Multi-parameter optimal setpoint analysis method and apparatus based on high-frequency machine vision learning

Abstract

The present invention relates to the technical field of image data processing. Disclosed are a multi-parameter optimal setpoint analysis method and apparatus based on high-frequency machine vision learning. The method comprises: S1, collecting data sources; S2, dividing the collected data sources into a test set and a training set; S3, determining influence factors; S4, performing correlation validation on the influence factors; S5, performing model construction; S6, performing model evaluation; and S7, performing model optimization. By means of collecting historical data of the same type of products and data of the same type of products in the same batch, the present invention improves the diversity of image data samples, thereby improving the generalization ability of a model, enhancing the robustness of the model, and improving the accuracy of model prediction and determination; and influence factors are fully considered during image data collection and model construction, and by means of performing correlation validation on the influence factors, different variables can be screened before modeling, such that the variables that make greater contributions to a model are selected, thereby improving the efficiency and accuracy of the model.

Description

Method and device for analyzing multi-parameter optimal setting value based on high-frequency machine vision learning TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, in particular to a method for analyzing multi-parameter optimal setting value based on high-frequency machine vision learning. BACKGROUND

[0002] High-frequency machine vision learning combines machine learning and computer vision technology, which can analyze and process a large amount of image data in real time. The method for analyzing multi-parameter optimal setting value aims to optimize the combination of multiple parameters to achieve the best performance of the system. By combining the two, efficient decision support can be achieved in complex application scenarios.

[0003] However, in the prior art, high-frequency machine learning and computer vision technology are used to collect image information of products, and then the collected parameters are compared and analyzed to determine the quality of the products. When detecting the quality of products in automated production, the model may overfit the training set due to insufficient training data or improper feature selection, making it difficult to set optimal parameters. In addition, due to the noise of image sensors and environmental interference, the quality of the collected images decreases. When collecting images of products on the production line, the images collected by the camera are prone to blur when the products follow the production line in motion, which also leads to a decrease in the quality of the collected images, ultimately resulting in a decrease in the accuracy of product quality detection.

[0004] SUMMARY

[0005] The present application aims to provide a method for analyzing multi-parameter optimal setting value based on high-frequency machine vision learning to solve the problem of overfitting the training set due to insufficient training data or improper feature selection, making it difficult to set optimal parameters, and the quality of the collected images decreases due to the noise of image sensors and environmental interference. When collecting images of products on the production line, the images collected by the camera are prone to blur when the products follow the production line in motion, which also leads to a decrease in the quality of the collected images, ultimately resulting in a decrease in the accuracy of product quality detection.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a method for analyzing multi-parameter optimal setting value based on high-frequency machine vision learning, comprising the following contents:

[0007] S1, collecting data sources: the data sources are composed of historical data of the same type of products and data of the same batch of the same type of products;

[0008] S2, dividing the collected data sources into a test set and a training set;

[0009] S3, determining an influence factor: dividing the influence factor according to human factors, environmental factors and equipment factors;

[0010] S4, verifying the correlation of the influence factor, the correlation verification adopts a variance inflation factor method to perform multiple collinearity verification, and filters factors with a VIF value less than 10, and the calculation formula of the VIF is as follows:

[0011] In the formula, R i 2 is a decision coefficient between the ith influence factor and other influence factors;

[0012] S5, model construction: using the collected samples of the training set to construct a model, and the random forest algorithm is used to select an optimal feature from the training set sample data as the splitting attribute of the node;

[0013] S6, model evaluation: using the collected samples of the test set to evaluate the model constructed;

[0014] S7, model optimization: used for optimizing the model that fails the model evaluation.

[0015] Preferably, in step S1, the ratio of the historical data of the same type of product in the data source to the same batch data of the same type of product is 1:1.

[0016] Preferably, in step S1, the same batch data of the same type of product is collected by using an image acquisition device of a production line.

[0017] Preferably, in step S2, the data sample ratio of the test set and the training set is 3:7.

