloud server, and downloading a broken cutter detection model from the cloud server;

s2, photographing all the to-be-machined tools in the tool magazine and calibrating the positions of tool tips of the to-be-machined tools;

s3, after the tool to be machined is machined, the tool to be machined is replaced to the corresponding position of the tool magazine, and the tool magazine is photographed to obtain an image of the tool to be machined;

s4, preprocessing the cutter image to be detected based on the marked cutter tip position information to obtain a cutter image with an auxiliary mark; wherein the auxiliary mark is an auxiliary line;

s5, extracting intersection point characteristics between the cutter point of the cutter to be detected and the auxiliary line;

s6, inputting the intersection point characteristics into a cutter breaking detection model trained in advance to obtain a cutter breaking detection result;

and S7, transmitting the tool image sample with the error detection to a cloud server to obtain an error sample set.

5. The method of claim 4, wherein the method of preprocessing the tool image comprises the steps of:

s41, cutting the cutter image by taking the marked cutter point position information as a reference to obtain a cutter image with a fixed size;

s42, carrying out gray scale processing on the cut image to obtain a gray scale image;

and S43, drawing an auxiliary line at a fixed position in the gray-scale image to obtain a tool image with an auxiliary mark.

6. The method according to claim 4 or 5, wherein the number of the auxiliary lines is 1 or more, and the auxiliary lines are parallel to each other at a predetermined inclination angle.

7. The method of claim 4, wherein the method of obtaining the pre-trained knife break detection model comprises the steps of:

s61, judging whether a trained broken cutter detection model exists at the cloud server side, if not, collecting image data sets of a broken cutter and a normal cutter as a training set, and turning to the step S62; if yes, go to step S63;

s62, training a broken cutter detection model on the cloud server based on the training set, and turning to S63;

and S63, when the number of samples in the error sample set on the cloud server is larger than a trainable threshold value C, performing incremental training based on the error sample set on the basis of the current tool breaking detection model, and updating the tool breaking detection model.

8. The method of claim 7, wherein the method of training the break detection model comprises the steps of:

s621, preprocessing the cutter images in the training set to construct a training set with auxiliary marks;

s622, extracting intersection point characteristics between the tool nose of each tool image in the training set with the auxiliary marks and the auxiliary lines;

and S623, training the classifier by taking the intersection point characteristics of the normal cutters in the training set as positive samples and the intersection point characteristics of the broken cutters in the training set as negative samples to obtain a broken cutter detection model.

Images