CN117474873B - Surface treatment system before brazing of high-chromium wear-resistant castings - Google Patents

Abstract

The invention belongs to the technical field of defect detection, in particular to a surface treatment system before brazing of a high-chromium wear-resistant casting, which aims at solving the problems that the actual situation of defects cannot be known only by three-dimensional model judgment, and whether a die casting can be used in the next process cannot be further analyzed, and the invention provides a scheme which comprises the following steps of surface defect detection, wherein the specific steps of the surface defect detection method are as follows: step one: six image acquisition cameras are uniformly arranged in the whole space of the casting to be detected, so that images of all angles of the casting are ensured to be acquired; step two: the method has the advantages that the method is high in calculation speed and accuracy, the casting machining condition can be effectively obtained, and subsequent procedures are guided.

Description

Surface treatment system before brazing of high-chromium wear-resistant castings

Technical Field

The invention relates to the technical field of defect detection, in particular to a surface treatment system before brazing of a high-chromium wear-resistant casting.

Background

Die casting is a part which is manufactured into a mechanical die casting machine through the pressure of a die, and parts with different shapes are formed through a series of heating pouring inlets on metal, and the part is also called a die casting; however, different types of defects are formed on the die casting after the processing is finished, and the defects need to be measured and judged to judge whether the defects are in an error range or not;

in the prior art, the defect measurement is carried out by manual measurement judgment, the use method is subjective to judge whether the die casting is qualified or not, the judgment speed is slow, the speed cannot be improved, and the labor cost is increased; the method for judging the quality of the die casting by using machine learning is higher than the method for judging the quality of the die casting by using manpower, but the machine learning needs longer time, so that the deep learning efficiency is high;

the existing market uses a three-dimensional model to judge defects on a die casting, only a plurality of defects on the die casting can be identified through the three-dimensional model, but the area of the defects on different surfaces is not known, and whether the defects are in a standard error range or not is not known, so that the real situation of the defects cannot be judged only through the three-dimensional model, and whether the die casting can be used in the next working procedure cannot be further analyzed.

Disclosure of Invention

The invention provides a surface treatment system before brazing of a high-chromium wear-resistant casting, which solves the defects in the prior art that a three-dimensional model is used for judging the casting, only a plurality of defects are contained on the casting through the three-dimensional model, but the area of the defects on different surfaces is not known, and whether the area is within a standard error range or not is not known, so that the real situation of the defects cannot be judged only through the three-dimensional model, and the defects that whether the casting can be used in the next working procedure or not cannot be further analyzed.

The invention provides the following technical scheme:

the surface treatment system before brazing of the high-chromium wear-resistant casting comprises surface defect detection, wherein the detection method for detecting the surface defect comprises the following specific steps:

step one: six image acquisition cameras are uniformly arranged in the whole space of the casting to be detected, so that images of all angles of the casting are ensured to be acquired;

step two: and respectively carrying out Gaussian filtering operation on the acquired images, and smoothing the images, wherein a distribution equation of the Gaussian filtering function in a two-dimensional space is as follows:wherein->The standard deviation of normal distribution can be adjusted according to specific detection environment working conditions, and the default value is 1.08;

step three: inputting the smoothed image into a FasterR-CNN network to start to detect image defects;

step four: extracting candidate frames from the featuremap generated in the steps by adopting a regional generation network (RegionProposalNetwork, RPN) in a FasterR-CNN-based network structure;

step five: simultaneously inputting the featuremap obtained in the third step and the recommended region information obtained in the fourth step into a pooling layer ropooling to obtain a feature map proposalfeaturemap of the recommended region;

step six: inputting the data obtained in the previous step into a full-connection layer softmax, classifying the image features by using a full-connection function, and judging the defect type;

step seven: calculating the defective area of the casting image;

step eight: establishing a transformation function of the image scale and the actual casting size, inversely transforming the calculated image defect area back to a physical space, calculating the actual defect area, judging whether the actual defect area is within an error range,

step nine, rotating the camera angles, respectively rotating 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees, and respectively acquiring casting images again from 6 directions, namely 6×5 images are generated by one casting, namely 6 azimuth 5 angles.

