CN108335288A - The crater image method for detecting abnormality of view-based access control model clarity and contours extract - Google Patents

Abstract

The present invention proposes a kind of variation of the width of the value and profile according to vision definition to differentiate the detection method of the affiliated abnormal class of the crater image；The present invention is from image vision angle, in conjunction with clarity and the computation performance of contours extract, first with clarity operator just classify, is classified again to weldering speed in conjunction with profile width, reduce operation time to a certain extent.And the used double light path visual perception device of the present invention, it is effective to improve vision profile and clarity extraction accuracy.

Description

Anomaly Detection Method of Melt Pool Image Based on Visual Clarity and Contour Extraction

technical field

The invention belongs to the field of molten pool vision, and in particular relates to a dual optical path visual sensing device triggered by an FPGA module and a molten pool image abnormality detection method for collaborative judgment of definition and contour.

Background technique

Welding is one of the important processing methods in the manufacturing industry and is widely used in material processing and structure manufacturing. Skilled welders can predict and control the weld quality by observing the surface information of the molten pool combined with experience. As intelligent robot welding gradually replaces manual welding, accurate sensing of molten pool information is an important prerequisite for intelligent control of the welding process. The stable and reliable visual sensing system ensures that the acquired image information of the molten pool is as comprehensive and accurate as possible.

The two-dimensional visual sensing of the molten pool mainly uses the visual sensing method to collect images of the molten pool, and through image processing and feature extraction, the relationship between image features and welding quality is analyzed and a control model is established. With the deepening of the research, researchers have established a reasonable visual sensing system based on the characteristics of the molten pool in different materials and different welding methods. Great progress has been made in welding quality control. Accurate sensing of molten pool information is an important prerequisite for intelligent control of the welding process. It is necessary to establish a stable and reliable visual sensing system to ensure that the acquired molten pool image information is as comprehensive and accurate as possible. If the visual sensor is directly used to capture the image of the molten pool during the welding process, the strong arc light will make the photosensitive element of the CCD reach light saturation, and the information of the molten pool is almost completely obliterated by the arc light.

In order to weaken the influence of arc light, this patent adopts a joint sensing device for molten pool vision and process parameters, and designs an imaging solution for the visual computing requirements based on definition and contour extraction. Since the calculation of the molten pool outline requires a strong boundary contrast and a high-bandpass exposure; the calculation of the visual clarity of the molten pool surface requires suppression of arc light and high-pass low exposure, so this patent uses a beam splitter to divide the molten pool beam into 850nm high-pass and 650nm Band-pass dual optical path visual sensing device. In this way, high-contrast molten pool boundaries and molten pool surface information with low arc light interference can be obtained at the same time, which improves the accuracy of subsequent molten pool contour extraction and definition calculation. After the image is obtained, in order to reduce the impact of various random noises and image distortion, this patent uses the method of intermediate frequency stretching to preprocess the image to suppress useless noise information, improve image quality, and facilitate the calculation of definition and contour extraction.

Contents of the invention

The present invention is based on visual clarity and contour extraction method for abnormal detection of molten pool image, comprising the following steps:

Step 1: Determine the possible abnormal categories of the molten pool image;

Step 2: Design process parameter cooperative sensing FPGA module, and design targeted imaging solutions according to different visual computing requirements;

Step 3: Collect positive samples and various negative samples for the optical path designed in step 2;

Step 4: Perform intermediate frequency stretching processing on all collected molten pool image samples and calculate their sharpness values, set thresholds for dividing various abnormal groups according to the distribution of results, and determine abnormal current parameters, abnormal voltage parameters and abnormal shielding gas parameters range of clarity;

Step 5: Extract the contour feature information of the normal group in the case of the same electrical parameters and shielding gas parameters in the results obtained in step 4, and perform data analysis to obtain the classification results of different welding speed levels, and realize the detection of abnormal welding pool welding speed.

Furthermore, according to different visual computing requirements described in step 2, a targeted imaging scheme is designed. The specific process is: if the calculation of the outline of the molten pool requires a strong boundary contrast, high exposure with a band pass can be used; the surface of the calculated molten pool has a clear vision High-pass low-exposure can be used to suppress arc light; therefore, a beam splitter is used to divide the molten pool beam into two beams, one with 850nm high-pass and one with 650nm band-pass, to form a dual-spectrum visual sensing device; and to ensure dual-spectrum sampling Synchronize.

Furthermore, performing intermediate frequency stretching processing on all collected melt pool image samples and calculating their sharpness values as described in step 4 includes the following steps:

Step 3-1: Perform intermediate frequency stretching preprocessing on the obtained melt pool image, and extract the more important detail components in the image;

The stretching formula of the intermediate frequency component is as follows:

In the formula, H(x, y) is the frequency domain of the stretched image, D(x, y) is the frequency domain of the original input image, d _l is the starting frequency of intermediate frequency stretching, d _h is the frequency of intermediate frequency stretching Cut-off frequency, m, n is the order of the filter;

Step 3-2: Select the sharpness operator of step 3-1 to calculate the sharpness of the image;

Step 3-3: manually delineate the classification results, and find the relationship between the sharpness value and the corresponding category;

Calculate the values of the above sharpness operators respectively, manually set the thresholds of various abnormal results, and obtain the highest discrimination accuracy rate of the energy gradient operator, and the results of the sharpness evaluation operator of the energy gradient are as follows:

That is, the sharpness value of the image obtained by adding the differences of the gray values of adjacent pixels in the row direction and the column direction of the image, where I(x, y) is the gray value of the image I at (x, y) degree value, q(I) is the sharpness value of the output result of this function.

