Casting casting head intelligent cleaning and surface quality control system based on feature extraction

Abstract

The invention provides a casting head intelligent cleaning and surface quality control system based on feature extraction, and relates to the technical field of artificial intelligence and intelligent machining. The system comprises a three-dimensional reconstruction module, a feature planning module, a process execution module and a quality control module. Wherein the three-dimensional reconstruction module collects laser-visual data of a casting head area on the surface of a casting to realize three-dimensional reconstruction of the casting head area on the surface of the casting; the feature planning module is used for extracting three-dimensional reconstruction data features of a casting head area on the surface of the casting, performing multi-target path planning under robot kinematics constraint and generating a real-time robot machining track; the process execution module executes compensation according to the real-time machining track and the process parameters of the robot to carry out machining and cleaning operation; and the quality control module extracts casting surface texture features, constructs a roughness evaluation model and realizes process execution compensation by reversely optimizing process parameters. According to the system, intelligent cleaning work of the casting head of the large casting can be more suitably carried out.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and intelligent processing technology, and particularly relates to a casting sprue intelligent cleaning and surface quality control system based on feature extraction. BACKGROUND

[0002] Under the dual drive of intelligent manufacturing strategy and industrial automation policy, industries are accelerating the intelligent upgrading of production systems. Among them, the process revolution in the casting field is also obvious - through the deployment of industrial robots to realize the automation replacement of casting sprue cleaning, not only a breakthrough in the traditional production mode is made, but also a double benefit is achieved. On the one hand, the working environment is significantly optimized through mechanical automation operation, effectively protecting the occupational health and safety of employees; on the other hand, intelligent equipment is used to replace manual work, achieving the production goal of reducing costs and increasing efficiency for enterprises.

[0003] In the existing automatic cleaning system of castings, most of them use special clamps to fix the castings, and based on the preset model, offline programming or manual trajectory teaching is performed to process the castings to clean the sprue area. This kind of processing scheme is usually used for castings with consistent and fixed cleaning area morphology, while for castings with different cleaning area morphology and different positions, it is difficult to perform intelligent cleaning. At the same time, the existing method is also difficult to control the surface quality after intelligent cleaning. Therefore, there is an urgent need for a more effective large casting sprue intelligent cleaning and quality control system to effectively solve the above problems. SUMMARY

[0004] The technical problem to be solved by the present application is to solve the shortcomings of the prior art, and to provide a casting sprue intelligent cleaning and surface quality control system based on feature extraction, which realizes intelligent cleaning of the casting sprue area and control of the surface quality.

[0005] To solve the above technical problems, the technical solution adopted by the present application is: a casting sprue intelligent cleaning and surface quality control system based on feature extraction, comprising a three-dimensional reconstruction module, a feature planning module, a process execution module and a quality control module.

[0006] The three-dimensional reconstruction module collects laser-vision data of the casting surface sprue area, and realizes three-dimensional reconstruction of the casting surface sprue area through dynamic coordinate conversion and point cloud registration;

[0007] The feature planning module extracts the three-dimensional reconstruction data features of the casting surface sprue area, and performs multi-objective path planning under the kinematic constraint of the robot to generate real-time machining trajectories of the robot;

[0008] The process execution module clamps the machining tool by the robot, and performs machining and cleaning operations according to the real-time machining trajectories of the robot and the process parameters.

[0009] The quality control module extracts casting surface texture features, constructs a roughness evaluation model, and realizes process execution compensation through reverse optimization of process parameters to control the roughness index and ensure that the quality consistency meets the standard.

[0010] Preferably, the dynamic coordinate conversion and point cloud registration method of the three-dimensional reconstruction module is:

[0011] The texture data and point cloud data of the casting pouring riser area are collected; the texture data and point cloud data are converted to the same coordinate system based on filtering and coordinate conversion algorithm; the coarse registration of point cloud data of each view is completed based on NDT algorithm; and the fine registration of texture data and point cloud data is realized based on ICP algorithm.

[0012] Preferably, the three-dimensional reconstruction module further optimizes splicing of three-dimensional reconstruction data by entropy weighting, and fuses texture data and point cloud data, so as to realize multi-modal registration, specifically:

[0013] The point density distribution geometric entropy and color variance texture entropy of the pouring riser area are calculated respectively, the former reflects the structural complexity, and the latter reflects the color change intensity, to determine the regional entropy value H total , as follows:

[0014] H total =λH geometry +(1-λ)H texture

[0015] Wherein, H geometry is the geometric entropy, H texture is the texture entropy, and λ is the entropy weight, which is 0.6 in experience;

[0016] The area with H total higher than the threshold value 1.5 is defined as a high-entropy area, the ICP weight is increased to optimize the registration using visual texture details, and the NDT weight is reduced to avoid excessive smoothing of the probability model to details; the area with H total lower than the threshold value 1.5 is defined as a low-entropy area, the NDT weight is increased to utilize the geometric stability of laser, and the ICP weight is reduced to avoid unnecessary calculation overhead;

