A vent hole identification and positioning method based on an Adaboost algorithm

Abstract

This invention discloses a ventilation hole identification and localization method based on the AdaBoost algorithm, comprising: deploying a detection robot with visual detection capabilities into the generator chamber, and acquiring video images of elliptical ventilation holes, single-row circular ventilation holes, and double-row circular ventilation holes through two cameras on its abdomen; processing the acquired images; performing Brenner gradient calculation, Tenengrad gradient calculation, Laplacian gradient calculation, and variance calculation on each frame of the acquired video images, and performing normalization processing; performing image recognition; image labeling; for images where the ventilation hole is about to enter or exit the field of view or where the image edge is only 0.3-0.7mm, the camera continuously acquires five images, and then locates the ventilation hole according to the label; establishing an SVM-AdaBoost model, importing the training set data into the model for training, testing the generated model using the test set data, and visualizing the training accuracy of the model; calculating the training error of the weak classifier to obtain the final classification judgment criterion.

Description

Technical Field

[0001] This invention belongs to the field of identification and positioning technology, and more specifically, relates to a method for identifying and locating ventilation holes based on the Adaboost algorithm. Background Technology

[0002] Large generators mainly consist of two modules: the rotor and the stator. During operation, the generator experiences frictional heat between the rotor and bearings, wear between components, fatigue damage caused by the constant centrifugal force of rotating parts, blockages caused by debris accumulation in ventilation holes, and insulation damage caused by aging or detachment of the stator core insulation, which can even lead to localized melting of the stator core. Regular inspection and maintenance are necessary to reduce the risk of costly forced shutdowns. In the past, due to the small gap between the stator and rotor, and the presence of wind baffles in the stator structure of some models, inspections of rotor ventilation holes and electromagnetic defect detection of the stator core required removing the rotor first. This resulted in long maintenance cycles, high costs, and considerable losses to the power plant during maintenance.

[0003] Current technologies often employ a crawling method on the stator surface for in-house inspection without removing the rotor. This method can check stator slot wedge tightness, ELCID checks, and visual inspection via cameras on all four sides. However, due to the relatively large air gap between the stator and rotor, stator-rotor misalignment, and small ventilation holes, the imaging effect inside the rotor ventilation holes is poor, making it impossible to achieve the goal of inspecting the rotor ventilation holes. Furthermore, rotor crawling technology is monopolized by foreign companies, and no related research results have been published domestically.

[0004] Therefore, a ventilation hole identification method is needed to achieve accurate rotor ventilation hole inspection. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for identifying and locating ventilation openings based on the Adaboost algorithm, comprising the following steps:

[0006] S100: Image Acquisition: The inspection robot with visual inspection function is placed into the generator chamber. With the help of positioning, the two cameras on the abdomen of the inspection robot are aligned with the rotor slot wedge. The inspection robot crawls for several minutes and acquires video images of elliptical ventilation holes, single-row circular ventilation holes, and double-row circular ventilation holes through the two cameras on the abdomen.

[0007] S200: Process the acquired images: (To ensure running efficiency and the sensitivity of evaluation results to changes in image features,) perform Brenner gradient calculation, Tenengrad gradient calculation, Laplacian gradient calculation and variance calculation on each frame of the acquired video images, and perform normalization processing.

[0008] S300: Image recognition: When the ventilation hole is located in the center of the image, the four gradient values ​​in the feature value all show a relative maximum value (which can well reflect the position information of the current camera relative to the slot wedge, and can be put into the same network model for training and recognition, but there is also a multi-peak phenomenon in the peak value, so a single threshold cannot be simply used as the judgment standard). The SVM-Adaboost method with multiple threshold intelligent tuning characteristics is selected to recognize the ventilation hole image.

[0009] S400: Image labeling: Each image is labeled. Images containing ventilation holes are labeled as 1, and images without ventilation holes are labeled as -1. 50% to 70% of the data are randomly selected as the training set for the model, and the remaining data is used as the test set.