[0018] Preferably, in step S5, the random forest algorithm includes the following contents:

[0019] S51, self-service sampling: randomly sampling k samples from the collected samples D of the training set with replacement as the training set K of a decision tree;

[0020] S52, random selection of features: randomly selecting m features (m << M) from the candidate M features as the candidate feature subset for decision-making at the current node; then selecting an optimal feature from the selected m features as the splitting attribute of the node using an information gain strategy, and the information gain Gain(A, K) of the feature A for the data set K is the difference between the entropy Ent(K) of the data set and the conditional entropy CEnt(A, K) of the feature A, and the formula is as follows: Ent(K) = -p1*log2p1-p2*log2p2 Gain(A, K) = Ent(K) - CEnt(A, K)

[0021] Wherein: p1, p2 are the proportion of unqualified and qualified samples in the current data set K; K v Indicates the number of samples of feature A taking value v;

[0022] S53, decision tree growth: during the formation of the decision tree, each node repeats step S2 for splitting, until there are no new features, the splitting ends;

[0023] S54, building a random forest: according to steps S1 to S3, N decision trees are built;

[0024] S55, the final prediction result is classified by voting the prediction results of all decision trees to determine an optimal feature parameter.

[0025] Preferably, in step S53, the decision tree growth adopts XGBoost algorithm to reduce the complexity of the decision tree, and obtains the optimal objective function Obj (k) of the decision tree. (k) As follows:

[0026] Wherein: is the sum of all samples g i belonging to the jth leaf node; g i is the first order derivative of the loss function; is the sum of all samples hi belonging to the jth leaf node; hi is the second order derivative of the loss function; I j ={i|q(x i )=j}, q(x i )=j indicates that the ith sample is located in the jth leaf node; T is the number of leaf nodes; w j is the weight value of the leaf node; λ, γ are hyperparameters for controlling the punishment degree;

[0027] The optimal objective function is used to measure the standard of tree structure, and the smaller the value is, the better the structure of the tree is.

[0028] Preferably, in step S6, the model evaluation adopts a confusion matrix and a ROC curve to evaluate the model constructed by the model, the confusion matrix is used to calculate the accuracy, precision and recall of the obtained data to evaluate the model, and the ROC curve is used to quantitatively judge the accuracy of the model.

[0029] Preferably, in step S7, the model optimization adopts a Bayesian optimization algorithm to optimize the constructed model.

[0030] Preferably, in step S7, the Bayesian optimization algorithm optimizes the constructed model, including the following contents:

[0031] S71, first, the prior knowledge is established by approximating the objective function through a Gaussian regression model;

[0032] S72, then, the prior knowledge is updated by observing the output of the objective function at different input points to form a posterior distribution;

[0033] S73, based on the posterior distribution, the next sampling is performed by using a PI acquisition function, the sampled data is added to the data set, the probability agent function is calculated, and the above steps are repeated, and the global optimal solution is selected after iteration to a set number of times.

[0034] The device for setting multiple parameters optimally based on high-frequency machine vision learning comprises a mounting mechanism, an adjusting mechanism and a cleaning mechanism, the mounting mechanism comprises a first sliding rail, a first sliding block is slidably connected in the first sliding rail, a first servo motor is fixedly connected to the top of the first sliding rail, a first adjusting screw rod is fixedly connected to the output end of the first servo motor, the first adjusting screw rod is in threaded connection with the first sliding block, a second sliding rail is fixedly connected to the side surface of the first sliding block, a second sliding block is slidably connected in the second sliding rail, a second servo motor is fixedly connected to one end of the second sliding rail, a second adjusting screw rod is fixedly connected to the output end of the second servo motor, the second adjusting screw rod is in threaded connection with the second sliding block, the adjusting mechanism comprises a third servo motor, the third servo motor is fixedly connected to the side surface of the second sliding block, a first connecting frame is fixedly connected to the output end of the third servo motor, a fourth servo motor is fixedly connected to the lower portion of the first connecting frame, a second connecting frame is fixedly connected to the output end of the fourth servo motor, a fifth servo motor is fixedly connected to the side surface of the second connecting frame, a mounting frame is fixedly connected to the output end of the fifth servo motor, an image acquisition device is mounted on the surface of the mounting frame, the cleaning mechanism comprises a fixed frame, the fixed frame is fixedly connected to the bottom of the second sliding rail, a mounting plate is fixedly connected to one end of the fixed frame, a plug-in board is inserted into the mounting plate, a wiping cotton layer is fixedly connected to the upper portion of the plug-in board.

[0035] Compared with the prior art, the device has the following beneficial effects:

[0036] 1. In this invention, by collecting historical data and batch data of similar products, the diversity of image data samples is improved, thereby enhancing the generalization ability, robustness, and accuracy of prediction and judgment of the model. When collecting image data and building the model, influencing factors are fully considered, and by verifying the correlation of influencing factors, different variables can be screened before modeling, and those variables that contribute more to the model can be selected, thereby improving the efficiency and accuracy of the model. It can also guide the optimization and adjustment of the model. Finally, the model is built through the random forest algorithm, and the optimal parameters are determined.