[0017] The registration objective function E is established, as follows:

[0018] E=ω NDT ·∑ε(T·p)+ω ICP ·∑d(T·p,q)

[0019] Wherein, ∑ε(T·p) is the NDT registration objective function, ∑d(T·p,q) -1 is the ICP registration objective function, and ω NDTNDT registration weight, ω ICP ICP registration weight, T is the transformation matrix in registration, p is an original point in the pouring riser area, and q is the corresponding point after registration of point p;

[0020] Optimize the transformation matrix T in registration by the LM algorithm to minimize the registration objective function E to achieve optimal registration;

[0021] The calculation of the objective function of NDT registration is as follows:

[0022] First, the target point cloud data is modeled by Gaussian distribution, and the multidimensional Gaussian distribution fitting is performed on the points in each voxel V i .

[0023]

[0024] Where n is the number of points in voxel V i , μ i is the mean vector of the points in voxel V i , Σ i is the covariance matrix of the points in voxel V i , and x j is the jth point in voxel V i .

[0025] A certain original point p is transformed by the transformation matrix T to become T·p, and point q is the point registered with it. It can be seen that the Gaussian distribution probability density of point q is fitted in a certain voxel. The NDT registration objective function formula is:

[0026]

[0027] The calculation of the objective function of ICP registration is as follows:

[0028] d(T·p,q)=α‖T·p-q‖+(1-α)‖I(T·p)-I(q)‖

[0029] Where I(·) is the color value, and α is the geometry-color weight.

[0030] Preferably, the feature planning module extracts three-dimensional reconstruction data features and performs multi-objective path planning under robot kinematics constraints to generate a specific method for generating real-time machining trajectories of the robot.

[0031] According to the real-time generated three-dimensional reconstruction data of the casting surface pouring riser area and the preset scheme of the machining position bounding box, the pouring riser area to be cleaned is selected, and after preprocessing, the planar, curved surface or other surface features are automatically extracted according to the process characteristics of different areas. Based on the feature points, the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0 is generated;

[0032] Based on the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0, the real-time machining trajectory P of the robot path in the current tool coordinate system T is calculated according to the current tool coordinate system T tool tool The trajectory will become the robot process execution path, so that the process tool carried by the robot can intelligently clean the actual topography of the current casting to be cleaned.

[0033] Preferably, the method for generating the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0 based on the feature point position is as follows:

[0034] Based on the robot motion constraint condition modeling and multi-objective optimization, the constraint condition setting includes the robot six-axis joint value limit interval wherein a = 1, 2, …, 6, are the minimum and maximum values of the robot six-axis joint values, respectively, and the end effector pose P0 is located in the set safe working space W, i.e. P0 ∈ W; the distance between the robot connecting rod and the environment is greater than the safety limit δ, i.e. Multi-objective optimization is realized based on the NSGA-II algorithm, and the optimization objectives include the path length min(∑‖P k+1 -P k ‖2) and safety max(d min (P k )), wherein P k is the kth path point in P0, P k+1 is the next path point adjacent to P k , and d min (P k ) is the distance between the robot connecting rod and the environment when a certain path point is executed.

[0035] Preferably, the specific calculation method for calculating the real-time machining trajectory P of the robot path in the current tool coordinate system T tool under the current tool coordinate system T tool is as follows:

[0036] The machining trajectory point P data is set as {P X , P Y , P Z , R X , R Y , R Z}, wherein P X , P Y , P Z represent the positions of the machining trajectory P on the three coordinate axes X, Y and Z; R X , R Y , R Z ​This represents the robot's posture, which is the angle of rotation around the X, Y, and Z axes of the current coordinate system.

[0037] In the rotation matrix Rot, N, O, and A are all coordinates derived from the robot flange coordinate system T0 to the current tool coordinate system T. tool The coordinate axes describe unit vectors, which are orthogonal and form a new right-handed coordinate system. A rotation matrix is ​​used to describe the relationship between the two coordinate systems at this point. The rotation matrix is ​​as follows:

[0038]

[0039] At this point, the rotation matrix Rot is used to modify the machining trajectory P. tool Multiplying by the left yields the actual machining trajectory P0, i.e.:

[0040] P0 = Rot × P tool

[0041] Therefore, the tool coordinate system T tool Real-time machining trajectory P of the robot path tool The calculation formula is:

[0042] P tool =Rot -1 ×P0.

[0043] Preferably, the specific method for the process execution module to perform compensation processing operations is as follows:

[0044] The robot grasps the processing tool and follows the real-time processing trajectory P. tool The processing parameters and process parameter compensation are used to clean the areas to be cleaned in each gating and riser area;

[0045] The machining process parameters are set according to the machining tools and process requirements, including the tool coordinate system T. tool Processing speed, tool feed direction, feed rate, cutting power, and coolant flow rate.