[0010] S500: The camera continuously captures five images of the location where the ventilation hole is about to enter or exit the field of view, or where the image edge is only 0.3-0.7mm. Then, the ventilation hole is located according to the label.

[0011] S600: Build an SVM-AdaBoost model, import the training set data into the model for training, test the generated model using the test set data, and visualize the training accuracy of the model; calculate the training error of the weak classifier to obtain the final classification judgment criterion.

[0012] Furthermore, the auxiliary positioning in step S100 is achieved through an auxiliary delivery device. The auxiliary delivery device can perform circumferential rotation positioning. There is a gripper mechanism at the front end of the auxiliary delivery device, in which the detection robot is placed. While the auxiliary delivery device rotates, it drives the detection robot to rotate. When the two cameras on the abdomen of the detection robot are facing the rotor slot wedge, the robot can be delivered.

[0013] Furthermore, (generators are typically composed of stators, rotors, end covers, and bearings, and are generally concentric cylindrical structures with a rotor core in the middle and a stator core on the outside. The generator rotor is a slender cylinder with slots.) Large generators use copper busbar ventilation holes. When the object to be identified is the copper busbar ventilation hole of a large generator, in step S100, due to the special structure of the copper busbar ventilation hole, two cameras must be equipped to simultaneously observe the upper and lower surfaces of the rotor ventilation hole. For the upper surface of the generator rotor, an autofocus camera is selected. The robot's forward and backward movements are detected through multiple operations, and the rotor surface is observed in real time. After the rotor ventilation hole enters the frame, the surface of the positioning hole and the bottom of the ventilation hole are identified, and clear images of the surface and bottom of the rotor ventilation hole are collected.

[0014] The robot crawls on the surface of the rotor with its abdomen facing the rotor. The first camera first locates the full outline of the ventilation, and the second camera can then see the bottom of the ventilation hole. Structurally, two cameras are used and arranged side by side on the bottom of the inspection robot. The first side camera observes the upper surface of the rotor ventilation hole, and the second camera observes the bottom of the rotor ventilation hole. The supplementary lighting module of the second camera can rotate and extend up and down to ensure sufficient light in the field of view of the bottom of the rotor ventilation hole.

[0015] Furthermore, the method for locating the ventilation hole based on the label in step S500 is as follows: when the recognition result is label 1 twice in a row, it is determined that the current position is above the ventilation hole, the image gradient value is recorded, and the image of the area is saved; when the recognition result is label -1 more than twice in a row, it is determined that the current position has left the ventilation hole, and the image with the largest Laplacian gradient value among the previously saved images is selected as the positioning result for the next step of processing.

[0016] Furthermore, the specific steps of step S600 are as follows:

[0017] Step S601: Define a training dataset. In the first iteration, set the weight of each of the n training samples to 1 / n.

[0018] Step S602: Learn SVM weak classifier for training samples. Sample the training set according to the sample weights to obtain the training set of component SVM classifier. Train an SVM weak classifier with radial basis function kernel function, using RBF-SVM as the weak classifier.

[0019] Step S603: Calculate the standard deviation of the training set as the parameter σ of the SVM weak classifier;

[0020] σ=sqrt(mean(var(d t )))

[0021] In the formula, d t For the training sample set, sqrt represents the square root operation, mean represents the arithmetic mean calculation, and var represents the variance calculation;

[0022] Step S604: Classify each feature using an SVM weak classifier, and calculate the training error rate of each SVM classification result, which is the sum of the weights of the misclassified samples.

[0023] Step S605: Calculate the weights of the SVM classifier;

[0024] Step S606: If the training error rate in step S602 is greater than 50%, retrain; if the training error rate is less than 50%, update the weights in the training sample set and re-enter steps S602 to S606 for the next iteration. When the number of iterations reaches the set value M, proceed to step S607.

[0025] Step S607: After the Mth iteration, sum the corresponding weights and obtain the decision function value of the strongest classifier through voting.

