Railroad plate drilling parameter intelligent optimization method and system based on deep learning

Abstract

The application provides a track plate drilling parameter intelligent optimization method and system based on deep learning, relates to the technical field of deep learning, and comprises the following steps: identifying a feature area in three-dimensional scanning data through a deep learning semantic segmentation network, establishing a local coordinate system, calculating a hole position deviation, and constructing a pose correction field, so that intelligent correction of drilling process parameters is realized, and an optimal drilling track is generated by using a state space search algorithm. The application can adapt to actual track plate geometric errors, improve drilling precision and efficiency, reduce drilling failure rate, and improve track plate manufacturing quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, and in particular to a track plate drilling parameter intelligent optimization method and system based on deep learning. BACKGROUND

[0002] As a key component in high-speed railway and urban rail transit systems, the processing precision of track plates directly relates to the stability of track structures and driving safety. Track plate drilling is an important process in its manufacturing process, which needs to ensure hole position precision, processing efficiency, and quality consistency. Traditional track plate drilling processes mainly rely on manual measurement and fixed molds for positioning, followed by operators controlling drilling equipment to complete processing. With the rapid development of rail transit construction, automated drilling equipment has gradually been applied to track plate production lines to achieve batch production through preset processing programs.

[0003] In recent years, with the development of computer vision and robotics, automatic drilling systems based on visual inspection have begun to be applied in track plate manufacturing. Such systems obtain track plate surface information through cameras or laser scanning devices, determine drilling positions through image processing algorithms, and complete processing tasks through numerical control systems controlling drilling actuators. This method has certain adaptability and automation compared to traditional processes.

[0004] Existing technologies cannot accurately identify and adapt to the actual state differences of track plates with different shapes and complex surface states, resulting in insufficient drilling position precision. Especially when there are manufacturing deviations or deformations in track plates, preset drilling parameters often cannot meet actual processing needs. Secondly, traditional drilling path planning lacks intelligent optimization capabilities and cannot dynamically adjust drilling parameters according to the pose state of actual track plates, causing low processing efficiency and low equipment utilization. SUMMARY

[0005] The present application provides a track plate drilling parameter intelligent optimization method and system based on deep learning, which can solve the problems in the prior art.

[0006] In a first aspect, the present application provides a track plate drilling parameter intelligent optimization method based on deep learning, comprising:

[0007] Obtaining three-dimensional scanning data of a track plate and preset drilling process data, using a deep learning semantic segmentation network to identify feature regions of the three-dimensional scanning data, and determining a reference surface region and a hole position region of the track plate according to the region probability distribution of each feature;

[0008] A local coordinate system of the rail plate is established according to the reference surface region, and geometric center calculation and normal vector estimation are performed on the hole site region in the local coordinate system to determine the spatial position data of each hole site, and by comparing the spatial position data with standard hole site data in the drilling process data, a position deviation vector is calculated and a pose correction field is constructed;

[0009] The drilling path parameters in the drilling process data are point-by-point corrected based on the pose correction field to obtain a drilling task point sequence adapted to the actual rail plate state;

[0010] The drilling task point sequence is taken as a state node, and the joint space motion range, tool posture accessibility and collision safety constraint of the drilling equipment are taken as state transition conditions, and an optimal drilling trajectory sequence satisfying all constraint conditions is generated through a state space search algorithm;

[0011] The drilling equipment is controlled to perform rail plate drilling operation according to the optimal drilling trajectory sequence.

[0012] A deep learning semantic segmentation network is used to identify feature regions of the three-dimensional scanning data, and the reference surface region and the hole site region of the rail plate are determined according to the region probability distribution corresponding to each feature.

[0013] The three-dimensional scanning data is voxelized to obtain voxelized representation data, the voxelized representation data is input into the deep learning semantic segmentation network, multi-level feature extraction is performed on the voxelized representation data through convolution operation, and the multi-level features are up-sampled and fused, so that the features of different levels are aligned in the spatial dimension and are spliced in the channel to obtain fused features.

[0014] Each spatial position of the fused features is classified and operated to calculate the region probability distribution of the position belonging to the reference surface category, the hole site category and the background category, and the category with the maximum region probability distribution is selected as the semantic label of the position.

[0015] Connected domain analysis is performed on voxels with the same category label and spatially connected according to the semantic label, the spatial range of each connected domain is identified, the connected domain corresponding to the reference surface category label is determined as the reference surface region, and each connected domain corresponding to the hole site category label is determined as the hole site region.

[0016] A local coordinate system of the rail plate is established according to the reference surface region, and geometric center calculation and normal vector estimation are performed on the hole site region in the local coordinate system to determine the spatial position data of each hole site, and by comparing the spatial position data with standard hole site data in the drilling process data, a position deviation vector is calculated and a pose correction field is constructed.

[0017] Point cloud data of the reference surface region is extracted, and plane fitting operation is performed on the point cloud data to obtain a fitted plane equation of the reference surface.

[0018] The normal vector of the reference plane is extracted based on the fitted plane equation and used as the first coordinate axis direction of the local coordinate system. The geometric center point is selected as the origin of the local coordinate system from the point cloud data of the reference plane region.

[0019] In a plane perpendicular to the first coordinate axis, the length direction of the track slab is defined as the second coordinate axis direction. The third coordinate axis direction is obtained by performing a vector cross product operation between the first and second coordinate axis directions, thus establishing the local coordinate system.

[0020] The point cloud data of each sub-region of the hole location region is transformed from the global coordinate system to the local coordinate system; the centroid of each transformed point cloud data of each sub-region of the hole location is calculated to obtain the geometric center coordinates of each sub-region of the hole location in the local coordinate system.

[0021] Surface fitting is performed on the point cloud data of each hole location sub-region, and the surface gradient direction of each hole location sub-region is calculated. The surface gradient direction is then used as the normal vector of each hole location sub-region.

[0022] The spatial location data is formed by combining the geometric center coordinates and the normal vector of each hole.

[0023] Based on the pose correction field, the drilling path parameters in the drilling process data are corrected point by point to obtain a drilling task point sequence adapted to the actual track slab state, including:

[0024] The design coordinates and design tool posture of each theoretical drilling point are extracted from the drilling process data;

[0025] A spatial interpolation function is established in the pose correction field, which characterizes the mapping relationship between the pose deviation at any spatial location and the coordinates of that location.