[0046] Preferably, the specific method by which the quality control module controls the consistency of the cleaned surface quality is as follows:

[0047] During the cleaning process, image data of the gating and riser area is acquired, and the surface texture features of the casting in the gating and riser area are extracted by a multi-scale convolutional filtering algorithm. A roughness evaluation model is constructed, and process execution compensation is achieved through reverse optimization of process parameters to control the consistency of surface quality.

[0048] Preferably, the specific method for extracting the surface texture features of the casting in the gating and riser area and constructing the roughness evaluation model is as follows:

[0049] The Gabor filter bank is designed, and the obtained image is respectively convolved with filters of different pixel scales λ and different directions θ to obtain response maps R corresponding to the filters λ,θ (x,y);

[0050] The feature vectors of the response maps R λ,θ (x,y) of the same pixel scale are respectively calculated and fused, including energy E λ,θ , variance and local contrast C λ,θ , and the statistical formula of the feature vector is:

[0051]

[0052] wherein N is the number of response maps of the same pixel scale, and Var is the variance of the response maps of the same pixel scale.

[0053] The real roughness Ra value of the pouring and feeding head region is measured by using a contact profiler, a sample set is constructed by grouping the feature vectors corresponding to the response maps, the feature dimension is reduced based on principal component analysis, and a roughness evaluation model is trained based on a support vector regression algorithm;

[0054] The roughness evaluation model function is:

[0055]

[0056] wherein z is the feature vector corresponding to the response map, f(z) is the predicted surface roughness, n SV is the number of support vectors, α m and is the Lagrange multiplier, K(z m ,z) is the kernel function, z m is the mth support vector, and b is the offset.

[0057] Preferably, the specific method for realizing process execution compensation by the reverse optimization process parameter is:

[0058] A response surface model is constructed to describe the relationship between the process parameters and the roughness, and the model formula is:

[0059] Ra=β0+β1v+β2f+β3Q+β 12 vf+β 11 v 2 +β 22 f 2 +β 33 Q 2

[0060] wherein β0, β1, β2, β3, β 12 , β 11 , β 22 , β33 are fitting coefficients, v is the machining speed, f is the grinding feed, and Q is the coolant flow rate;

[0061] The reverse optimization process parameter target is to maintain the roughness Ra target In the target range: Ra target ±Δ, when the roughness exceeds the target range, the Adam optimizer is used to update the process parameters along the negative gradient direction of the roughness, and the process execution module is quickly calculated and issued to ensure the consistency of the cleaned surface quality.

[0062] Preferably, the large casting sprue intelligent cleaning and quality control system based on feature extraction further comprises a host computer, a robot controller, a laser scanner and a binocular camera.

[0063] The host computer stores a host program, which performs the functions of the three-dimensional reconstruction module, the feature planning module, the process execution module and the quality control module when the program is loaded, and realizes intelligent cleaning of the large casting sprue area.

[0064] The robot controller is used to control the robot to execute the real-time machining trajectory P tool and feedback process execution data sent by the host computer.

[0065] The laser scanner and the binocular camera are respectively used to collect laser data and image data of the casting sprue area to be cleaned, so as to realize three-dimensional reconstruction of the casting sprue area.

[0066] The beneficial effects produced by the above technical solutions are that the casting sprue intelligent cleaning and surface quality control system based on feature extraction provided by the present application has the following advantages:

[0067] (1) It is applied to the intelligent cleaning process of the large casting sprue area, and can control the surface quality consistency. The topography of the area to be cleaned is different at different positions.

[0068] (2) During the machining process, the laser scanner and the binocular camera are used to collect laser-vision data for multi-source data fusion, and the three-dimensional reconstruction of the current casting to be cleaned is realized in real time through dynamic coordinate conversion and point cloud registration.

[0069] (3) During the machining process, the features are extracted according to the three-dimensional reconstruction model of the casting, and the multi-objective path planning under the kinematic constraint of the robot is carried out to generate the real-time machining trajectory of the robot.

[0070] (4) Different machining tools and matching process execution parameters are selected for the industrial robot, and the machining operation is carried out according to the real-time machining trajectory, so that the intelligent cleaning of the casting sprue area is realized.

[0071] (5) Based on multi-scale convolution filter and support vector regression algorithm, a surface roughness evaluation model is constructed, and process parameter reverse optimization is realized to achieve process execution compensation.

[0072] (6) Machine learning technology is adopted to extract features, reconstruct three-dimensional data of the casting, and plan the machining path in real time under the kinematic constraint of the robot. The corresponding robot machining process is matched to become a complete solution for the cleaning area of the large casting pouring and sprue. Compared with the traditional scheme, the cleaning process of the system has higher efficiency and better real-time performance, and can more appropriately perform intelligent cleaning work on the large casting pouring and sprue. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 A structure block diagram of the casting pouring and sprue intelligent cleaning and surface quality control system based on feature extraction provided by the embodiment of the present application is provided.

[0074] Figure 2 A structure schematic diagram of the three-dimensional reconstruction module provided by the embodiment of the present application is provided.