[0026]

[0027] Among them, h m For d t The SVM weak classifier is obtained by training an SVM weak classifier with σ as the parameter, using the training sample set; α m The values ​​are the weights of the weak classifier, m = 1, 2, ..., M, representing the current iteration number. sign() is the sign function. When the value inside the parentheses is greater than "0", the output value is set to "1", indicating that the judgment result is correct. When the value inside the parentheses is not greater than "0", the output value is set to "-1", indicating that the judgment result is incorrect.

[0028] Furthermore, the specific method for defining the training dataset in step S601 is as follows:

[0029] Define a training dataset T = {(x1,y1),(x2,y2)...(x...} n ,y n )},(x i ,y i ) is one of the training data, x i Representing sample i, y i Let the label of the sample be {1, -1}, and w be the sample weight; in the first iteration, for all n training samples, the weight of each sample is set to 1 / n, and the weight distribution D1 = (w 11 ,w 12 ,…w 1i …,w 1n ),in:

[0030] Furthermore, the classifier formula used in step S602 is as follows:

[0031] RBF-SVM is used as a weak classifier, and the RBF kernel function R(x) is used. n -c i This can be represented as:

[0032]

[0033] In the formula, x n -ci Let be the Euclidean norm, ci be the center of the Green's function, and σ be the variance of the Green's function.

[0034] Furthermore, the method for calculating the training error rate of each SVM classification result in step S604, i.e., the sum of the weights of the misclassified samples, is as follows:

[0035]

[0036] In the formula, m = 1, 2, ..., M represents the current iteration round number, e m h is the training error rate. m For d t The SVM weak classifier is obtained by training an SVM weak classifier with σ as the parameter, using the training sample set as the training sample set, where yi is the label, n is the number of training samples, and w mi It is the weight of the i-th sample in the m-th iteration.

[0037] Furthermore, the method for calculating the weights of the SVM classifier in step S604 is as follows;

[0038]

[0039] In the formula, α m e represents the weights of the weak classifier. m The training error rate is m, and the number of iterations is m.

[0040] e m When ≤1 / 2, α m ≥0, and α m With e m As the value decreases, the value increases. The later classifier is more focused on correctly classifying the data that the previous classifier misclassified. Each weak classifier only focuses on a small portion of the data. They work together, and when voting, weights are assigned based on the classification error rate of the weak classifiers. The lower the error rate, the higher the weight.

[0041] Furthermore, the method for updating the weights in the training sample set in step S606 is as follows:

[0042] D m+1 =(w m+1,1 ,w m+1,2 ,…w m+1,i …,w m+1,n )

[0043]

[0044]

[0045] Where m = 1, 2, ..., M represents the current iteration round, w represents the sample weight, and α m y represents the weights of the weak classifier.i For the label of the sample, h m For d t The SVM weak classifier is obtained by training an SVM weak classifier with σ as the parameter, using the training sample set as the training sample set. i Z represents sample i. m It is the normalization factor that makes D m+1 It reverts to a probability distribution.

[0046] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0047] 1. The method of the present invention constructs a ventilation hole recognition model based on runtime features and machine learning, which can quickly and efficiently identify ventilation holes and save the image directly above the ventilation hole for the next step of detection, and can simultaneously meet the needs of recognition and localization of multiple types of ventilation holes;

[0048] 2. The method of the present invention uses Brenner gradient calculation, Tenengrad gradient calculation, Laplacian gradient calculation and variance calculation to extract features as model input. This dataset is more multidimensional than a single dataset. At the same time, the use of SVM and Adaboost composite recognition model can greatly improve the accuracy of algorithm recognition. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a ventilation hole identification and positioning method based on the Adaboost algorithm according to the present invention.

[0050] Figure 2 This is a flowchart illustrating step S600 of a ventilation hole identification and positioning method based on the Adaboost algorithm according to the present invention.

[0051] Figure 3 This is a schematic diagram of the robot collecting images of the ventilation holes of the copper busbar in step S100 of an embodiment of the ventilation hole identification and localization method based on the Adaboost algorithm of the present invention.