[0026] Input the design coordinates of each theoretical borehole point into the spatial interpolation function to calculate the position correction vector and attitude correction matrix at each theoretical borehole point;

[0027] Apply the corresponding position correction vector to the design coordinates of each theoretical drilling point, and obtain the target coordinates of each corrected drilling point on the actual track slab surface through coordinate transformation calculation;

[0028] Apply the corresponding attitude correction matrix to the designed tool attitude at each theoretical drilling point, and obtain the target tool attitude at each corrected drilling point through rotation transformation operation;

[0029] Based on the target coordinates and target tool posture of each corrected drilling point, calculate the homogeneous transformation matrix of the tool posture at each corrected drilling point;

[0030] According to the machining sequence constraint in the drilling process data, the tool pose homogeneous transformation matrix of each modified drilling point is sequentially organized to form the drilling task point sequence.

[0031] The drilling task point sequence is taken as a state node, and the joint space motion range, tool posture accessibility and collision safety constraint of the drilling equipment are taken as state transition conditions. An optimal drilling trajectory sequence satisfying all constraint conditions is generated by a state space search algorithm, including:

[0032] The joint angle solution set corresponding to each drilling task point in the drilling task point sequence is obtained. The joint angles in the joint angle solution set are sequentially judged whether they are located between the upper limit value and the lower limit value of the corresponding joint angle. The reachable drilling task points whose joint angles meet the range constraint are screened out.

[0033] The included angle between the actual tool axis direction of the end effector at each reachable drilling task point and the target tool axis direction required by the drilling task point is calculated. The executable drilling task points whose included angles are less than a preset posture deviation threshold are screened out.

[0034] The geometric boundary of the motion component of the drilling equipment and the spatial envelope range of the obstacles in the working environment are obtained. For adjacent point pairs between the executable drilling task points, the joint angles from the starting point to the target point are discretely interpolated to generate a plurality of intermediate joint states.

[0035] The spatial position of the geometric boundary of the motion component at each intermediate joint state is calculated. It is judged whether the spatial position overlaps with the spatial envelope range of the obstacles. It is determined whether the motion path between the adjacent point pairs exists collision.

[0036] The motion path without collision is taken as a feasible motion path. The weighted sum of the joint angle change amount corresponding to each feasible motion path is taken as the motion cost of the feasible motion path.

[0037] Starting from the starting drilling task point, the path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm. The path combination with the minimum sum of motion costs of all feasible motion paths is selected as the optimal drilling trajectory sequence.

[0038] Starting from the starting drilling task point, the path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm. The path combination with the minimum sum of motion costs of all feasible motion paths is selected as the optimal drilling trajectory sequence, including:

[0039] creating a priority queue, elements in the priority queue including a current arrived drilling task point identifier, a drilling task point access sequence that has passed and a corresponding total motion cost;

[0040] popping out a head element with a minimum total motion cost from the priority queue as a current expansion element;

[0041] judging whether the access sequence of the current expansion element has included all executable drilling task points, if yes, marking the current expansion element as a candidate solution and skipping subsequent expansion steps, if not, performing subsequent expansion steps;

[0042] searching for adjacent drilling task points directly connected with the drilling task point identifier of the current expansion element and not appearing in the access sequence of the current expansion element in the collision-free connection relationship;

[0043] for each adjacent drilling task point, obtaining a motion cost of a feasible motion path from the drilling task point of the current expansion element to the adjacent drilling task point, adding the motion cost to the total motion cost of the current expansion element to obtain an expanded total motion cost;

[0044] repeating the step of popping out the head element from the priority queue until the current expansion element popped out is marked as a candidate solution;

[0045] selecting an optimal candidate solution with a minimum expanded total motion cost from all candidate solutions, and combining a path corresponding to the optimal candidate solution as the optimal drilling trajectory sequence.

[0046] A second aspect of the embodiment of the present application provides a deep learning-based track plate drilling parameter intelligent optimization system, comprising:

[0047] a first unit configured to acquire three-dimensional scanning data of a track plate and preset drilling process data, perform feature region identification on the three-dimensional scanning data by using a deep learning semantic segmentation network, and determine a reference surface region and a hole site region of the track plate according to a region probability distribution corresponding to each feature;

[0048] a second unit configured to establish a local coordinate system of the track plate according to the reference surface region, perform geometric center calculation and normal vector estimation on the hole site region in the local coordinate system, determine spatial position data of each hole site, calculate a position deviation vector by comparing the spatial position data with standard hole site data in the drilling process data, and construct a pose correction field;

[0049] a third unit configured to perform point-by-point correction on drilling path parameters in the drilling process data based on the pose correction field, to obtain a drilling task point sequence adapted to an actual track plate state;

[0050] The fourth unit is configured to generate an optimal drilling trajectory sequence meeting all constraint conditions by a state space search algorithm, taking the drilling task point sequence as a state node and taking the joint space motion range of the drilling device, tool posture reachability and collision safety constraint as state transition conditions.

[0051] The fifth unit is configured to control the drilling device to perform the track plate drilling operation according to the optimal drilling trajectory sequence.

[0052] The third aspect of the embodiment of the application,

[0053] An electronic device is provided, comprising:

[0054] A processor;

[0055] A memory for storing processor-executable instructions;

[0056] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0057] The fourth aspect of the embodiment of the application,

[0058] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0059] The beneficial effects of the present application are as follows:

[0060] The feature region of the three-dimensional scanning data is identified by the deep learning semantic segmentation network, which can automatically and accurately identify the reference surface region and the hole position region of the track plate, overcome the subjectivity and inefficiency of the traditional manual identification method, and improve the accuracy of the drilling position identification.

[0061] The local coordinate system of the track plate is established based on the identified reference surface region, and the geometric center calculation and normal vector estimation of the hole position region are performed on this basis, so that the spatial position data of each hole position is accurately obtained, and the deformation, displacement and other state deviations of the track plate in the actual environment are effectively dealt with.

[0062] By constructing a pose correction field and correcting the path parameters in the drilling process data point by point, the drilling parameters can adapt to the actual state of the track plate, ensuring the drilling precision and reducing the drilling error caused by the deviation of the workpiece.

[0063] The drilling task point sequence is taken as a state node, and the joint space motion range of the drilling device, the tool posture reachability and the collision safety constraint are comprehensively considered, and the optimal drilling trajectory sequence is generated by a state space search algorithm, so that the problem of drilling trajectory planning under complex constraint conditions is solved. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 A flowchart of an embodiment of the track plate drilling parameter intelligent optimization method based on deep learning of the present application is shown in FIG. 1.