The step nine further comprises the following substeps,

step ten: for defect types with similar characteristics, firstly, counting the defect types with similar characteristics according to actual working conditions, then adding a classification network for the characteristics, strengthening the identification capability of the classification network on the defects, for the defects which are mutually influenced and coupled together, acquiring a multi-defect segmentation effect based on a weak supervision method, dividing each defect subarea under the condition of coupling of multiple defects, extracting the characteristics of the subareas independently, identifying the defect types, and reducing the mutual influence among the defects;

step eleven: aiming at the new pictures obtained after each rotation angle, similarity calculation is carried out on the new pictures and the pictures before rotation, the pictures with high similarity do not enter a deep learning network any more, and only the pictures with low similarity are further calculated;

step twelve: the parallel architecture is adopted to optimize the operation speed of the implementation end of the FasterR-CNN, so that the real-time performance of the detection system is improved;

step thirteen: non-linear fitting is carried out on the characteristics of the collected effective pictures and the camera angles corresponding to the physical space, the zero pole of a non-linear fitting function is solved, and the non-linear fitting function is inversely transformed to the physical space, so that the optimal angle of the camera is obtained;

step fourteen: fusing the detection results, taking the optimal camera angle as the center, wherein the weight of the corresponding defect characteristic value is maximum, the farther the angle is from the optimal angle, the smaller the weight proportion is, and the falling speed of the weight proportion refers to a Gaussian functionWherein->Is standard deviation of normal distribution, can be adjusted according to specific detection environment working conditions, and takes 1 as default value and 1 as default value>Is the difference between the current angle and the optimal angle.

In step 4, the RPN specifically includes the following steps:

generating anchor frames, mapping each point of the featuremap back to the original image, taking the central point as an anchor point, and then selecting k anchor frames with different sizes and proportions around the anchor point, wherein 3 scales and 3 aspect ratios are selected, and if 1 anchor point corresponds to 9 anchor frames, one m×n featuremap can produce m×n×9 anchor frames;

judging whether each anchor frame contains a defect, extracting the anchor frame containing the defect, wherein the step is a classification problem, and classifying by adopting CNN;

fine tuning the anchor frame by utilizing bounding box regression to generate a new anchor frame closest to the real frame, defining (x, y) as coordinate values, (w, h) as the image scale size,is true value +.>For predictive value +.>For the original anchor frame, the regression procedure is as follows:

representing a linear function at four coordinate scales, representing a scale transformation and a translation transformation of an original anchor frame, and solving a transformation matrix w of the linear function as follows:

tx, ty represents a coordinate translation coefficient, tw, th represents a scale transform coefficient;

adopting a non-maximum value inhibition method to calculate weights of the parts with overlapped anchor frames and eliminating anchor frames with smaller weights;

finally, the recommended defect feature region proposals given by the RPN network is obtained.

In step six, the full join function isAnd zi is input, j is the number of full connection elements, and meanwhile, the accurate position of the defect area is calculated by a resampling boundary box regression method.

In the seventh step, the specific steps of calculating the defective area of the casting image are as follows:

inputting the 6 images with different angles obtained in the first step into the FasterR-CNN network to obtain 6 groups of defects and positions thereof;

re-projecting the defect location into the original image;

extracting defects at boundary positions in 6 original images according to the relative positions of cameras and the shape characteristics of castings;

summing the defect areas at the borders to obtain the whole defect area, and keeping the defect areas at the non-borders of the image unchanged to obtain all defect types and the area sizes thereof.

The method can accurately identify the defects of the castings under various angles, calculate the actual area of the defects, judge whether the defects are in an error range, and is high in calculation speed and accuracy, the processing condition of the castings can be effectively acquired, and subsequent procedures are guided.

Drawings

Fig. 1 is a schematic flow chart of a detection method of the present application.

Detailed Description

Embodiments of the present invention will be described below with reference to the accompanying drawings in the embodiments of the present invention.

In describing embodiments of the present invention, it should be noted that, unless explicitly stated and limited otherwise, the terms "coupled" and "mounted" should be interpreted broadly, and for example, "coupled" may or may not be detachably coupled; may be directly connected or indirectly connected through an intermediate medium. In addition, "communication" may be direct communication or may be indirect communication through an intermediary. Wherein, "fixed" means that the relative positional relationship is not changed after being connected to each other. References to orientation terms, such as "inner", "outer", "top", "bottom", etc., in the embodiments of the present invention are merely to refer to the orientation of the drawings and, therefore, the use of orientation terms is intended to better and more clearly illustrate and understand the embodiments of the present invention, rather than to indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and are therefore not to be construed as limiting the embodiments of the present invention.