[0052] In all the accompanying drawings, the same reference numerals denote the same technical features, specifically: 1-inspection robot, 2-camera number one, 3-camera number two, 4-ventilation hole, 5-rotor. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0054] A method for identifying and locating ventilation openings based on the Adaboost algorithm includes the following steps:

[0055] S100: The inspection robot with visual inspection function is placed into the generator chamber. With the help of positioning, the two cameras on the abdomen of the inspection robot are aligned with the rotor slot wedge. Then, the operator remotely controls the inspection robot to crawl at a constant speed for several minutes. The two cameras on the abdomen collect video images of elliptical ventilation holes, single-row circular ventilation holes, and double-row circular ventilation holes respectively.

[0056] Assisted positioning is achieved through an auxiliary delivery device, which can perform circumferential rotation positioning. The front end of the auxiliary delivery device has a gripper mechanism in which the detection robot is placed. As the auxiliary delivery device rotates, it drives the detection robot to rotate. When the two cameras on the belly of the detection robot are facing the rotor slot wedge, the robot can be delivered.

[0057] Generators typically consist of components such as a stator, rotor, end covers, and bearings, forming a concentric cylindrical structure. The rotor core is in the center, and the stator core is on the outside. The generator rotor is a slender cylinder with slots. Large generators often use copper busbar ventilation holes. Due to the special structure of the copper busbar ventilation holes, two cameras must be used to simultaneously observe the upper and lower surfaces of the rotor ventilation holes. For the upper surface, an autofocus camera is selected. Through multiple operations, the robot's forward and backward movements are detected to observe the rotor surface in real time. After the rotor ventilation holes enter the frame, the camera identifies and positions the surface and bottom of the ventilation holes, capturing clear images of both the surface and bottom of the rotor ventilation holes.

[0058] When not operating above the hole, a clear image cannot be obtained because the distance between the camera lens and the rotor core surface is less than the minimum working distance. After the rotor ventilation hole enters the frame, the image of the upper surface of the rotor can be clearly observed after automatic focusing. The image is clearest when the camera lens is directly above the ventilation hole. After saving, a predetermined program is triggered. A preferred program setting is: after 1 second, the second fixed-focus camera arrives directly above the hole, pauses for 0.5 seconds, of which 160ms is used for exposure to obtain a clear image of the bottom of the ventilation hole, thus completing the image acquisition of the ventilation hole.

[0059] The robot crawls on the surface of the rotor with its abdomen facing the rotor. The first camera first locates the full outline of the ventilation, and the second camera can then see the bottom of the ventilation hole. Structurally, two cameras are used and arranged side by side on the bottom of the inspection robot. The first side camera observes the upper surface of the rotor ventilation hole, and the second camera observes the bottom of the rotor ventilation hole. The supplementary lighting module of the second camera can rotate and extend up and down to ensure sufficient light in the field of view of the bottom of the rotor ventilation hole.

[0060] S200: Process the collected data;

[0061] (To ensure operational efficiency and the sensitivity of evaluation results to changes in image features,) Brenner gradient calculation, Tenengrad gradient calculation, Laplacian gradient calculation, and variance calculation are performed frame by frame on the acquired video images, and normalization processing is performed.

[0062] (1) Brenner gradient function

[0063] The Brenner gradient function calculates the square of the gray-level difference between two adjacent pixels. The function is defined as follows:

[0064] D(f)=∑ y ∑ x |f(x+2,y)-f(x,y)| 2

[0065] Where: f(x,y) represents the gray value of the pixel (x,y) corresponding to image f, and D(f) is the image sharpness calculation result;

[0066] (2) Tenengrad gradient function

[0067] The Tenengrad gradient function uses the Sobel operator to extract the gradient values ​​in the horizontal and vertical directions respectively. The image sharpness based on the Tenengrad gradient function is defined as follows:

[0068] D(f)=Σ y Σ x |G(x,y)|(G(x,y)>T)

[0069]

[0070] Among them G x and G y Let x and y represent the convolution of the Sobel operators in the horizontal and vertical directions at the pixel (x, y) corresponding to image f, respectively. The following Sobel operator template is used to detect edges, and T is the given edge detection threshold.