[0065] Figure 2 A flowchart of generating an optimal drilling trajectory sequence is shown in FIG. 2. DETAILED DESCRIPTION

[0066] To make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0067] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0068] Reference Figure 1 and Figure 2 The track plate drilling parameter intelligent optimization method based on deep learning of the embodiments of the present application comprises:

[0069] Obtaining three-dimensional scanning data of a track plate and preset drilling process data, using a deep learning semantic segmentation network to identify feature regions of the three-dimensional scanning data, determining a reference surface region and a hole site region of the track plate according to the region probability distribution corresponding to each feature;

[0070] Establishing a local coordinate system of the track plate according to the reference surface region, and calculating the geometric center and estimating the normal vector of the hole site region in the local coordinate system to determine the spatial position data of each hole site, comparing the spatial position data with the standard hole site data in the drilling process data to calculate the position deviation vector and construct a pose correction field;

[0071] Based on the pose correction field, the drilling path parameters in the drilling process data are corrected point by point to obtain a drilling task point sequence that adapts to the actual state of the track plate;

[0072] Taking the drilling task point sequence as a state node, taking the joint space motion range, tool posture accessibility and collision safety constraints of the drilling equipment as state transition conditions, and generating an optimal drilling trajectory sequence that satisfies all the constraint conditions through a state space search algorithm;

[0073] Controlling the drilling equipment to perform track plate drilling operation according to the optimal drilling trajectory sequence.

[0074] In an alternative embodiment, a deep learning semantic segmentation network is used to identify feature regions of the three-dimensional scanning data, and the reference surface region and the hole site region of the track plate are determined according to the region probability distribution corresponding to each feature, comprising:

[0075] The three-dimensional scanning data is voxelized to obtain voxelized representation data, the voxelized representation data is input into a deep learning semantic segmentation network, multi-level feature extraction is performed on the voxelized representation data through convolution operation, upsampling and feature fusion are performed on the multi-level features, the features of different levels are aligned in the spatial dimension and are spliced in the channel to obtain fused features;

[0076] Each spatial position of the fused features is subjected to classification operation, the region probability distribution of the position belonging to the reference surface class, the hole site class and the background class is calculated, and the class with the maximum region probability distribution is selected as the semantic label of the position;

[0077] Connected component analysis is performed on the voxels with the same class label and spatially connected according to the semantic label, the spatial range of each connected component is identified, the connected component corresponding to the reference surface class label is determined as the reference surface region, and each connected component corresponding to the hole site class label is determined as the hole site region.

[0078] In this embodiment, a deep learning semantic segmentation network is used to identify feature regions of the three-dimensional scanning data of the track plate, mainly including steps of voxelizing the three-dimensional scanning data, deep semantic segmentation network feature extraction, feature fusion, classification operation and connected component analysis, so as to accurately identify the reference surface region and the hole site region of the track plate.

[0079] Three-dimensional scanning data of the track plate is obtained, which is usually collected by a three-dimensional laser scanner or a structured light scanner, and contains geometric shape information of the track plate, which is represented as a large amount of three-dimensional point cloud data. Since the original three-dimensional point cloud data is irregular in form, it is difficult to directly apply a deep learning network for processing, and voxelization processing is required.

[0080] The voxelization processing process is to convert irregular three-dimensional point cloud data into regular three-dimensional grid structure. In specific implementation, the size of the voxel is first determined, for example, set to 2mm x 2mm x 2mm, and then a uniform grid is established in three-dimensional space, and each point cloud data is assigned to the corresponding voxel according to its spatial position. If there is a point in the voxel, the voxel is marked as occupied; if there is no point, it is marked as idle. For voxels containing multiple points, the average position or other statistical features of these points can be calculated as the attribute value of the voxel. After voxelization processing, the voxelized representation data obtained is a three-dimensional tensor, where each element represents the voxel state and attribute of the corresponding spatial position.

[0081] The voxelized characterization data is input into a pre-trained deep learning semantic segmentation network, which adopts a three-dimensional convolutional neural network structure including an encoder and a decoder. The encoder part extracts features from the input data layer by layer through multi-layer three-dimensional convolution operations, and each layer of convolution is followed by batch normalization and a ReLU activation function to enhance the non-linear expression ability of the network. The convolution kernel size is generally set to 3x3x3, the step is 1, and the padding is 1 to maintain the spatial dimension of the feature map. To reduce computational complexity and extract higher-level features, a max-pooling operation is applied after every few convolution layers to halve the spatial resolution of the feature map.

[0082] Specifically, the encoder contains multiple stages, each containing 2-3 convolution layers, and the number of channels of the output feature map of each stage is 64, 128, 256, and 512, respectively, forming a hierarchical feature representation from low to high. Low-level features mainly contain local information such as edges and textures, while high-level features contain more abstract semantic information. These features at different levels are important for accurately identifying the datum plane and the hole location area, so feature fusion is needed in the decoder.

[0083] The decoder part restores the low-resolution feature map to the original resolution through upsampling and fuses it with the feature map of the corresponding layer of the encoder. The upsampling is implemented using deconvolution (transposed convolution) with a kernel size of 2x2x2 and a step of 2, which expands the spatial dimension of the feature map by a factor of two. Feature fusion uses a skip connection method to concatenate the feature maps of the corresponding layers of the encoder and the upsampled feature maps in the channel dimension, combining the spatial detail information of the low layer and the semantic information of the high layer.

[0084] During feature fusion, the feature maps of different levels are first unified to the same spatial resolution through upsampling, and then concatenated in the channel dimension. For example, the fourth stage feature map (8x8x8x512) is upsampled to 16x16x16x512, and concatenated with the third stage feature map (16x16x16x256) to obtain a fusion feature of 16x16x16x768. In this way, layer-by-layer upsampling and feature fusion with the corresponding encoder layer are performed, and finally a fusion feature with the same spatial resolution as the input data is obtained, with a total of 768 channels.

[0085] The classification operation is performed on each spatial position of the fused feature, and a 1x1x1 convolution kernel is used to map the fused feature to a class probability. Assuming that the rail plate recognition task includes three categories: a reference surface category, a hole position category, and a background category, the number of convolution kernel output channels is 3. The Softmax function is applied to normalize the three channel values of each position to obtain the probability distribution of each category for the position. According to the probability distribution, the category with the maximum probability is selected as the semantic label of the position to form a complete semantic segmentation result.