Furthermore, the collection of positive samples and various negative samples under the condition of dual optical paths described in step 3 specifically includes abnormal current, abnormal voltage, abnormal shielding gas and abnormal welding speed.

Furthermore, in step 5, the specific steps for extracting the contour feature information of the normal group under the condition that the electrical parameters and the shielding gas parameters in the results obtained in step 4 are consistent are as follows:

Step 5-1: ROI setting of molten pool;

First, set a large ROI to remove the redundant background area; second, according to the overall characteristics of the molten pool, that is, the head is brighter and easily interfered by arc light, and the semi-solidified area at the tail is lower in brightness and less interference, it can be divided into ROI1 and ROI2; according to the imaging characteristics of the head and tail regions, small-scale grayscale stretching is performed respectively, and the grayscale stretching formula is as follows:

In the formula, for small scale stretching A smaller value can remove the highlighted area, for large scale stretching A larger value can enhance the overall contrast of the tail;

Step 5-2: block pretreatment of molten pool;

Gaussian filtering and opening operation are performed on the gray-scale stretched image in ROI1 to reduce the interference of arc light at the edge, and then the Canny operator is used to perform low-threshold edge detection and filter out too small edge contours; because the interference in the ROI2 area is small , the gray distribution is relatively uniform, so the ROI2 area is directly segmented by the Otsu method threshold value and the Canny operator is used for high threshold edge detection;

Step 5-3: Contour merging and connection;

Merge the edge contours detected by ROI1 and ROI2, and search the 8-neighborhood and 16-neighborhood of the image, connect the small-scale fractures in the contour, and then detect the contour endpoints for large-scale fractures. Connect, and finally get the outline of the molten pool connection.

The beneficial effects of the present invention are as follows: this patent starts from the visual angle of the image, combines the sharpness and the calculation characteristics of the outline extraction, first uses the sharpness operator to carry out the initial classification, and then combines the outline width to reclassify the welding speed, to a certain extent Reduced computing time. Moreover, the dual optical path visual perception device adopted in this patent can effectively improve the extraction accuracy of visual outline and definition.

Description of drawings

Fig. 1 is that the present invention triggers and synchronizes 850 high-pass, 650 band-pass images;

Fig. 2 is the classification result of the clarity of the present invention;

Fig. 3 is the extraction of molten pool profile of the present invention;

Fig. 4 is the classification of molten pool welding speed in the present invention.

Detailed ways

Referring to Fig. 1, 2, 3, 4, the specific implementation steps of this patent are as follows:

Step 1, design process parameters cooperative sensing FPGA module, build dual optical path molten pool visual sensor to collect molten pool image;

Step 2: Convert the collected 850 high-pass images to the frequency domain through Fourier transform, enhance the details in the frequency domain, and use the inverse Fourier transform to convert the image from the frequency domain back to the time domain again, and obtain the stretched intermediate frequency Result graph;

Step 3, select various sharpness operators to evaluate the sharpness of the stretched image, and classify according to the category of the experimental samples, and find the sharpness operator with the best classification effect;

Step 4, the classification effect of the energy gradient operator in step 3 is the best, as shown in Figure 2, the energy gradient operator can classify the abnormal phenomena of the Ar=0 group, the current group is too low, and the voltage group is too low, according to Manually set the threshold division, the clarity value of abnormal shielding gas q>5, abnormal current: q<0.6, low voltage group: 0.6<q<2, and normal group of electrical parameters and shielding gas parameters: 2<q<5. And the recognition accuracy of each of the above parts is abnormal shielding gas: 90.0%, abnormal current: 98.44%, abnormal voltage: 91.33%, normal electrical parameters and shielding gas parameters: 98.68%. Overall accuracy: 95.23%;

Step 5, when using the test image for real-time anomaly detection, if the value of sharpness q is greater than 5 or less than 2, suspend welding and check the corresponding abnormality, if the sharpness value 2<q<5, then perform contour Extract, and further detect whether there is an abnormal phenomenon of welding speed;

In step 6, feature information of the molten pool is extracted according to the obtained molten pool profile, such as: molten pool width, molten pool length, molten pool drag angle, and molten pool aspect ratio. According to the extracted width information, the hard threshold segmentation can be used to distinguish the welding speed level under the same welding process and the same welding electrical parameters under the condition of constant camera angle. The higher the welding speed, the smaller the width of the molten pool and the longer the length of the molten pool. On the contrary, the lower the welding speed, the larger the width of the molten pool and the smaller the length of the molten pool. By analyzing the data, the hard segmentation thresholds of different welding speed levels are obtained, from th1 to th6. Among them, the corresponding welding speed of grade 1 is 2mm/s, the corresponding welding speed of grade 4 is 4mm/s, the corresponding welding speed of grade 3 is 6mm/s, the corresponding welding speed of grade 4 is 8mm/s, and the corresponding welding speed of grade 5 is 12mm/s , grade six corresponds to a welding speed of 16mm/s, as shown in Figure 4.