[0071] The Sobel operator template used is defined as follows:

[0072]

[0073] (3) Laplacian gradient function

[0074] The image sharpness based on the Laplacian gradient function is defined as follows: The convolution of the Laplacian operator is calculated for each pixel, and then the results are accumulated pixel by pixel.

[0075] D(f) = ∑, ∑ x |G(x,y)|(G(x,y)>T)

[0076] Where G(x,y) represents the convolution of the Laplacian operator at pixel (x,y) corresponding to image f, and T is the given edge detection threshold;

[0077] The Laplacian operator is a second-order differential operator in n-dimensional Euclidean space, defined as follows:

[0078]

[0079] The Laplacian operator template used is defined as follows:

[0080]

[0081] (4) Variance calculation

[0082] The variance function is as follows:

[0083] D(f)=∑ y ∑ x |f(x,y)-μ)| 2

[0084] Where: f(x,y) represents the gray value of the pixel (x,y) corresponding to image f, D(f) is the image sharpness calculation result, and μ is the average gray value of the entire image.

[0085] S300: When the ventilation hole is located in the center of the image, the four gradient values ​​in the feature value all have a relative maximum value, which can well reflect the position information of the current camera relative to the slot wedge. It can be put into the same network model for training and recognition. However, there is also a multi-peak phenomenon in the peak value. Therefore, a single threshold cannot be simply used as the judgment standard. The method based on SVM-Adaboost (which has multiple threshold intelligent tuning characteristics) is selected to recognize the ventilation hole image.

[0086] S400: Image Labeling: Each image is labeled. Images containing ventilation holes are labeled as 1, and images without ventilation holes are labeled as -1. 50% to 70% of the data are randomly selected as the training set, and the remaining data is used as the test set. The training set will be used for simulation training of the ventilation hole recognition model, and the test set will be used for model optimization testing.

[0087] S500: When the ventilation hole is about to enter or exit the field of view, or when the image edge is only 0.3-0.7mm, the camera continuously captures five images. When the recognition result is label 1 twice consecutively, it is determined that the position is above the ventilation hole, and the image gradient value is recorded and the image of this area is saved. When the recognition result is label -1 more than twice consecutively, it is determined that the position has left the ventilation hole, and the image with the largest Laplacian gradient value among the previously saved images is selected as the positioning result for the next step of processing.

[0088] S600: Build an SVM-AdaBoost model, import the training set data into the model for training, test the generated model using the test set data, and visualize the model's training accuracy; calculate the training error of the weak classifier, i.e., the sum of the weights of the misclassified samples, to obtain the final classification criterion. The specific steps are as follows:

[0089] S601: Define a training dataset, T = {(x1,y1),(x2,y2)...(x...} n ,y n )},(x i ,y i ) is one of the training data, x i Representing a sample, y i Let {1, -1} be the label of this sample.

[0090] Define a training dataset T = {(x1,y1),(x2,y2)...(x...} n ,y n )},(x i ,y i ) is one of the training data, x i Representing sample i, y i Let {1, -1} be the label of the sample, and w be the sample weight;

[0091] In the first iteration, for all n training samples, the weight of each sample is set to 1 / n, and the weight distribution is D1 = (w 11 ,w 12 ,…w 1i …,w 1n ),in:

[0092] S602: Learn SVM weak classifiers on training samples. Sample the training set according to the sample weights to obtain the training set of component SVM classifiers. Train an SVM weak classifier with radial basis function kernel function, using RBF-SVM as the weak classifier.

[0093] Using RBF-SVM as a weak classifier, the RBF kernel function can be expressed as:

[0094]

[0095] In the formula, x n -c i Let be the Euclidean norm, ci be the center of the Green's function, and σ be the variance of the Green's function.