In embodiments of the present invention, the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more such feature.

In the embodiment of the present invention, "and/or" is merely an association relationship describing an association object, and indicates that three relationships may exist, for example, a and/or B may indicate: a exists alone, A and B exist together, and B exists alone. In addition, the character "/" herein generally indicates that the front and rear associated objects are an "or" relationship.

Reference in the specification to "one embodiment" or "some embodiments" or the like means that a particular feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of the invention. Thus, appearances of the phrases "in one embodiment," "in some embodiments," "in other embodiments," and the like in the specification are not necessarily all referring to the same embodiment, but mean "one or more but not all embodiments" unless expressly specified otherwise. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless expressly specified otherwise.

The application provides a high chromium wear-resisting foundry goods surface treatment system before brazing, including the surface defect detection can discern the defect position fast to calculate the defect area according to defect position area, judge whether its size is in the error range, specific step is as follows, as shown in figure 1:

step one: six image acquisition cameras are uniformly arranged in the whole space of the casting to be detected, so that images of all angles of the casting are ensured to be acquired.

Step two: respectively to collectCarrying out Gaussian filter operation on the image, and smoothing the image, wherein a distribution equation of the Gaussian filter function in a two-dimensional space is as follows:wherein->The standard deviation of normal distribution can be adjusted according to specific detection environment working conditions, and the default value is 1.08.

Step three: inputting the smoothed image into a FaterR-CNN network to start to detect image defects, and inputting the image into a CNN convolutional neural network to obtain a characteristic image featuremap of the image, wherein the CNN consists of an input layer, a convolutional layer, an active layer, a pooling layer and a full-connection layer, and a VGG deep convolutional network is adopted, ((convolutional layer+active layer) ×2+pooling layer) ×2+ ((convolutional layer+active layer) ×3+pooling layer) ×2+ (convolutional layer+active layer) ×3.

Step four: in order to acquire the position of the image defect, the image subregion is required to be continuously selected to judge whether the defect exists, but the operation is extremely long in time consumption and low in efficiency, and a region generation network (RegionProposalNetwork, RPN) is adopted in the FaterR-CNN-based network structure to extract a candidate frame from the featuremap generated in the steps, so that the calculation speed of the model is increased;

the RPN comprises the following specific steps:

generating anchor frames, mapping each point of the featuremap back to the original image, taking the central point as an anchor point, then selecting k anchor frames with different sizes and proportions around the anchor point, selecting 3 scales and 3 aspect ratios, and producing m multiplied by n multiplied by 9 anchor frames by one m multiplied by n multiplied by one m multiplied by 9 anchor frame by 1 anchor point;

judging whether each anchor frame contains a defect, extracting the anchor frame containing the defect, wherein the step is a classification problem, and classifying by adopting CNN;

the anchor frame is finely adjusted by utilizing bounding box regression to generate a new anchor frame closest to the real frame, wherein (x, y) is defined as a coordinate value, and (w, h) is defined as an image rulerThe degree of the degree,is true value +.>For predictive value +.>For the original anchor frame, the regression procedure is as follows:

representing a linear function under four coordinate scales, and representing scale transformation and translation transformation of an original anchor frame;

the transformation matrix w of the linear function is solved as follows:

tx, ty represents a coordinate translation coefficient, tw, th represents a scale transform coefficient;

adopting a non-maximum value inhibition method to calculate the weight of the part with overlapped anchor frames, eliminating anchor frames with smaller weight,

finally, the recommended defect feature region proposals given by the RPN network is obtained.

Step five: and (3) inputting the featuremap obtained in the step (III) and the recommended region information obtained in the step (IV) into a pooling layer ROIPooling at the same time to obtain a feature map proposalfeaturemap of the recommended region.

Step six: the defect identification and positioning are carried out, the data obtained in the previous step are input into a full-connection layer softmax, the image characteristics are classified by utilizing a full-connection function, and the defect type is judged;

the full join function isAnd zi is input, j is the number of full connection elements, and meanwhile, the accurate position of the defect area is calculated by a resampling boundary box regression method.