Claims (5)

1. the crater image method for detecting abnormality of view-based access control model clarity and contours extract, which is characterized in that include the following steps：

Step 1：Determine the abnormal class that crater image is likely to occur；

Step 2：Design technology parameter collaborative perception FPGA module calculates demand, design specific aim imaging side according to different visions Case；

Step 3：The light path acquisition positive sample design step 2 and all kinds of negative samples；

Step 4：Intermediate frequency stretch processing is carried out to all crater image samples of acquisition and calculates its definition values, according to result point Cloth, setting divide all kinds of threshold values organized extremely, determine current parameters exception, voltage parameter is abnormal and protects gas abnormal parameters Clarity range；

Step 5：Normal group of the electrical parameter in step 4 acquired results in the case of consistent with protection gas parameter is subjected to contour feature letter The extraction of breath carries out data analysis to obtain the classification results of different weldering speed grades, realizes the detection of molten bath weldering speed exception.

2. molten bath vision according to claim 1 and technological parameter collaborative perception device, which is characterized in that described in step 2 Calculate demand according to different visions, design specific aim is imaged scheme, and detailed process is：It is stronger such as to calculate molten bath profile needs The exposure of band logical height can be used in boundary contrast；It calculates weld pool surface vision definition to need to inhibit arc light, the low exposure of high pass can be used Light；Therefore molten bath light beam is divided into two bundles using Amici prism, it is a branch of to use 850nm high passes, is a branch of using 650nm band logicals, shape Spectral Visual sensing device in pairs；And ensure that double spectrum samples synchronize.

3. the molten bath Outlier Detection Algorithm of view-based access control model clarity according to claim 1, which is characterized in that step 4 institute All crater image samples to acquisition stated carry out intermediate frequency stretch processing and calculate its definition values, include the following steps：

Step 3-1：The pretreatment that intermediate frequency stretching is carried out to the crater image obtained, extracts details more important in image point Amount；

It is as follows that intermediate frequency component stretches formula：

In formula, H (x, y) is the image frequency domain after stretching, and D (x, y) is the frequency domain of original input image, d_lThe starting stretched for intermediate frequency Frequency, d_hBe intermediate frequency stretch by frequency, m, n are the exponent number of filter；

Step 3-2：Optional step 3-1 clarity operator calculates the clarity of image；

Step 3-3：Classification results delimited manually, find definition values and play corresponding class relations；

The value for calculating separately above-mentioned clarity operator, manually sets the threshold value of all kinds of abnormal results, obtains energy gradient operator Differentiate that accuracy rate highest, the clarity evaluation operator result of energy gradient are as follows：

The clarity for the image that image row direction is added with the difference of the gray value of neighbor pixel on column direction It is worth, in formula, I (x, y) is gray values of the image I at (x, y), and q (I) is the output result definition values of the function.

4. the molten bath Outlier Detection Algorithm according to claim 1 based on clarity and contours extract, which is characterized in that Positive sample and all kinds of negative samples are acquired under the conditions of double light path described in step 3, specifically includes current anomaly, and electric voltage exception is protected It protects gas exception and weldering speed is abnormal.

5. the molten bath Outlier Detection Algorithm according to claim 1 based on clarity and contours extract, which is characterized in that Normal group of the electrical parameter in step 4 acquired results in the case of consistent with protection gas parameter is subjected to contour feature letter described in step 5 The extraction of breath the specific steps are：

Step 5-1：The ROI in molten bath is arranged；

First, big ROI is set and removes extra background area；Secondly, according to the overall permanence in molten bath, i.e., head brightness it is higher and It is easy to be interfered by arc light, half resolidified region brightness of tail portion is relatively low and interferes less, can be classified as ROI1 and ROI2；According to head The imaging characteristic in portion and tail region, carries out small scale gray scale stretching respectively, and gray scale stretching formula is as follows：

In formula, small scale is stretchedValue it is smaller can remove highlight regions, for large scale stretch's It is worth larger, the contrast of tail portion entirety can be enhanced；

Step 5-2：The partitioning pretreatment in molten bath；

Gaussian filtering and opening operation are carried out to weaken the interference of edge arc light to the image that gray scale stretching in ROI1 is crossed, then used Canny operators carry out Low threshold edge detection and filter out too small edge contour；Since the interference in the regions ROI2 is smaller, gray scale point Cloth is relatively uniform, therefore directly carries out Da-Jin algorithm Threshold segmentation to the regions ROI2 and carry out high threshold side with Canny operators Edge detects；

Step 5-3：Profile merges and connection；

ROI1 and the ROI2 edge contour detected are merged, and 8 neighborhoods and 16 neighborhoods of image are scanned for, is connected The fracture of profile Small and Medium Sized later carries out the fracture of large scale the detection of profile endpoint, abutting end point is attached, Finally obtain the profile of molten bath connection.