[0096] S603: Calculate the standard deviation of the training set as the parameter σ of the SVM weak classifier;

[0097] σ=sqrt(mean(var(d t )))

[0098] In the formula, d t For the training sample set, sqrt represents the square root operation, mean represents the arithmetic mean calculation, and var represents the variance calculation;

[0099] S604: Classify each feature using an SVM weak classifier, and calculate the training error rate of each SVM classification result, which is the sum of the weights of the misclassified samples.

[0100]

[0101] In the formula, m = 1, 2, ..., M represents the current iteration round number, e m h is the training error rate. m For d t The SVM weak classifier is obtained by training an SVM weak classifier with σ as the parameter, using the training sample set as the training sample set, where yi is the label, n is the number of training samples, and w mi It is the weight of the i-th sample in the m-th iteration.

[0102] S605: Calculate the weights of the SVM classifier;

[0103]

[0104] In the formula, α m e represents the weights of the weak classifier. m The training error rate;

[0105] e m When ≤1 / 2, α m ≥0, and α m With em As the value decreases, the value increases. The later classifier is more focused on correctly classifying the data that the previous classifier misclassified. Each weak classifier only focuses on a small portion of the data. They work together, and when voting, weights are assigned based on the classification error rate of the weak classifiers. The lower the error rate, the higher the weight.

[0106] S606: If the training error rate in step S602 is greater than 50%, retrain; if the training error rate is less than 50%, update the weights in the training sample set and proceed to step S601 for the next iteration. When the number of iterations reaches the set value M, proceed to step S606.

[0107] The method for updating the weights in the training sample set is as follows:

[0108] D m+1 =(w m+1,1 ,w m+1,2 ,…w m+1,i …,w m+1,n )

[0109]

[0110]

[0111] Where m = 1, 2, ..., M represents the current iteration round, w represents the sample weight, and α m y represents the weights of the weak classifier. i For the label of the sample, h m For d t The SVM weak classifier is obtained by training an SVM weak classifier with σ as the parameter, using the training sample set as the training sample set. i Z represents sample i. m It is the normalization factor that makes D m+1 It reverts to a probability distribution.

[0112] S607: After M iterations, sum the corresponding weights and obtain the decision function value of the strongest classifier through voting.

[0113]

[0114] Among them, h m α is the SVM weak classifier obtained by training an SVM weak classifier with dt as the training sample set and σ as the parameter; m The values ​​are the weights of the weak classifier, m = 1, 2, ..., M, representing the current iteration round number and M the iteration count. sign() is the sign function. When the value inside the parentheses is greater than "0", the output value is set to "1", indicating that the judgment result is correct. When the value inside the parentheses is not greater than "0", the output value is set to "-1", indicating that the judgment result is incorrect.

[0115] S700: Experimental Analysis

[0116] Ventilation hole image data were collected on the constructed experimental platform. After grayscale processing, the gradient data were calculated. There were 700 elliptical hole-type slot wedge images, 400 through-hole-type slot wedge images, 600 circular hole-type slot wedge images, and 500 background images, totaling 2200 images. The images were allocated according to Table 1 during training.

[0117] Table 1

[0118]

[0119] The number of iterations was set to 3000, the learning rate was set to 0.01, and the Tenengrad gradient, Laplacian gradient and variance of the image were used as input parameters. Each image was labeled as 1 (the image contains ventilation holes) or -1 (the image does not contain ventilation holes, i.e., pure background image). After training for 2000 times, the algorithm based on SAMMER.R began to converge and reached a stable state at around 3000 times.

[0120] In actual online testing, false recognition is prone to occur when ventilation holes are about to enter or exit the field of view. However, this is suppressed by the adjustment mechanism in the positioning algorithm. The positioning algorithm performs well, and the positioning accuracy of the three types of ventilation holes can reach more than 95%, with an average recognition time of 0.02 seconds.