[0075] Figure 3 A structure schematic diagram of the feature planning module provided by the embodiment of the present application is provided.

[0076] Figure 4 An effect schematic diagram of the feature planning module provided by the embodiment of the present application is provided, wherein (a) is a casting physical map, (b) is a three-dimensional reconstruction effect diagram, (c) is a feature extraction effect diagram, and (d) is a path planning effect diagram.

[0077] Figure 5 A structure schematic diagram of the process execution module provided by the embodiment of the present application is provided.

[0078] Figure 6 A structure schematic diagram of the quality control module provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0079] The specific embodiments of the present application are further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0080] In the embodiment, the casting pouring and sprue intelligent cleaning and surface quality control system based on feature extraction includes a three-dimensional reconstruction module, a feature planning module, a process execution module, a quality control module, and an upper computer, a robot controller, a laser scanner, and a binocular camera, as shown in the figure. Figure 1

[0081] The three-dimensional reconstruction module collects laser-vision data of the casting surface pouring and sprue area, and realizes three-dimensional reconstruction of the casting surface pouring and sprue area through dynamic coordinate conversion and point cloud registration.​

[0082] The feature planning module extracts the three-dimensional reconstruction data features of the casting surface pouring and riser area, and performs multi-objective path planning under the kinematic constraints of the robot to generate real-time machining trajectories of the robot;

[0083] The process execution module clamps the machining tool by the robot, and performs machining and cleaning operations according to the real-time machining trajectories of the robot and the process parameters;

[0084] The quality control module extracts the surface texture features of the casting, constructs a roughness evaluation model, and realizes process execution compensation by reverse optimization of process parameters to control the roughness index and ensure the quality consistency to meet the standards;

[0085] The upper computer stores the upper program, and when the program is loaded, the functions of the three-dimensional reconstruction module, the feature planning module, the process execution module and the quality control module are executed to realize intelligent cleaning of the large casting pouring and riser area;

[0086] The robot controller is used to control the robot to execute the real-time machining trajectory P tool and feedback process execution data sent by the upper computer;

[0087] The laser scanner and the binocular camera are respectively used to collect laser data and image data of the casting pouring and riser area to be cleaned to realize three-dimensional reconstruction of the casting pouring and riser area.

[0088] In this embodiment, as shown in Figure 2 The method of dynamic coordinate conversion and point cloud registration of the three-dimensional reconstruction module is:

[0089] The texture data and point cloud data of the casting pouring riser area are collected to comprehensively capture the appearance features and three-dimensional space information of the area. The texture data can clearly present the color and texture details of the surface of the pouring riser, and provide rich visual basis; the point cloud data accurately records the three-dimensional coordinate information of the area, and truly reflects the geometric shape, laying a solid data foundation for subsequent analysis; the texture data and the point cloud data are converted to the same coordinate system based on filtering and coordinate conversion algorithm, effectively eliminating the deviation caused by factors such as device position and angle during data collection. The filtering process can remove noise points and redundant information in the point cloud data, improving the purity and accuracy of the data; coordinate conversion realizes the unification of the two different types of data in spatial position, ensuring that they can be registered and analyzed in the same reference frame, creating conditions for the fusion application of data; the rough registration of point cloud data at different angles is completed based on the NDT algorithm, narrowing the gap in spatial position between point cloud data at different angles. The NDT algorithm can roughly align the point cloud data at different angles without accurate initial correspondence. This process effectively reduces the positional deviation between point cloud data at different angles, significantly reducing the number of iterations and computational cost of subsequent fine registration; the fine registration of texture data and point cloud data is realized based on the ICP algorithm, significantly improving the matching accuracy of the two types of data. The ICP algorithm continuously iterates and optimizes to minimize the distance error between corresponding points, so that the local point cloud data and texture data obtained at different angles are highly consistent in spatial position, ensuring that the texture information can be accurately mapped to the three-dimensional point cloud model, providing high-precision data support for subsequent three-dimensional reconstruction, quality detection and other applications.

[0090] In this embodiment, the three-dimensional reconstruction module also performs entropy value weighted optimization splicing on the three-dimensional reconstruction data, fuses texture data and point cloud data, and realizes multi-modal registration, specifically:

[0091] The point density distribution geometric entropy and color variance texture entropy of the pouring riser area are calculated respectively, the former reflects the structural complexity, and the latter reflects the color change intensity, to determine the region entropy value H total The formula is as follows:

[0092] H total =λH geometry +(1-λ)H texture

[0093] Wherein, H geometry is the geometric entropy, H texture is the texture entropy, and λ is the entropy weight, which is 0.6 in experience;

[0094] The region with H total higher than the threshold value 1.5 is defined as a high-entropy region, the ICP weight is improved to utilize the visual texture details to optimize registration, and the NDT weight is reduced to avoid excessive smoothing of the probability model to details; the region with Htotal The area below the threshold 1.5 is defined as a low-entropy area, and the NDT weight is increased to utilize the laser geometric stability, and the ICP weight is reduced to avoid unnecessary computational overhead;