[0037] 2. In this invention, when building a model, the built model is evaluated through model evaluation, and the accuracy of model detection is quantitatively judged, which plays a detection role for the established model. When the established model does not reach the optimal setting value, the unqualified model is optimized through model optimization, and the optimized model is iteratively evaluated through the model evaluation step so that the obtained parameter setting value is the optimal setting value.

[0038] 3. In this invention, by precisely adjusting the installation position of the image acquisition device on the production line, the accuracy of image data acquisition by the image acquisition device is improved. When dust adheres to the lens of the image acquisition device during the acquisition process, the second servo motor drives the second slider to slide along the second slide rail, so that the lens of the image acquisition device is cleaned by wiping the surface of the cotton layer, thus ensuring the accuracy of image data acquisition by the image acquisition device. Attached Figure Description

[0039] Figure 1 is a flowchart of the analysis method and apparatus for optimal setting values ​​of multiple parameters based on high-frequency machine vision learning according to the present invention.

[0040] Figure 2 is a three-dimensional structural diagram of the installation mechanism in the analysis method and device for multi-parameter optimal setting values ​​based on high-frequency machine vision learning of the present invention.

[0041] Figure 3 is a three-dimensional structural diagram of the cleaning mechanism in the analysis method and device for multi-parameter optimal setting values ​​based on high-frequency machine vision learning of the present invention.

[0042] Figure 4 is a three-dimensional structural diagram of the adjustment mechanism in the analysis method and device for multi-parameter optimal setpoints based on high-frequency machine vision learning of the present invention.

[0043] Figure 5 is a front view schematic diagram of the image acquisition device in the analysis method and device for optimal setting values ​​of multiple parameters based on high-frequency machine vision learning of the present invention.

[0044] In the figure: 1, mounting mechanism; 11, No. 1 slide rail; 12, No. 1 sliding block; 13, No. 1 servo motor; 14, No. 1 adjusting screw; 15, No. 2 slide rail; 16, No. 2 sliding block; 17, No. 2 servo motor; 18, No. 2 adjusting screw; 2, adjusting mechanism; 21, No. 3 servo motor; 22, No. 1 connecting frame; 23, No. 4 servo motor; 24, No. 2 connecting frame; 25, No. 5 servo motor; 26, mounting frame; 27, image acquisition device; 3, cleaning mechanism; 31, fixed frame; 32, mounting plate; 33, plugboard; 34, wiping cotton layer. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] Referring to FIG. 1, the method for analyzing multi-parameter optimal set value based on high-frequency machine vision learning includes the following contents:

[0047] Step one, collecting data sources: the data sources are composed of historical data of the same type of product and same batch data of the same type of product, the ratio of the historical data of the same type of product to the same batch data of the same type of product in the data source is 1:1, and the same batch data of the same type of product is collected by using the image acquisition device of the production line.

[0048] Step two, dividing the collected data sources into test set and training set, the data sample ratio of the test set and the training set is 3:7.

[0049] Step three, determining the influence factor: dividing the influence factor according to human factors, environmental factors and equipment factors.

[0050] Step four, verifying the correlation of the influence factor, the correlation verification adopts the method of variance inflation factor to verify the multicollinearity, and filters the factors with VIF value less than 10, and the calculation formula of VIF is as follows:

[0051] In the formula: R i 2 is the determination coefficient between the ith influence factor and other influence factors.

[0052] Step five, model construction: using the collected samples of the training set to construct the model, the random forest algorithm is used for model construction, the random forest algorithm is used to select an optimal feature from the training set sample data as the splitting attribute of the node, and the random forest algorithm includes the following contents:

[0053] S51, self-service sampling: randomly sampling k samples from the training set D with replacement as the training set K of a decision tree;

[0054] S52, feature random selection: randomly selecting m features (m << M) from the candidate M features as the candidate feature subset for decision-making at the current node; then selecting an optimal feature as the splitting attribute of the node from the selected m features using the information gain strategy. The information gain Gain(A, K) of feature A for the data set K is the difference between the entropy Ent(K) of the data set and the conditional entropy CEnt(A, K) of feature A, as follows: Gain(A, K) = Ent(K) - CEnt(A, K)

[0055] where: p1, p2 are the proportions of unqualified and qualified samples in the current data set K; v represents the number of samples with feature A taking value v;

[0056] S53, decision tree growth: during the formation of the decision tree, each node repeats step S2 for splitting until there are no new features, and the splitting ends;