[0121] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and locating air vents based on an Adaboost algorithm, characterized in that, The method comprises the following steps: S100: image acquisition: a detection robot with visual detection function is put into the generator bore, auxiliary positioning is performed to make the two cameras on the abdomen of the detection robot face the rotor slot wedge, the detection robot is run for several minutes, and video images of the elliptical ventilation hole, the single-row circular ventilation hole and the double-row circular ventilation hole are acquired by the two cameras on the abdomen; when the camera lens is located directly above the ventilation hole, the image is the clearest, and after saving, a predetermined program is triggered, the program is set as: 1 second later, the second fixed-focus camera reaches directly above the hole, and pauses for 0.5 seconds, of which 160 ms is used for exposure to obtain a clear image of the bottom of the ventilation hole, and the image acquisition of the ventilation hole is completed; The robot climbs on the rotor surface, the abdomen faces the rotor, and the first camera is positioned on the ventilation profile, and the second camera can see the bottom of the ventilation hole; In terms of structure, two cameras are used and arranged side by side at the bottom of the detection robot, the first side camera observes the upper surface of the rotor ventilation hole, and the second camera observes the bottom of the rotor ventilation hole; the light supplement module matched with the second camera can rotate and stretch up and down to ensure that the light in the field of view of the bottom of the rotor ventilation hole is sufficient; S200: processing the acquired images: the acquired video images are processed frame by frame by Brenner gradient calculation, Tenengrad gradient calculation, Laplacian gradient calculation and variance calculation, and normalized; S300: image recognition: when the ventilation hole is located in the center of the image, the four gradient values in the characteristic value all appear relative maximum values, and a method based on SVM-Adaboost with multiple threshold intelligent setting characteristics is selected to recognize the ventilation hole image; S400: image labeling: each image is labeled, the image containing the ventilation hole is labeled as 1, and the image not containing the ventilation hole is labeled as -1; 50% to 70% of the data are randomly selected as the training set of the model, and the remaining data are used as the test set; S500: the position of the image in which the ventilation hole is about to enter or exit the field of view or the image edge is only 0.3 to 0.7 mm, the camera continuously acquires five images, and then the ventilation hole is positioned according to the label; S600: an SVM-AdaBoost model is established, the training set data are imported into the model for training, the test set data are used to test the generated model, and the training accuracy of the model is visualized; The training error of the weak classifier is calculated to obtain the final classification judgment standard.

2. The method of claim 1, wherein the method is based on an Adaboost algorithm. The auxiliary positioning in step S100 is realized by an auxiliary placing device, the auxiliary placing device can be rotationally positioned, the front end of the auxiliary placing device has a gripping mechanism, the detection robot is placed in the gripping mechanism, the auxiliary placing device rotates while driving the detection robot to rotate, the detection robot is rotated to the position in which the two cameras on the abdomen face the rotor slot wedge, and the robot placing action can be performed.

3. The method of claim 1, wherein the method is based on an Adaboost algorithm. Large generator uses copper row vent hole, when the identification object is the copper row vent hole of large generator, two cameras must be equipped to observe the upper surface and bottom of rotor vent hole due to the special structure of copper row vent hole in step S100, an automatic focusing camera is selected for the upper surface of generator rotor, the forward and backward movement of detection robot is detected through multiple operations, the surface condition of rotor is observed in real time, the surface and bottom of rotor vent hole are identified and positioned after the rotor vent hole enters the picture, and the clear image of surface and bottom of rotor vent hole is collected.

4. The method of claim 1, wherein the method is based on an Adaboost algorithm. The method for vent hole positioning according to the label in step S500 is as follows: when the identification result is label 1 for two times in succession, it is determined that the position is above the vent hole at this time, the image gradient value is started to be recorded, and the image in the region is saved; when the identification result is label-1 for more than two times in succession, it is determined that the position has left the upper surface of the vent hole, the image with the maximum Laplacian gradient value in the saved image is selected as the positioning result, and is used for the next step processing.