[0086] Connected component analysis is performed on voxels with the same category label and spatial connectivity to identify the spatial range of each connected component. A three-dimensional connected component labeling algorithm, such as a three-dimensional version of breadth-first search or region growing algorithm, is used to sequentially access each voxel. For voxels that have not been accessed and have a target category label, the algorithm recursively or iteratively accesses neighboring voxels from the voxel. If a neighboring voxel has the same category label, it is added to the current connected component. This process continues until no further expansion is possible. In this way, all spatially connected voxels belonging to the same category are grouped into a connected component. For connected components corresponding to the reference surface category label, the reference surface region of the rail plate is determined. For each connected component corresponding to the hole position category label, the hole position region of the rail plate is determined.

[0087] Through the above steps, the feature region recognition of the rail plate three-dimensional scanning data is completed, and the reference surface region and the hole position region are accurately located, providing an important spatial positioning basis for subsequent rail plate processing.

[0088] In an alternative embodiment, a local coordinate system of the rail plate is established based on the reference surface region, and geometric center calculation and normal vector estimation are performed on the hole position region in the local coordinate system to determine the spatial position data of each hole position, including:

[0089] Point cloud data of the reference surface region is extracted, and plane fitting operation is performed on the point cloud data to obtain a fitted plane equation of the reference surface.

[0090] The normal vector of the reference surface is extracted as the direction of the first coordinate axis of the local coordinate system based on the fitted plane equation, and a geometric center point is selected as the origin of the local coordinate system in the point cloud data of the reference surface region.

[0091] In a plane perpendicular to the direction of the first coordinate axis, the length direction of the rail plate is determined as the direction of the second coordinate axis, and the third coordinate axis direction is obtained by vector cross multiplication operation of the first coordinate axis direction and the second coordinate axis direction to establish the local coordinate system.

[0092] The point cloud data of each sub-region of the hole location region is transformed from the global coordinate system to the local coordinate system; the centroid of each transformed point cloud data of each sub-region of the hole location is calculated to obtain the geometric center coordinates of each sub-region of the hole location in the local coordinate system.

[0093] Surface fitting is performed on the point cloud data of each hole location sub-region, and the surface gradient direction of each hole location sub-region is calculated. The surface gradient direction is then used as the normal vector of each hole location sub-region.

[0094] The spatial location data is formed by combining the geometric center coordinates and the normal vector of each hole.

[0095] When determining the spatial location data of holes on the track plate, it is first necessary to establish a local coordinate system to accurately locate each hole. The reference surface area on the track plate serves as the basis for establishing the local coordinate system, while the hole location area is the target area whose spatial position needs to be determined.

[0096] Extracting point cloud data from the reference surface region is the first step in establishing a local coordinate system. A 3D laser scanning device is used to perform a full-range scan of the track slab, obtaining a complete point cloud dataset including the reference surface region. To improve subsequent processing efficiency, the original point cloud data can be downsampled using a voxel mesh filtering method. An appropriate voxel size (e.g., 0.5mm × 0.5mm × 0.5mm) is set, retaining one representative point within each voxel, thereby reducing the data volume while preserving geometric features.

[0097] When performing plane fitting on the point cloud data of the reference plane region, the RANSAC (Random Sample Consensus) algorithm is used for fitting calculation. This algorithm iteratively selects three points from the point cloud data to determine a plane, calculates the distance from the remaining points to the plane, and if the distance is less than a preset threshold (e.g., 0.2 mm), the point is considered an interior point. After multiple iterations (e.g., 1000 times), the plane with the most interior points is selected as the best fitting result, resulting in the plane equation Ax + By + Cz = 0, where the vector (A, B, C) is the plane normal vector.

[0098] Based on the fitted plane equation, the normal vector N=(A,B,C) is extracted as the direction of the first coordinate axis Z-axis of the local coordinate system. To ensure consistency of direction, the positive direction of the Z-axis is defined to face outward from the track plate. The geometric center of the point cloud in the reference plane region is used as the origin O of the coordinate system. The specific calculation method is to take the average of the coordinates of all points in the reference plane region: O=(∑xi / n,∑yi / n,∑zi / n), where n is the total number of points.

[0099] In determining the direction of the second coordinate axis, the length direction of the track plate is taken as the X-axis direction by using the actual installation orientation of the track plate. The maximum variation direction of the projected point set can be calculated as the X-axis direction by projecting the reference surface region point cloud in a plane perpendicular to the Z-axis through a principal component analysis (PCA) method. After obtaining the X-axis direction vector X, the third coordinate axis Y-axis direction is obtained through vector cross multiplication operation Y=Z×X, thereby establishing a complete local coordinate system.

[0100] Converting the point cloud data of the hole site region from the global coordinate system to the local coordinate system is a key step in subsequent processing. Let the coordinates of a point in the global coordinate system be Pg=(xg,yg,zg), and the coordinates in the local coordinate system be Pl=(xl,yl,zl). The conversion relationship is: Pl=R·(Pg-O), where R is a rotation matrix composed of the direction vectors of the three axes of the local coordinate system, and O is the coordinates of the origin of the local coordinate system in the global coordinate system.

[0101] For each hole site sub-region, point cloud data processing needs to be performed respectively. First, the converted point cloud data is segmented using the Euclidean clustering algorithm to separate the point cloud of each hole site sub-region. Then, the centroid coordinates of each hole site sub-region point cloud are calculated respectively as the geometric center of the hole site: C=(∑xi / m,∑yi / m,∑zi / m), where m is the number of points in the point cloud of the hole site sub-region.

[0102] Surface fitting is performed on each hole site sub-region to determine the normal vector. A quadratic surface fitting method is used to represent the hole site surface as z=ax²+by²+cxy+dx+ey+f. The parameters a, b, c, d, and e are solved by least squares method to obtain the fitted surface. The gradient direction at any point (x, y, z) on the surface can be represented as: , and after normalization, it is the normal vector direction of the point. The gradient direction at the geometric center of the hole site is taken as the normal vector N of the entire hole site.