[0095] A registration objective function E is established, as shown in the following formula:

[0096] E = ω NDT ·∑ε(T·p) + ω ICP ·∑d(T·p,q)

[0097] Wherein, ∑ε(T·p) is an NDT registration objective function, ∑d(T·p,q) -1 is an ICP registration objective function, ω NDT is an NDT registration weight, ω ICP is an ICP registration weight, T is a transformation matrix during registration, p is an original point in the pouring and feeding area, and q is a corresponding point after registration of the point p;

[0098] The transformation matrix T during registration is optimized by the LM algorithm to minimize the registration objective function E to achieve optimal registration;

[0099] The calculation of the objective function of the NDT registration is as follows:

[0100] First, the target point cloud data is modeled by Gaussian distribution, and a multidimensional Gaussian distribution fitting is performed on the points in each voxel V i :

[0101]

[0102] Wherein, n is the number of points in the voxel V i , μ i is the mean vector of the points in the voxel V i , Σ i is the covariance matrix of the points in the voxel V i , and x j is the jth point in the voxel V i .

[0103] A certain original point p is changed to T·p after the transformation matrix T, and the point q is a registered point thereof, and the Gaussian distribution probability density fitted in a certain voxel of the point q is known, and the NDT registration objective function formula is:

[0104]

[0105] The calculation of the objective function of the ICP registration is as follows:

[0106] d(T·p,q) = α‖T·p-q‖ + (1-α)‖I(T·p)-I(q)‖

[0107] Where I(·) is the color value and α is the geometric-color weight.

[0108] In this embodiment, as Figure 3 , 4 As shown, the feature planning module extracts features from the 3D reconstruction data and performs multi-objective path planning under robot kinematic constraints to generate the robot's real-time processing trajectory. The specific method is as follows:

[0109] Based on the real-time generated 3D reconstruction data of the casting surface gating and riser area and the preset scheme of the processing position bounding box, the gating and riser area to be cleaned is selected. After preprocessing, plane, curved surface or other surface features are automatically extracted according to the process characteristics of different areas. Based on the feature points, the real-time processing trajectory P0 of the current casting in the robot flange coordinate system T0 is generated.

[0110] Based on the real-time machining trajectory P0 of the casting in the robot flange coordinate system T0, and according to the tool coordinate system T currently used by the robot... tool Calculate the real-time machining trajectory P of the robot path in the tool coordinate system. tool This trajectory will become the robot's process execution path, enabling the robot's tooling to intelligently clean the area of ​​the casting to be cleaned according to its actual shape.

[0111] The method for generating the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0 based on feature points is as follows:

[0112] Modeling and multi-objective optimization based on robot motion constraints; constraint settings include limit ranges for joint values ​​of the robot's six axes. Where a = 1, 2, ..., 6, These represent the minimum and maximum values ​​of the robot's six axis joints, respectively; the end effector pose P0 is located within the defined safe workspace W, i.e., P0∈W; and the distance between the robot links and the environment. Greater than the safety limit δ, that is This algorithm, NSGA-II, is used to perform multi-objective optimization, with the optimization objective being the path length min(∑‖P). k+1 -P k ‖2) and safety max(d min (P k ), where P k Let P be the k-th path point in P0. k+1 To be with P k The adjacent next path point, d min (P k ) represents the distance between the robot link and the environment when a certain path point is executed.

[0113] The coordinate system T of the currently used tooltool Calculate the real-time machining trajectory P of the robot path in the tool coordinate system. tool The specific calculation method is as follows:

[0114] Set the machining trajectory point P data as {P X P Y P Z R X R Y R Z}, where P X P Y P Z Represents the position of the machining trajectory P on the three coordinate axes X, Y, and Z; R X R Y R Z This represents the robot's posture, which is the angle of rotation around the X, Y, and Z axes of the current coordinate system.

[0115] In the rotation matrix Rot, N, O, and A are all coordinates derived from the robot flange coordinate system T0 to the current tool coordinate system T. tool The coordinate axes describe unit vectors, which are orthogonal and form a new right-handed coordinate system. A rotation matrix is ​​used to describe the relationship between the two coordinate systems at this point. The rotation matrix is ​​as follows:

[0116]

[0117] At this point, the rotation matrix Rot is used to modify the machining trajectory P. tool Multiplying by the left yields the actual machining trajectory P0, i.e.:

[0118] P0 = Rot × P tool

[0119] Therefore, the tool coordinate system T tool Real-time machining trajectory P of the robot path tool The calculation formula is:

[0120] P tool =Rot -1 ×P0.

[0121] In this embodiment, as Figure 5 As shown, the specific method by which the process execution module performs compensation to carry out processing operations is as follows:

[0122] The robot grasps the processing tool and follows the real-time processing trajectory P. tool The processing parameters and process parameter compensation are used to clean the areas to be cleaned in each gating and riser area;

[0123] The machining process parameters are set according to the machining tools and process requirements, including the tool coordinate system T.tool , machining speed, tool feed direction, feed amount, cutting power, and coolant flow rate.