[0057] S54, building a random forest: according to steps S1 to S3, N decision trees are built;

[0058] S55, the final prediction result is classified by voting the prediction results of all decision trees to determine an optimal feature parameter;

[0059] The XGBoost algorithm is used to reduce the complexity of the decision tree to obtain the optimal objective function Obj (k) of the decision tree as follows: (k)

[0060] where: is the sum of all samples g i belonging to the jth leaf node; g i is the first-order derivative of the loss function; is the sum of all samples hi belonging to the jth leaf node; hi is the second-order derivative of the loss function; I j = {i | q(x i ) = j}, q(x i ) = j indicates that the ith sample is located in the jth leaf node; T is the number of leaf nodes; w j is the weight value of the leaf node; λ, γ are hyperparameters for controlling the punishment degree; ​

[0061] The optimal objective function is used to measure the standard of tree structure, and the smaller value indicates the better structure of the tree.

[0062] Step six, model evaluation: using the test set of collected samples to evaluate the model constructed by the model, the model evaluation uses the confusion matrix and the ROC curve to evaluate the model constructed by the model, the confusion matrix is used to calculate the accuracy, precision and recall of the obtained data to evaluate the model, and the ROC curve is used to quantitatively judge the accuracy of the model detection.

[0063] Step seven, model optimization: used for optimizing the unqualified model of model evaluation, the model optimization uses the Bayesian optimization algorithm to optimize the constructed model, and the Bayesian optimization algorithm includes the following contents:

[0064] S71, first, the Gaussian regression model is used to approximate the objective function to establish prior knowledge;

[0065] S72, then, by observing the output of the objective function at different input points, the prior knowledge is updated to form the posterior distribution;

[0066] S73, based on the posterior distribution, the next sampling is carried out by using the PI acquisition function, the sampled data is added to the data set, the probability agent function is calculated, and the above steps are repeated, and the global optimal solution is selected after iteration to the set number of times.

[0067] As shown in Figures 2-5: the device for multi-parameter optimal setting value based on high-frequency machine vision learning includes a mounting mechanism 1, an adjusting mechanism 2 and a cleaning mechanism 3, the mounting mechanism 1 includes a first sliding rail 11, a first sliding block 12 is slidably connected in the first sliding rail 11, and a first servo motor 13 is fixedly connected to the top of the first sliding rail 11, a first adjusting lead screw 14 is fixedly connected to the output end of the first servo motor 13, and the first adjusting lead screw 14 is threadedly connected with the first sliding block 12, a second sliding rail 15 is fixedly connected to the side of the first sliding block 12, a second sliding block 16 is slidably connected in the second sliding rail 15, and a second servo motor 17 is fixedly connected to one end of the second sliding rail 15, a second adjusting lead screw 18 is fixedly connected to the output end of the second servo motor 17, and the second adjusting lead screw 18 is threadedly connected with the second sliding block 16, the adjusting mechanism 2 includes a third servo motor 21, the third servo motor 21 is fixedly connected to the side of the second sliding block 16, and a first connecting frame 22 is fixedly connected to the output end of the third servo motor 21, a fourth servo motor 23 is fixedly connected to the lower part of the first connecting frame 22, a second connecting frame 24 is fixedly connected to the output end of the fourth servo motor 23, a fifth servo motor 25 is fixedly connected to the side of the second connecting frame 24, and a mounting frame 26 is fixedly connected to the output end of the fifth servo motor 25, an image acquisition device 27 is mounted on the surface of the mounting frame 26, and the cleaning mechanism 3 includes a fixed frame 31, the fixed frame 31 is fixedly connected to the bottom of the second sliding rail 15, and a mounting plate 32 is fixedly connected to one end of the fixed frame 31, a plug-in board 33 is inserted into the mounting plate 32, and a wiping cotton layer 34 is fixedly connected to the upper part of the plug-in board 33.