[0103] Finally, the geometric center coordinates C and the normal vector N of each hole site are combined to form complete spatial position data. For larger diameter holes, the effective diameter can also be calculated by extracting multiple points at the edge of the hole, fitting a circle or ellipse, and obtaining diameter information to further enhance the completeness of the spatial position data.

[0104] In practical applications, to improve the calculation accuracy, the point cloud data can be preprocessed, such as removing outliers and noise points. Outliers can be removed by statistical outlier filtering method. The average distance of each point to its nearest neighbor points is calculated. If the distance exceeds 3 times the standard deviation of the global average distance, it is considered as an outlier and removed. In addition, smoothing methods such as bilateral filtering can be used to reduce the influence of noise on the calculation results while maintaining the clarity of the edge features.

[0105] In an alternative embodiment, the point-by-point correction of the drilling path parameters in the drilling process data based on the pose correction field to obtain a drilling task point sequence adapted to the actual track plate state comprises:

[0106] extracting the design coordinates and the design tool pose of each theoretical drilling point from the drilling process data;

[0107] establishing a spatial interpolation function in the pose correction field, the spatial interpolation function representing the mapping relationship between the pose deviation at any spatial position and the coordinate of the position;

[0108] inputting the design coordinates of each theoretical drilling point into the spatial interpolation function to calculate the position correction vector and the pose correction matrix at each theoretical drilling point;

[0109] applying the corresponding position correction vector to the design coordinates of each theoretical drilling point to obtain the target coordinates of each corrected drilling point on the actual track plate surface through coordinate transformation operation;

[0110] applying the corresponding pose correction matrix to the design tool pose of each theoretical drilling point to obtain the target tool pose of each corrected drilling point through rotation transformation operation;

[0111] calculating the tool pose homogeneous transformation matrix of each corrected drilling point according to the target coordinates and the target tool pose of each corrected drilling point;

[0112] organizing the tool pose homogeneous transformation matrix of each corrected drilling point in sequence according to the machining sequence constraint in the drilling process data to form the drilling task point sequence.

[0113] Extracting the design coordinates and the design tool pose of each theoretical drilling point from the drilling process data. The drilling process data is usually stored in the form of a database or a standard format file, which contains the theoretical coordinate values (x, y, z) and tool pose information of each drilling point of the track plate. The tool pose can be represented by a unit vector (hx, hy, hz) indicating the drilling direction, or by Euler angles (α, β, γ) indicating the spatial pose of the tool. The extraction process can be completed by a data parsing program to convert structured data into an operable array form for subsequent processing.

[0114] A spatial interpolation function is established in the pose correction field. This function characterizes the mapping relationship between the pose deviation at any spatial location and the coordinates of that location. The pose correction field is constructed through the aforementioned calibration point measurements and includes the position deviation vector and attitude deviation matrix of each point in space. To achieve correction calculations for arbitrary locations, a three-dimensional spatial interpolation method, such as radial basis function (RBF) interpolation, is used. For position correction, the function F(x,y,z)=(Δx,Δy,Δz) is established; for attitude correction, the function G(x,y,z)=L is established, where L is the rotation correction matrix. The establishment of the interpolation function considers the spatial distribution characteristics of the calibration points, ensuring a smooth correction transition effect in the region between calibration points.

[0115] The design coordinates of each theoretical borehole point are input into a spatial interpolation function to calculate the position correction vector and attitude correction matrix at each theoretical borehole point. For each theoretical borehole point P_i(x_i,y_i,z_i), the position correction vector of that point is calculated using the position correction interpolation function F.

[0116] ΔP_i=F(x_i,y_i,z_i)=(Δx_i,Δy_i,Δz_i); The attitude correction matrix L_i=G(x_i,y_i,z_i) for this point is calculated using the attitude correction interpolation function G. This step generates personalized correction parameters for each theoretical drilling point to adapt to the deformation state of the actual track slab.

[0117] Apply corresponding position correction vectors to the design coordinates of each theoretical drilling point, and obtain the target coordinates of each corrected drilling point on the actual track slab surface through coordinate transformation calculations. The formula for calculating the corrected drilling point coordinates P_i' is as follows:

[0118] P_i'=P_i+ΔP_i, that is, (x_i',y_i',z_i')=(x_i+Δx_i,y_i+Δy_i,z_i+Δz_i).

[0119] This transformation maps the drilling points in the theoretical design space to the corresponding positions on the actual track slab surface, ensuring the precise positioning of the drilling locations.

[0120] A corresponding attitude correction matrix is ​​applied to the designed tool attitude at each theoretical drilling point, and the target tool attitude at each corrected drilling point is obtained through rotation transformation. Assuming the designed tool attitude at the theoretical drilling point is represented by a unit direction vector h_i, the corrected tool attitude h_i' is calculated using matrix multiplication: h_i' = L_i·h_i. If Euler angles are used to represent the attitude, they must first be converted to a rotation matrix, and after applying the attitude correction, the result is converted back to Euler angles. This step ensures that the tool entry angle remains perpendicular to the actual track plate surface or meets specific process angle requirements.

[0121] According to the target coordinates and target tool poses of the modified drilling points, the tool pose homogeneous transformation matrix of each modified drilling point is calculated. The homogeneous transformation matrix provides a unified mathematical representation, which is convenient for subsequent robot or numerical control system to perform pose control.

[0122] According to the processing sequence constraints in the drilling process data, the tool pose homogeneous transformation matrix of each modified drilling point is organized in order to form a drilling task point sequence. The processing sequence constraints may consider various factors, such as minimizing the tool travel path, avoiding interference, ensuring processing quality, etc. The final drilling task point sequence can be represented as {T_1, T_2,..., T_n}, where the subscript represents the execution order. This sequence will be directly sent to the robot control system or numerical control system to guide the drilling equipment to complete all drilling tasks along the optimized path.

[0123] In practical applications, to improve system efficiency, path smoothing processing can be added after the generation of the task point sequence to ensure smooth and continuous motion trajectories between task points, avoiding sharp acceleration changes. At the same time, according to the drilling process requirements, specific approach strategies, exit strategies and drilling parameter settings can be added at each drilling point to further improve drilling quality and efficiency.

[0124] Through the above detailed steps, the conversion process from the theoretical drilling path to the drilling task point sequence adapted to the actual track plate state is completed, ensuring the accuracy and reliability of the drilling operation.