[0124] In this embodiment, as shown in Figure 6 , the specific method for the quality control module to control the consistency of the cleaning surface quality is:

[0125] During the execution of the cleaning process, image data of the riser area is obtained, the texture features of the casting surface in the riser area are extracted through a multi-scale convolution filtering algorithm, a roughness evaluation model is constructed, and process execution compensation is realized through reverse optimization of process parameters to control the consistency of the surface quality.

[0126] During the execution of the cleaning process, image data of the riser area is obtained, which can monitor the surface state changes of the area in the machining process in real time, providing instant visual data for subsequent quality analysis. By extracting the texture features of the casting surface in the riser area through a multi-scale convolution filtering algorithm, the details of the texture can be captured from different scale levels, preserving both the macro texture distribution features and the microscopic texture changes, greatly improving the comprehensiveness and accuracy of feature extraction. The construction of the roughness evaluation model can quantify the extracted texture features into measurable roughness indicators, realizing objective evaluation of the casting surface quality and avoiding the subjectivity and errors of manual evaluation. Through reverse optimization of process parameters to realize process execution compensation, the cleaning process parameters such as robot running speed and cutting force can be adjusted in a timely manner according to the roughness evaluation results, effectively correcting the deviations in the machining process, and finally realizing precise control of the consistency of the surface quality of the riser area, ensuring that the surface quality of each casting can meet the preset standards.

[0127] The specific method for extracting the texture features of the casting surface in the riser area and constructing the roughness evaluation model is:

[0128] A Gabor filter bank is designed, and the obtained image is respectively convolved with filters of different pixel scales λ and different directions θ to obtain response maps R λ,θ (x,y) corresponding to each filter.

[0129] The filter generation formula is:

[0130]

[0131] Where x is the horizontal pixel coordinate of the image, y is the vertical pixel coordinate of the image, x' = xcosθ + ysinθ is the actual horizontal coordinate considering the direction of the filter, y' = -xsinθ + ycosθ is the actual vertical coordinate considering the direction of the filter, γ = 0.5 is the aspect ratio, and σ = λ / 2.

[0132] The response maps Rλ,θ (x, y) of the feature vector, including energy E λ,θ , variance and local contrast C λ,θ , the feature vector statistical formula is:

[0133]

[0134]

[0135] Wherein, N is the number of response map under the same pixel scale, Var is the variance of the response map under the same pixel scale.

[0136] The real roughness Ra value of the pouring head area is measured by using a contact profiler, a sample set is constructed by grouping the feature vectors corresponding to the response map, the feature dimension is reduced based on principal component analysis, and a roughness evaluation model is trained based on support vector regression algorithm;

[0137] Roughness evaluation model function:

[0138]

[0139] Wherein, z is the feature vector corresponding to the response map, f(z) is the predicted surface roughness, n SV The number of support vectors is α m and The Lagrange multiplier is K(z m , z) is the kernel function, z m is the mth support vector, and b is the offset.

[0140] The specific method of the reverse optimization process parameter to realize process execution compensation is:

[0141] A response surface model is constructed to describe the relationship between process parameters and roughness, and the model formula is:

[0142] Ra=β0+β1v+β2f+β3Q+β 12 vf+β 11 v 2 +β 22 f 2 +β 33 Q 2

[0143] Wherein, β0, β1, β2, β3, β 12 , β 11 , β 22 , β 33 are fitting coefficients, v is the processing speed, f is the grinding feed, and Q is the cooling liquid flow;

[0144] The goal of the reverse optimization process parameter is to maintain the roughness Ratarget In the target range: Ra target When the roughness exceeds the target range, the process parameters are updated along the negative gradient direction of the roughness using the Adam optimizer, the process parameters are calculated and sent to the process execution module, and the cleaning surface quality consistency is ensured through process execution compensation.

[0145] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the present application.

Claims

1. A feature extraction based intelligent gating and riser cleaning and surface quality control system for castings, characterized by: The system comprises a three-dimensional reconstruction module, a feature planning module, a process execution module and a quality control module. The three-dimensional reconstruction module collects laser-vision data of the casting surface and the pouring riser area, and realizes three-dimensional reconstruction of the casting surface and the pouring riser area through dynamic coordinate conversion and point cloud registration. The feature planning module extracts three-dimensional reconstruction data features of the casting surface and the pouring riser area, and performs multi-objective path planning under the kinematic constraint of a robot to generate real-time machining trajectories of the robot. The process execution module clamps a machining tool by the robot, and performs machining and cleaning operations according to the real-time machining trajectories of the robot and process parameters. The quality control module extracts casting surface texture features, constructs a roughness evaluation model, and realizes process execution compensation through reverse optimization of process parameters to control the roughness index and ensure the quality consistency to meet the standards.