[0068] In use, the present application improves the diversity of image data samples by collecting historical data of the same type of product and data of the same batch of the same type of product, thereby improving the generalization ability of the model, enhancing the robustness of the model, and improving the accuracy of prediction and judgment of the model, fully considering the influence factors when collecting image data and constructing the model, and verifying the correlation of the influence factors, so that different variables can be screened before modeling, and variables that contribute more to the model are selected, thereby improving the efficiency and accuracy of the model, and guiding the optimization and adjustment of the model, and finally constructing the model through the random forest algorithm and determining the optimal parameters;

[0069] When constructing the model, the constructed model is evaluated through model evaluation, and the accuracy of model detection is quantitatively judged, so as to detect the established model, when the established model does not reach the optimal setting value, the model evaluation unqualified model is optimized through model optimization, and the optimized model is cyclically evaluated through the model evaluation step, so that the obtained parameter setting value is the optimal setting value;

[0070] When the image information of the same batch of the same type of products is collected by the image acquisition device of the production line, the image data of the same batch of the same type of products is collected by the image acquisition device 27. When the image acquisition device 27 is installed on the production line, the first servo motor 13 drives the first adjusting screw 14 to rotate, so that the first sliding block 12 slides along the first sliding rail 11, thereby adjusting the height of the second sliding rail 15, and further adjusting the installation height of the image acquisition device 27. Similarly, the second servo motor 17 drives the second adjusting screw 18 to rotate, and drives the second sliding block 16 to slide along the second sliding rail 15, thereby adjusting the left and right positions of the image acquisition device 27. The installation position of the image acquisition device 27 on the production line is accurate, which improves the accuracy of the image data collection of the image acquisition device 27. Similarly, the first connecting frame 22 is moved by the third servo motor 21, the second connecting frame 24 is moved by the fourth servo motor 23, and the mounting frame 26 is moved by the fifth servo motor 25, which further accurately adjusts the installation position of the image acquisition device 27, so that the image acquisition device 27 is in the best image data collection position. At the same time, by adjusting the position of the image acquisition device 27, it can accurately collect image data of different products.

[0071] When the lens of the image acquisition device 27 is attached with dust during the collection process, the second servo motor 17 drives the second sliding block 16 to slide along the second sliding rail 15, so that the lens of the image acquisition device 27 passes through the surface of the wiping cotton layer 34. The wiping cotton layer 34 cleans the lens of the image acquisition device 27 to ensure the accuracy of the image data collection of the image acquisition device 27. The wiping cotton layer 34 is installed in the plug-in board 33 and the mounting plate 32, which is convenient for cleaning or replacing the wiping cotton layer 34.

[0072] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent replacements to part of the technical features, as long as they are within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the scope of the present application shall be included in the protection scope of the present application.

Claims

1. A method for analyzing multi-parameter optimal set point based on high frequency machine vision learning, characterized in that, The method comprises the following steps: S1, collecting data sources: the data sources are composed of historical data of the same type of product and the same batch data of the same type of product; S2, dividing the collected data sources into a test set and a training set; S3, determining the influence factors: dividing the influence factors according to human factors, environmental factors and equipment factors; S4, correlation verification is performed on the influence factors: the correlation verification is performed on the influence factors by using the variance inflation factor method to verify the multiple collinearity, and factors with a VIF value less than 10 are screened, and the calculation formula of the VIF is as follows: wherein: R i 2 Ri is the coefficient of determination between the ith influencing factor and the other influencing factors; S5, model construction: using the collected samples of the training set to construct a model, wherein the model construction adopts a random forest algorithm, and the random forest algorithm is used to select an optimal feature from the training set sample data as a splitting attribute of the node; S6, model evaluation: using the collected samples of the test set to evaluate the model constructed in the model construction; S7, model optimization: used for optimizing the model that fails the model evaluation.

2. The method of claim 1, wherein the method is based on high frequency machine vision learning. In step S1, the ratio of the historical data of the same type of product to the same batch data of the same type of product in the data source is 1:

1.

3. The method of claim 1, wherein the method is based on high frequency machine vision learning. In step S1, the same batch data of the same type of product is collected by using an image acquisition device of a production line.

4. The method of claim 1, wherein the method is based on high frequency machine vision learning. In step S2, the data sample ratio of the test set and the training set is 3:

7.

5. The method of claim 1, wherein the method is based on high frequency machine vision learning for multi-parameter optimal setting value analysis. In step S5, the random forest algorithm comprises the following steps: S51, self-service sampling: randomly sampling k samples from the collected samples D of the training set with replacement as the training set K of a decision tree; S52, random selection of features: randomly selecting m features (m<<M) from the candidate M features as the candidate feature subset for decision-making at the current node; then selecting an optimal feature from the selected m features as the splitting attribute of the node by using an information gain strategy, wherein the information gain Gain(A, K) of the feature A for the data set K is the difference between the entropy Ent(K) of the data set and the conditional entropy CEnt(A, K) of the feature A, and the formula is as follows: Ent(K) = -p1*log2p1 - p2*log2p2 Gain(A, K) = Ent(K) - CEnt(A, K) In the formula, p1 and p2 are the proportions of unqualified and qualified samples in the current data set K; K v represents the number of samples with the value v of the feature A; S53, decision tree growth: in the process of forming a decision tree, each node repeats step S2 for splitting until there is no new feature, and the splitting ends; S54, constructing a random forest: constructing N decision trees according to steps S1 to S3; S55, the final prediction result is obtained by voting the prediction results of all decision trees to determine an optimal feature parameter.