[0125] In an alternative embodiment, the drilling task point sequence is taken as a state node, and the joint space motion range, tool pose accessibility and collision safety constraints of the drilling equipment are taken as state transition conditions to generate an optimal drilling trajectory sequence that satisfies all constraint conditions through a state space search algorithm, including:

[0126] Obtain the joint angle solution set corresponding to each drilling task point in the drilling task point sequence; judge whether the joint angle in the joint angle solution set is located between the upper limit value and the lower limit value of the corresponding joint angle, and filter out the reachable drilling task points whose joint angles meet the range constraints;

[0127] Calculate the included angle between the actual tool axis direction of the end effector at each reachable drilling task point and the target tool axis direction required by the drilling task point, and filter out the executable drilling task points whose included angles are less than a preset pose deviation threshold;

[0128] Obtain the geometric boundaries of the motion components of the drilling equipment and the spatial envelope range of obstacles in the working environment, discretize and interpolate the joint angles between the adjacent points of the executable drilling task points from the starting point to the target point, and generate a number of intermediate joint states;

[0129] The spatial positions of the geometric boundaries of the moving part in each intermediate joint state are calculated, and it is determined whether the spatial positions overlap with the spatial envelope of the obstacle, and whether the motion path between adjacent point pairs exists collision;

[0130] The motion path without collision is taken as a feasible motion path, the joint angle change amount corresponding to each feasible motion path is calculated, and the weighted sum of the joint angle change amounts is taken as the motion cost of the feasible motion path;

[0131] Starting from the initial drilling task point, a path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship through a state space search algorithm, and the path combination with the minimum sum of motion costs of all feasible motion paths is selected as the optimal drilling trajectory sequence.

[0132] In the operation process of the multi-axis drilling device, in order to improve the drilling efficiency and ensure the drilling quality, it is necessary to generate an optimal drilling trajectory sequence meeting various constraint conditions according to the sequence of drilling task points.

[0133] Obtain the joint angle solution set corresponding to each drilling task point in the sequence of drilling task points. For each drilling task point, a plurality of possible joint angle solutions are obtained by inverse kinematics calculation. The inverse kinematics calculation can use an analytical method or an iterative method, wherein the analytical method is suitable for robots with specific structures, and the iterative method such as the Jacobian matrix method is suitable for complex structures. Taking a six-axis drilling robot as an example, each drilling task point can correspond to a plurality of joint angle solutions, for example, the jth drilling task point can have v solutions, denoted as {θj1, θj2,..., θjv}, wherein θjv represents the vth joint angle solution vector of the jth task point, containing the angle values of all joints.

[0134] Next, it is determined whether the joint angles in the joint angle solution set are located between the upper limit value and the lower limit value of the corresponding joint, and the reachable drilling task points with joint angles meeting the range constraints are selected. For each joint angle solution vector θjv, it is checked whether each joint angle θjvk in it meets the constraint condition: θmin,k ≤ θjvk ≤ θmax,k, wherein θmin,k and θmax,k represent the lower limit value and the upper limit value of the kth joint, respectively. Only when all joint angles meet the constraint condition, the solution is considered as a valid solution, and the corresponding drilling task point is considered as reachable.

[0135] An angle between the actual tool axis direction of the end effector at each reachable drilling task point and the target tool axis direction required by the drilling task point is calculated, and a drillable drilling task point with an angle less than a preset pose deviation threshold is screened out. Assuming that the target tool axis direction is a unit vector vt, and the actual tool axis direction corresponding to the joint angle solution θjv is a unit vector va, the angle a can be calculated by the formula a = arccos(vt·va). The preset pose deviation threshold is usually set to a small angle value, such as 0.5 degrees or 1 degree, to ensure drilling accuracy.

[0136] After determining the drillable drilling task points, the geometric boundaries of the moving parts of the drilling equipment and the spatial envelope range of the obstacles in the working environment are obtained. The moving parts of the drilling equipment can be modeled by a series of geometric shapes (such as cylinders, cuboids, etc.), and the spatial envelope range of the obstacles can be represented by point cloud data or a predefined geometric model.

[0137] For adjacent point pairs between the drillable drilling task points, the joint angles from the starting point to the target point are discretely interpolated to generate a number of intermediate joint states. The interpolation method usually uses linear interpolation, that is, uniform sampling according to a certain step size between the starting joint state θs and the target joint state θg. The selection of the step size should balance the calculation efficiency and the accuracy of collision detection, for example, it can be set to sample once every 10 milliseconds or every 0.1 radian of joint angle change.

[0138] Then, the spatial positions of the geometric boundaries of the moving parts at each intermediate joint state are calculated, and it is determined whether the spatial position overlaps with the spatial envelope range of the obstacles to determine whether there is a collision in the motion path between the adjacent point pairs. The specific method is that for each intermediate joint state, the positions and attitudes of each link are calculated according to the forward kinematics, and then it is checked whether these geometric boundaries intersect with the obstacles in the environment. The collision detection algorithm can use the hierarchical bounding box technology or the GJK algorithm to improve the detection efficiency.

[0139] The motion path without collision is taken as a feasible motion path, and the joint angle change corresponding to each feasible motion path is calculated, and the weighted sum of the joint angle change is taken as the motion cost of the feasible motion path. The motion cost can be represented as: C = w1|Δθ1| + w2|Δθ2| +... + wv|Δθv|, where Δθj represents the angle change of the jth joint, and wj is the corresponding weight coefficient. The weight coefficient is usually related to the motion inertia, energy consumption or importance of the joint, for example, the weight of the base joint is usually larger, and the weight of the end small joint is smaller.

[0140] Finally, starting from the initial drilling task point, a path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm, and a path combination with the minimum total motion cost of all feasible motion paths is selected as the optimal drilling trajectory sequence. The algorithm can be A* algorithm or Dijkstra algorithm. The algorithm uses a heuristic function to guide the search direction and improve efficiency. In the search process, a priority queue can be used to store the paths to be explored, and the paths are sorted according to the current cost plus the estimated cost to the target, and the paths with lower cost are explored preferentially.

[0141] In practical applications, algorithm parameters can be adjusted for specific drilling tasks. For example, for high-precision drilling of aviation structural parts, the attitude deviation threshold can be set smaller; for batch drilling tasks that need to be completed quickly, the weight coefficients in the motion cost calculation can be adjusted appropriately to emphasize execution efficiency. At the same time, the selection of the state space search algorithm can also be adjusted flexibly according to the task size. For small-scale tasks, exhaustive search can be used, and for large-scale tasks, heuristic or optimization algorithms such as genetic algorithm or simulated annealing algorithm can be used to find approximate optimal solutions within a reasonable time.