5. The method of claim 1, wherein the method is based on an Adaboost algorithm. The specific steps of step S600 are as follows: Step S601: define a set of training data, when the first iteration, set the weight of each sample as 1 / n for all n training samples; Step S602: learn the SVM weak classifier for the training sample, sample in the training set according to the sample weight, obtain the training set of component SVM classifier, train an SVM weak classifier with radial basis kernel function, and adopt RBF-SVM as the weak classifier; Step S603: calculate the standard deviation of the training set as the parameter σ of the SVM weak classifier; sigma = sqrt(mean(var(d t ))) In the formula, d t is a training sample set, sqrt represents a square root operation, mean represents an arithmetic average calculation, and var represents a variance calculation; Step S604: classify each feature with the SVM weak classifier, and calculate the training error rate of each SVM classification result, that is, the sum of the weights of the misclassified samples; Step S605: calculate the weight of the SVM classifier; Step S606: when the training error rate in step S602 is greater than 50%, retrain; When the training error rate is less than 50%, update the weight in the training sample set, re-enter steps S602-S606, and perform the next iteration, and when the iteration number reaches the set value M, enter step S607; Step S607: after M iterations, add the corresponding weights, and obtain the decision function value of the strongest classifier through voting method; wherein h m is the d t is the SVM weak classifier trained by the training sample set with the parameter σ; α m is the weight of the weak classifier, m = 1, 2, …, M, represents the current iteration number, and sign() is a sign function. When the value in the bracket is greater than "0", the output value is set to "1", representing that the judgment result is correct. When the value in the bracket is not greater than "0", the output value is set to "-1", representing that the judgment result is incorrect.

6. The method of claim 5, wherein the method is based on an Adaboost algorithm. The specific method for defining the training data set in step S601 is as follows: Define a training dataset T = {(x1,y1),(x2,y2)...(x...} n ,y n )},(x i ,y i ) is one of the training data, x i Representing sample i, y i Let {1, -1} be the label of the sample, and w be the sample weight; in the first iteration, for all n training samples... In this case, each sample weight is set to 1 / n, and the weight distribution D1 = (w 11 ,w 12 ,…w 1i …,w 1n ), where:

7. The method of claim 5, wherein the method is based on an Adaboost algorithm. The classifier formula used in step S602 is as follows: RBF-SVM is used as the weak classifier, and the RBF kernel function R(x n -c i ) can be expressed as: where x n -c i is the Euclidean norm, c; is the center of the Green function, and σ is the variance of the Green function.

8. The method of claim 6, wherein the method is based on an Adaboost algorithm. The method for calculating the training error rate of each SVM classification result in step S604, that is, the sum of the weights of the misclassified samples, is as follows: where m = 1, 2,..., M, represents the current iteration round, e m is the training error rate, h m is the training sample set, and σ is the parameter t is the SVM weak classifier obtained by training the SVM weak classifier with σ as the parameter, y i is the label, n is the training sample amount, and w mi is the weight of the i-th sample in the m-th iteration.

9. The method of claim 8, wherein the method is based on an Adaboost algorithm. The method for calculating the weight of the SVM classifier in step S604 is as follows: wherein α m is the weight of the weak classifier, e m is the training error rate, and m is the number of iterations. e m ≤1 / 2, a m ≥0, and a m increases as e m decreases, the latter classifier is more committed to the data that the former classifier misclassified, each weak classifier only focuses on a small part of the data, and works together to finally vote according to the classification error rate of the weak classifier, and the lower the error rate, the higher the weight.

10. The method of claim 9, wherein the method is based on an Adaboost algorithm. The method for updating the weight in the training sample set in step S606 is as follows: D m+1 = (W m+1,1 , w m+1,2 , …w m+1,i …,w m+1,n ) where m = 1, 2,..., M represents the current iteration round, w represents the sample weight, a m is the weight of the weak classifier, y i is the label of the sample, h m is the d t is the training sample set, and the SVM weak classifier obtained by training the SVM weak classifier with σ as the parameter, x i represents the i-th sample, Z m is the normalization factor, so that D m+1 becomes a probability distribution again.

Images