2. The feature extraction based intelligent cleaning of casting sprues and runners and surface quality control system as claimed in claim 1 wherein: The method for dynamic coordinate conversion and point cloud registration of the three-dimensional reconstruction module is as follows: Texture data and point cloud data of the casting pouring riser area are collected, the texture data and the point cloud data are converted to the same coordinate system based on filtering and coordinate conversion algorithm, rough registration of point cloud data of each view is completed based on NDT algorithm, and precise registration of the texture data and the point cloud data is realized based on ICP algorithm.

3. The feature extraction based intelligent sprue and skull cleaning and surface quality control system of claim 2, wherein: The three-dimensional reconstruction module further optimizes and splices the three-dimensional reconstruction data by entropy weighting, and fuses the texture data and the point cloud data, so as to realize multi-modal registration, specifically as follows: The point density distribution geometric entropy and color variance texture entropy are calculated for the pouring and gating region respectively, the former reflects the structural complexity, and the latter reflects the color change intensity, to determine the region entropy value H total As follows: H total = λH geometry + (1 - λ)H texture Where H geometry is the geometric entropy, H texture is the texture entropy, and λ is the entropy weight, which is an empirical value of 0.

6. H total Regions higher than threshold 1.5 are defined as high-entropy regions, and the ICP weight is increased to optimize registration using visual texture details, and the NDT weight is reduced to avoid excessive smoothing of the probability model by details; H total Regions lower than threshold 1.5 are defined as low-entropy regions, and the NDT weight is increased to utilize laser geometric stability, and the ICP weight is reduced to avoid unnecessary computational overhead; A registration objective function E is established, as shown in the following formula: E = ω NDT • ∑ε(T·p) + ω ICP • ∑d(T·p, q) where ∑ε(T·p) is the NDT registration objective function, ∑d(T·p, q) -1 is the ICP registration objective function, ω NDT is the NDT registration weight, ω ICP is the ICP registration weight, T is the transformation matrix in registration, p is one original point in the pouring basin area, and q is the corresponding point after registration of the point p. The transformation matrix T during registration is optimized by the LM algorithm to minimize the registration objective function E to achieve optimal registration. The calculation of the objective function of NDT registration is as follows: Firstly, the target point cloud data is modeled by Gaussian distribution, and the multidimensional Gaussian distribution fitting is performed on the points in each voxel V i : where n is the number of voxels V i the number of midpoints, μ i the mean vector of the points in a voxel V i the covariance matrix of the points in a voxel V i the jth point in a voxel V i the jth point in a voxel V j the jth point in a voxel V i the jth point in a voxel V A certain origin p is changed to T·p after the transformation matrix T, and q is the registered point, so the Gaussian distribution probability density of the point q in a certain voxel is fitted, and the NDT registration objective function formula is as follows: The calculation of the objective function of ICP registration is as follows: d(T·p,q)=α‖T·p-q‖+(1-α)‖I(T·p)-I(q)‖ Wherein, I(·) is the color value, and a is the geometry-color weight.

4. The feature extraction based intelligent gating and riser cleaning and surface quality control system of claim 1, wherein: The specific method for the feature planning module to extract three-dimensional reconstruction data features and perform multi-objective path planning under the kinematic constraint of a robot to generate real-time machining trajectories of the robot is as follows: According to the three-dimensional reconstruction data of the casting surface and the pouring riser area generated in real time and the preset scheme of the machining position bounding box, the pouring riser area to be cleaned is selected, the planar, curved surface or other surface features are automatically extracted according to the process characteristics of different regions after preprocessing, and the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0 is generated based on the feature point position. Based on the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0, the real-time machining trajectory P tool of the robot path in the tool coordinate system T tool currently used by the robot is calculated, and this trajectory will become the robot process execution path, so that the process tool carried by the robot can intelligently clean the actual topography of the current casting to be cleaned.

5. The feature extraction based intelligent sprue and runner cleaning and surface quality control system of claim 4, wherein: The method for generating the real-time machining trajectory P0 of the current casting in the robot flange coordinate system T0 based on the feature point position is as follows: Modeling and multi-objective optimization based on robot motion constraints; constraint settings include robot six-axis joint value limit interval Wherein, a = 1, 2, …, 6, Respectively, the minimum and maximum values of the joint values of the six axes of the robot, the pose P0 of the end effector is located in the set safety workspace W, i.e. P0 ∈ W; the distance between the robot link and the environment Is greater than the safety limit δ, i.e. Multi-objective optimization is realized based on NSGA-II algorithm, and the optimization objectives include path length min(∑‖P k+1 -P k ‖2) and safety max(d min (P k )), wherein P k is the kth path point in P0, P k+1 is the next path point adjacent to P k , and d min (P k ) is the distance between the robot link and the environment when a certain path point is executed; The tool coordinate system T tool is calculated according to the currently used tool coordinate system T tool The specific calculation method of the real-time machining trajectory P tool of the robot path in the tool coordinate system is as follows: Set the machining trajectory point P data as {P X , P Y , P Z , R X , R Y , R Z}, wherein P X , P Y , P Z represent the position of the machining trajectory P on the three coordinate axes X, Y, Z; R X , R Y , R Z represent the robot pose, that is, the angle of rotation around the X, Y, Z three coordinate axes in the current coordinate system respectively; N, O, A in the rotation matrix Rot are all in the coordinate system T of the robot flange, and the coordinate system T is used for the current tool coordinate system tool The unit vectors described by the coordinate axes are orthogonal and form a new right-handed coordinate system. The rotation matrix describes the change relationship between the two coordinate systems at this time, and the rotation matrix is as follows: At this point, the rotation matrix Rot is used to modify the machining trajectory P. tool Multiplying by the left yields the actual machining trajectory P0, i.e.: P0 = Rot x P tool Therefore, the tool coordinate system T tool The real-time machining trajectory P tool The calculation formula is: P tool = Rot -1 x P0.