Step seven: the casting image defect area is calculated, and the steps are as follows:

inputting the 6 images with different angles obtained in the first step into the FasterR-CNN network to obtain 6 groups of defects and positions thereof;

re-projecting the defect location into the original image;

extracting defects at boundary positions in 6 original images according to the relative positions of cameras and the shape characteristics of castings;

and summing the defect areas at the borders to obtain the whole defect area. And the defect area of the non-bordering part of the image is kept unchanged, and finally, all defect types and the area size of the defect types can be obtained.

Step eight: and establishing a transformation function of the image scale and the actual casting size, inversely transforming the calculated image defect area back to a physical space, calculating the actual defect area, and judging whether the actual defect area is in an error range or not.

Step nine: further, in order to solve the influence of light reflection phenomena generated by greasy dirt, rust and the like in defects on image acquisition, a polarization filter is arranged from a hardware angle, possible reflected light in a specific direction is shielded, after the steps one to eight are completed from an algorithm angle, a camera angle is rotated, 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees are respectively rotated, casting images are acquired from 6 directions respectively, namely, 6×5 images are generated by one casting in total, namely, 6 directions and 5 angles are generated.

The choice of the rotation camera angle may not always be optimal and experiments and analysis are required to determine the most suitable angle.

Step ten: for defect types with similar characteristics, in order to increase detection accuracy, firstly, statistics is carried out on which defect types are similar in characteristics according to actual working conditions, and then a classification network is added for the characteristics, so that the identification capability of the classification network on the defects is enhanced.

For defects which are mutually influenced and coupled together, a multi-defect segmentation effect is obtained based on a weak supervision method, each defect subarea under the condition of coupling of multiple defects is divided, then characteristic extraction is carried out on the subareas separately, defect types are identified, and the mutual influence among the defects is reduced.

Step eleven: firstly, in order to solve the problem of data storage requirement, aiming at the new pictures acquired after each rotation angle, similarity calculation is carried out on the new pictures and the pictures before rotation, the pictures with high similarity do not enter a deep learning network any more, and only the pictures with low similarity are further calculated, so that the number of the pictures processed by each casting is reduced, and the data storage pressure is reduced;

then, the parallel architecture is adopted to optimize the operation speed of the implementation end of the FasterR-CNN, so that the real-time performance of the detection system is improved;

and then, carrying out nonlinear fitting on the characteristics of the acquired effective pictures and the camera angles corresponding to the physical space, solving the zero pole of a nonlinear fitting function, and inversely transforming to the physical space to obtain the optimal angle of the camera.

Finally, fusing the detection results, taking the optimal camera angle as the center, wherein the weight of the corresponding defect characteristic value is maximum, the farther the angle is from the optimal angle, the smaller the weight proportion, and the falling speed of the weight proportion refers to a Gaussian functionWherein->Is standard deviation of normal distribution, can be adjusted according to specific detection environment working conditions, and takes 1 as default value and 1 as default value>For the difference value between the current angle and the optimal angle, the real defect type and the defect size of the position can be accurately judged according to the fused detection result.

In summary, the method can accurately identify defects of the castings under various angles, calculate the actual area of the defects and judge whether the actual area of the defects is in an error range, and has the advantages of high calculation speed, high accuracy, capability of effectively acquiring the processing condition of the castings and guiding subsequent procedures.

Claims (3)

1. The surface treatment system before brazing of the high-chromium wear-resistant casting comprises surface defect detection, wherein the detection method for detecting the surface defect comprises the following specific steps:

step one: six image acquisition cameras are uniformly arranged in the whole space of the casting to be detected, so that images of all angles of the casting are ensured to be acquired;

step two: and respectively carrying out Gaussian filtering operation on the acquired images, and smoothing the images, wherein a distribution equation of the Gaussian filtering function in a two-dimensional space is as follows:wherein->The standard deviation of normal distribution is 1.08;

step three: inputting the smoothed image into a FasterR-CNN network to start to detect image defects;

step four: extracting a candidate frame from the featuremap generated in the step by using an area generation network RPN in a FaterR-CNN-based network structure;

in step four, the RPN specifically includes the following steps:

generating anchor frames, mapping each point of the featuremap back to the original image, taking the central point as an anchor point, then selecting k anchor frames with different sizes and proportions around the anchor point, wherein the anchor frames comprise 3 scales and 3 aspect ratios, and the anchor frames correspond to 9 anchor frames, so that the featuremap of m multiplied by n can produce m multiplied by n multiplied by 9 anchor frames;

judging whether each anchor frame contains a defect, extracting the anchor frame containing the defect, and classifying by adopting CNN;

fine tuning the anchor frame by utilizing bounding box regression to generate a new anchor frame closest to the real frame, defining (x, y) as coordinate values, (w, h) as the image scale size,to be a true value of the value,for predictive value +.>For the original anchor frame, the regression procedure is as follows:

d _* (P) represents a linear function at four coordinate scales, representing a scale transformation and a translation transformation of the original anchor frame, wherein x, y, w, h;

adopting a non-maximum value inhibition method to calculate weights of the parts with overlapped anchor frames and eliminating anchor frames with smaller weights;

finally, obtaining a recommended defect feature region proposals given by the RPN network;

step five: simultaneously inputting the featuremap obtained in the third step and the recommended region information obtained in the fourth step into a pooling layer ropooling to obtain a feature map proposalfeaturemap of the recommended region;

step six: inputting the data obtained in the fifth step into a full-connection layer softmax, classifying the image features by using a full-connection function, and judging the defect type;

step seven: calculating the defective area of the casting image;

step eight: establishing a transformation function of the image scale and the actual casting size, inversely transforming the calculated image defect area back to a physical space, calculating the actual defect area, judging whether the actual defect area is within an error range,

step nine, rotating the camera angles, respectively rotating 30 degrees, 60 degrees, 90 degrees, 120 degrees and 150 degrees, respectively acquiring casting images from 6 directions again, namely, 6×5 images are generated by one casting, namely, 6 azimuth 5 angles are generated, the step nine also comprises the following substeps,

step ten: for defect types with similar characteristics, firstly, counting which defect types are similar in characteristics according to actual working conditions, then adding a classification network for the characteristics, strengthening the identification capability of the classification network on the defects, acquiring a multi-defect segmentation effect for the defects which are mutually influenced and/or coupled together by adopting a weak supervision method, dividing each defect subarea under the condition of coupling of multiple defects, extracting the characteristics of the subareas independently, identifying the defect types, and reducing the mutual influence among the defects; step eleven: aiming at the new pictures obtained after each rotation angle, similarity calculation is carried out on the new pictures and the pictures before rotation, the pictures with high similarity do not enter a deep learning network any more, and only the pictures with low similarity are further calculated;

step twelve: the parallel architecture is adopted to optimize the operation speed of the implementation end of the FasterR-CNN, so that the real-time performance of the detection system is improved;

step thirteen: non-linear fitting is carried out on the characteristics of the collected effective pictures and the camera angles corresponding to the physical space, the zero pole of a non-linear fitting function is solved, and the non-linear fitting function is inversely transformed to the physical space, so that the optimal angle of the camera is obtained;

step fourteen: fusing the detection results, taking the optimal camera angle as the center, wherein the weight of the corresponding defect characteristic value is maximum, the farther the angle is from the optimal angle, the smaller the weight proportion is, and the falling speed of the weight proportion refers to a Gaussian functionWherein->Is the standard deviation of normal distribution, and the value is 1 +.>Is the difference between the current angle and the optimal angle.

2. A high chromium wear resistant cast pre-braze surface treatment system in accordance with claim 1, wherein in step six, the full junction boxThe number is，Z _i And j is the number of the full connection elements for input, and meanwhile, calculating the accurate position of the defect area by adopting a bounding box regression method again.

3. The surface treatment system before brazing of high-chromium wear-resistant castings according to claim 1, wherein in the seventh step, the specific steps of calculating the defective areas of the castings image are as follows:

inputting the 6 images with different angles obtained in the first step into the FasterR-CNN network to obtain 6 groups of defects and positions thereof;

re-projecting the defect location into the original image;

extracting defects at boundary positions in 6 original images according to the relative positions of cameras and the shape characteristics of castings;

summing the defect areas at the borders to obtain the whole defect area, and keeping the defect areas at the non-borders of the image unchanged to obtain all defect types and the area sizes thereof.