6. The method of claim 5, wherein the method is based on high frequency machine vision learning. In step S53, the decision tree growth employs an XGBoost algorithm to reduce the complexity of the decision tree to obtain an optimal objective function Obj of the decision tree (k) , as follows: (k) ​ In the formulae: the sum of all samples g belonging to the jth leaf node i i the first derivative of the loss function; H j i∈Ij hi+λ, the sum of all samples hi belonging to the jth leaf node​​ hi is the second derivative of the loss function; I j = {i | q(x i ) = j}, q(x i ) = j means that the ith sample is located in the jth leaf node; T is the number of leaf nodes; w j is the weight value of the leaf node; λ, γ are hyperparameters for controlling the penalty strength; The optimal objective function is used to measure the standard of the tree structure, and the smaller the value is, the better the tree structure is.

7. The method of claim 1, wherein the method is based on high frequency machine vision learning for multi-parameter optimal setting value analysis. In step S6, the model evaluation adopts a confusion matrix and a ROC curve to evaluate the model constructed in the model construction, the confusion matrix is used to calculate the accuracy, precision and recall of the obtained data to evaluate the model, and the ROC curve is used to quantitatively judge the accuracy of the model detection.

8. The method of claim 1, wherein the method is based on high frequency machine vision learning for multi-parameter optimal setting value analysis. In step S7, the model optimization adopts a Bayesian optimization algorithm to optimize the constructed model.

9. The method of claim 8, wherein the method is based on high frequency machine vision learning for multi-parameter optimal setting value analysis. In step S7, the Bayesian optimization algorithm for optimizing the constructed model comprises the following steps: S71, first, a Gaussian regression model is used to approximate the objective function to establish prior knowledge; S72, then, the prior knowledge is updated by observing the output of the objective function at different input points to form a posterior distribution; S73, based on the posterior distribution, using PI acquisition function for next sampling, adding the sampled data to the data set, and then calculating the probability agent function, and repeating the above steps, and selecting the global optimal solution after iteration to the set number of times.

10. Apparatus for multi-parameter optimal set point based on high frequency machine vision learning, characterized in that The multi-parameter optimal setting value analysis method based on high-frequency machine vision learning in any one of claims 1-9 is used, comprising a mounting mechanism (1), an adjusting mechanism (2) and a cleaning mechanism (3), the mounting mechanism (1) comprises a first sliding rail (11), the first sliding rail (11) is slidably connected with a first sliding block (12) inside, and a first servo motor (13) is fixedly connected to the top of the first sliding rail (11), the output end of the first servo motor (13) is fixedly connected with a first adjusting lead screw (14), the first adjusting lead screw (14) is threadedly connected with the first sliding block (12), the side of the first sliding block (12) is fixedly connected with a second sliding rail (15), the second sliding rail (15) is slidably connected with a second sliding block (16) inside, and a second servo motor (17) is fixedly connected to one end of the second sliding rail (15), the output end of the second servo motor (17) is fixedly connected with a second adjusting lead screw (18), the second adjusting lead screw (18) is threadedly connected with the second sliding block (16), the adjusting mechanism (2) comprises a third servo motor (21), the third servo motor (21) is fixedly connected with the side of the second sliding block (16), and the output end of the third servo motor (21) is fixedly connected with a first connecting frame (22), the lower part of the first connecting frame (22) is fixedly connected with a fourth servo motor (23), the output end of the fourth servo motor (23) is fixedly connected with a second connecting frame (24), the side of the second connecting frame (24) is fixedly connected with a fifth servo motor (25), the output end of the fifth servo motor (25) is fixedly connected with a mounting frame (26), and the surface of the mounting frame (26) is mounted with an image acquisition device (27), the cleaning mechanism (3) comprises a fixed frame (31), the fixed frame (31) is fixedly connected with the bottom of the second sliding rail (15), one end of the fixed frame (31) is fixedly connected with a mounting plate (32), the mounting plate (32) is inserted with a plug-in board (33) inside, and the upper part of the plug-in board (33) is fixedly connected with a wiping cotton layer (34).

Images