[0142] In an alternative embodiment, starting from the initial drilling task point, a path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm, and a path combination with the minimum total motion cost of all feasible motion paths is selected as the optimal drilling trajectory sequence, including:

[0143] A priority queue is created, and the elements in the priority queue include the current drilling task point identifier, the access sequence of the drilling task points that have passed, and the corresponding total motion cost.

[0144] The first element with the minimum total motion cost is popped out from the priority queue as the current expansion element.

[0145] It is judged whether the access sequence of the current expansion element contains all executable drilling task points. If it contains, the current expansion element is marked as a candidate solution and the subsequent expansion step is skipped. If it does not contain, the subsequent expansion step is performed.

[0146] Adjacent drilling task points directly connected with the drilling task point identifier of the current expansion element and not appearing in the access sequence of the current expansion element are retrieved in the collision-free connection relationship.

[0147] For each adjacent drilling task point, obtain the motion cost of the feasible motion path from the drilling task point of the current expansion element to the adjacent drilling task point, and add the motion cost to the total motion cost of the current expansion element to obtain the total motion cost after expansion;

[0148] Repeat the step of popping the head element from the priority queue until the current expansion element popped is marked as a candidate solution;

[0149] Select the optimal candidate solution with the minimum total motion cost after expansion from all candidate solutions, and combine the path corresponding thereto as the optimal drilling trajectory sequence.

[0150] The search process is started by creating a priority queue, each element of which contains three key information: the identification of the currently arrived drilling task point, the access sequence of the drilling task points that have passed, and the corresponding total motion cost. When the queue is initialized, the starting drilling task point is inserted as the first element in the queue, and at this time the access sequence only contains the starting point and the total motion cost is zero.

[0151] Subsequently, the head element with the minimum total motion cost is popped from the priority queue as the current expansion element. The sorting rule of the priority queue is based on the total motion cost, which ensures that the path with the minimum cost is selected for expansion each time, which conforms to the principle of optimal search.

[0152] Next, it is judged whether the access sequence of the current expansion element contains all executable drilling task points. If all task points are contained, it indicates that a complete path is found, and the expansion element is marked as a candidate solution and the subsequent expansion step is skipped; if not all task points are contained, the expansion operation is continued.

[0153] For elements that need to continue to expand, the adjacent drilling task points that are directly connected to the current drilling task point and have not appeared in the current access sequence are retrieved in the pre-established collision-free connection relationship. The connection relationship can be pre-calculated by the kinematics of the robot arm and the collision detection algorithm, and stored in the form of an adjacency matrix or an adjacency list.

[0154] For each connectable adjacent drilling task point, obtain the motion cost of the feasible motion path from the current task point to the adjacent point. The motion cost can be measured by indicators such as path length, motion time or energy consumption. Add this motion cost to the total motion cost of the current expansion element to obtain the total motion cost after expansion. Then, construct a new expansion element containing the identification of the newly arrived drilling task point, the updated access sequence (add the new task point to the end of the sequence) and the updated total motion cost, and insert this new element into the priority queue.

[0155] The step of popping the head element from the priority queue is repeatedly performed until the current extended element popped is marked as a candidate solution, that is, a complete path containing all drilling task points is found. Due to the characteristics of the priority queue, if there are multiple paths containing all task points, the paths will be found in order from small to large total motion cost.

[0156] Finally, the optimal candidate solution with the smallest total motion cost is selected from all candidate solutions, and the path corresponding to the optimal candidate solution is combined as the optimal drilling trajectory sequence. The sequence will guide the robot to complete all drilling tasks in the optimal order.

[0157] In actual application scenarios, assume that a sheet metal part needs to be drilled 10 times, and there are multiple possible motion paths between each drilling point. Through the above algorithm, the shortest drilling sequence can be quickly found considering the motion constraints of the robot and environmental obstacles.

[0158] The priority queue can be implemented using a binary heap or a Fibonacci heap to improve the efficiency of queue operations. For each extension step, if the current access sequence length is K and there are Y drilling task points, at most (Y-K) adjacent points need to be checked. By using appropriate pruning strategies, such as discarding an extended element when its total motion cost exceeds the known optimal solution, the efficiency of the algorithm can be further improved.

[0159] When dealing with large-scale drilling tasks, heuristic functions can be introduced to estimate the cost from the current state to the target state, extending the basic search algorithm to search, further improving search efficiency. The heuristic function can be designed based on the minimum spanning tree cost or the nearest neighbor distance between the remaining unvisited points.

[0160] Through this state space search-based method combined with robot kinematics and path planning techniques, the optimal execution sequence for multi-point drilling tasks in complex industrial environments can be effectively found, improving production efficiency and reducing energy consumption.

[0161] The track plate drilling parameter intelligent optimization system based on deep learning of the embodiment of the application comprises:

[0162] The first unit is configured to obtain three-dimensional scanning data of a track plate and preset drilling process data, identify a feature region of the three-dimensional scanning data by using a deep learning semantic segmentation network, and determine a reference surface region and a hole site region of the track plate according to a region probability distribution corresponding to each feature.

[0163] The second unit is configured to establish a local coordinate system of the track plate according to the reference surface region, and perform geometric center calculation and normal vector estimation on the hole site region in the local coordinate system to determine spatial position data of each hole site, calculate a position deviation vector and construct a pose correction field by comparing the spatial position data with standard hole site data in the drilling process data.

[0164] The third unit is configured to correct drilling path parameters in the drilling process data point by point based on the pose correction field to obtain a drilling task point sequence adapted to an actual track plate state.

[0165] The fourth unit is configured to take the drilling task point sequence as a state node, take joint space motion range, tool posture accessibility and collision safety constraints of a drilling device as state transition conditions, and generate an optimal drilling trajectory sequence meeting all constraint conditions through a state space search algorithm.

[0166] The fifth unit is configured to control the drilling device to perform track plate drilling operation according to the optimal drilling trajectory sequence.