6. The feature extraction based casting sprue and runner intelligent cleaning and surface quality control system as claimed in claim 4, wherein: The specific method for the process execution module to perform machining operation by compensation is as follows: The robot grabs the machining tool according to the real-time machining track P tool The machining process parameters and the process parameter execution compensation are used to clean each riser and runner area to be cleaned. The machining process parameters are set according to the machining tool and process requirements, including tool coordinate system T tool , machining speed, tool feed direction, feed amount, cutting power and cooling liquid flow.

7. The feature extraction based casting sprue and runner intelligent cleaning and surface quality control system as claimed in claim 6, wherein: The specific method for the quality control module to control the consistency of the cleaning surface quality is as follows: During the cleaning process execution, image data of the pouring riser area is acquired, casting surface texture features of the pouring riser area are extracted through a multi-scale convolution filtering algorithm, a roughness evaluation model is constructed, process execution compensation is realized through reverse optimization of process parameters, and the surface quality consistency is controlled.

8. The feature extraction based casting sprue and runner intelligent cleaning and surface quality control system as claimed in claim 7, wherein: The specific method for extracting the texture features of the casting surface in the pouring and gating region and constructing the roughness evaluation model is as follows: A Gabor filter bank is designed, and the acquired image is convolved with filters of different pixel scales λ and different directions θ, respectively, to obtain a response map R corresponding to each filter λ,θ (x, y); The response map R is calculated and fused at the same pixel scale respectively λ,θ The feature vector of (x, y), including energy E λ,θ , variance And local contrast C λ,θ The statistical formula of the feature vector is: Wherein, N is the number of response maps under the same pixel scale, and Var is the variance of the response map under the same pixel scale. The real roughness Ra value of the pouring and gating region is measured by using a contact profilometer, a sample set is constructed by grouping the corresponding feature vectors of the response maps, the feature dimension is reduced based on principal component analysis, and a roughness evaluation model is trained based on a support vector regression algorithm; The roughness evaluation model function is as follows: where z is a feature vector corresponding to the response map, f(z) is the predicted surface roughness, n SV is the number of support vectors, a m and is the Lagrange multiplier, K(z m ,z) is the kernel function, z m is the mth support vector, and b is the bias.

9. The feature extraction based casting sprue and runner intelligent cleaning and surface quality control system as claimed in claim 8, wherein: The specific method for realizing process execution compensation by reverse optimization of the process parameters is as follows: A response surface model is constructed to describe the relationship between the process parameters and the roughness, and the model formula is as follows: Ra = β0+ β1v + β2f + β3Q + β 12 vf+β 11 v 2 +β 22 f 2 +β 33 Q 2 wherein β0, β1, β2, β3, β 12 , β 11 , β 22 , β 33 are fitting coefficients, v is the machining speed, f is the grinding feed, and Q is the coolant flow rate; The reverse optimization process parameter target is to maintain the roughness Ra target In the target range: Ra target When the roughness exceeds the target range, use the Adam optimizer to update the process parameters in the negative gradient direction of the roughness, quickly calculate and issue to the process execution module, and ensure the cleaning surface quality consistency through process execution compensation.

10. The feature extraction based intelligent sprue and skull cleaning and surface quality control system of any of claims 2-9, wherein: The large casting pouring and gating intelligent cleaning and quality control system based on feature extraction further comprises an upper computer, a robot controller, a laser scanner and a binocular camera. The upper computer stores an upper program, and when the program is loaded, the functions of the three-dimensional reconstruction module, the feature planning module, the process execution module and the quality control module are executed to realize intelligent cleaning of the pouring and gating region of the large casting. The robot controller is used to control the robot to execute the real-time machining trajectory P sent by the host computer tool and feedback process execution data; The laser scanner and the binocular camera are respectively used to collect laser data and image data of the pouring and gating region of the casting to be cleaned to realize three-dimensional reconstruction of the pouring and gating region of the casting.

Images