[0167] A third aspect of the embodiment of the present application,

[0168] An electronic device is provided, comprising:

[0169] a processor;

[0170] a memory for storing processor-executable instructions;

[0171] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0172] A fourth aspect of the embodiment of the present application,

[0173] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0174] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0175] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to 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 of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent optimization of track panel drilling parameters based on deep learning, characterized in that, The method comprises the following steps: acquiring three-dimensional scanning data of a track plate and preset drilling process data, using a deep learning semantic segmentation network to identify feature regions of the three-dimensional scanning data, and determining a reference surface region and a hole site region of the track plate according to a region probability distribution corresponding to each feature; establishing a local coordinate system of the track plate according to the reference surface region, and performing geometric center calculation and normal vector estimation on the hole site region in the local coordinate system to determine spatial position data of each hole site, and calculating a position deviation vector and constructing a pose correction field by comparing the spatial position data with standard hole site data in the drilling process data; based on the pose correction field, performing point-by-point correction on drilling path parameters in the drilling process data to obtain a drilling task point sequence that adapts to the actual state of the track plate, including: extracting design coordinates and design tool poses of each theoretical drilling point from the drilling process data; establishing a spatial interpolation function in the pose correction field, the spatial interpolation function representing a mapping relationship between a pose deviation amount at any spatial position and coordinates of the position; inputting the design coordinates of each theoretical drilling point into the spatial interpolation function to calculate a position correction vector and a pose correction matrix at each theoretical drilling point; applying the corresponding position correction vector to the design coordinates of each theoretical drilling point to obtain target coordinates of each corrected drilling point on the actual track plate surface through coordinate transformation operation; applying the corresponding pose correction matrix to the design tool pose of each theoretical drilling point to obtain a target tool pose of each corrected drilling point through rotation transformation operation; calculating a tool pose homogeneous transformation matrix of each corrected drilling point according to the target coordinates and the target tool pose of each corrected drilling point; organizing the tool pose homogeneous transformation matrices of each corrected drilling point in sequence according to the machining sequence constraint in the drilling process data to form the drilling task point sequence; taking the drilling task point sequence as a state node, and taking joint space motion range, tool pose accessibility and collision safety constraint of a drilling device as state transition conditions, an optimal drilling trajectory sequence that satisfies all constraint conditions is generated through a state space search algorithm; controlling the drilling device to perform track plate drilling operation according to the optimal drilling trajectory sequence.

2. The method of claim 1, wherein, The method for identifying feature regions of the three-dimensional scanning data using a deep learning semantic segmentation network and determining a reference surface region and a hole site region of the track plate according to a region probability distribution corresponding to each feature comprises the following steps: perform voxelization processing on the three-dimensional scanning data to obtain voxelized representation data, input the voxelized representation data into a deep learning semantic segmentation network, perform multi-level feature extraction on the voxelized representation data through convolution operation, up-sample and fuse the multi-level features, align the features at different levels in the spatial dimension and perform channel splicing to obtain fused features; perform classification operation on each spatial position of the fused features to calculate region probability distributions of the position belonging to a reference surface category, a hole site category and a background category, and select the category with the largest region probability distribution as the semantic label of the position; According to the semantic label, voxels with the same category label and spatially connected are subjected to connected component analysis, the spatial range of each connected component is identified, the connected component corresponding to the reference surface category label is determined as the reference surface region, and each connected component corresponding to the hole site category label is determined as the hole site region.

3. The method of claim 1, wherein, A local coordinate system of the track plate is established according to the reference surface region, and the geometric center calculation and normal vector estimation of the hole site region are performed in the local coordinate system to determine the spatial position data of each hole site, including: Point cloud data of the reference surface region is extracted, plane fitting operation is performed on the point cloud data, and a fitting plane equation of the reference surface is obtained; According to the fitting plane equation, the normal vector of the reference surface is extracted as the direction of the first coordinate axis of the local coordinate system, and a geometric center point in the point cloud data of the reference surface region is selected as the origin of the local coordinate system; In a plane perpendicular to the direction of the first coordinate axis, the length direction of the track plate is determined as the direction of the second coordinate axis, the third coordinate axis direction is obtained through the vector cross product operation of the first coordinate axis direction and the second coordinate axis direction, and the local coordinate system is established; The point cloud data of each hole site sub-region in the hole site region is converted from the global coordinate system to the local coordinate system; The point cloud data of each hole site sub-region after conversion is subjected to centroid calculation respectively to obtain the geometric center coordinates of each hole site sub-region in the local coordinate system; The point cloud data of each hole site sub-region is subjected to surface fitting respectively, the surface gradient direction of each hole site sub-region is calculated, and the surface gradient direction is taken as the normal vector of each hole site sub-region; The geometric center coordinates and the normal vector of each hole site are combined to form the spatial position data.

4. The method of claim 1, wherein, The drilling task point sequence is taken as a state node, the joint space motion range of the drilling device, the tool posture reachability and the collision safety constraint are taken as state transition conditions, and an optimal drilling trajectory sequence satisfying all constraint conditions is generated through a state space search algorithm, including: Obtain the joint angle solution set corresponding to each drilling task point in the drilling task point sequence; Determine whether the joint angle in the joint angle solution set is located between the upper limit value and the lower limit value of the corresponding joint angle one by one, and filter out reachable drilling task points whose joint angles meet the range constraint; Calculate the included angle between the actual tool axis direction of the end effector at each reachable drilling task point and the target tool axis direction required by the drilling task point, and filter out executable drilling task points whose included angles are less than a preset posture deviation threshold; Obtain the spatial envelope range of the motion component geometric boundary of the drilling device and the obstacles in the working environment, discretize and interpolate the joint angle from the starting point to the target point for each adjacent point pair between the executable drilling task points, and generate a plurality of intermediate joint states; Calculate the spatial position of the motion component geometric boundary at each intermediate joint state, determine whether the spatial position overlaps with the spatial envelope range of the obstacles, and determine whether there is a collision in the motion path between the adjacent point pairs. The motion path without collision is taken as a feasible motion path, the joint angle change amount corresponding to each feasible motion path is calculated, and a weighted sum of the joint angle change amount is taken as a motion cost of the feasible motion path; A path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm starting from a starting drilling task point, and a path combination with a minimum sum of motion costs of all feasible motion paths is selected as the optimal drilling trajectory sequence.

5. The method of claim 4, wherein, The path combination sequentially passing through all executable drilling task points is searched in the collision-free connection relationship by a state space search algorithm starting from a starting drilling task point, and a path combination with a minimum sum of motion costs of all feasible motion paths is selected as the optimal drilling trajectory sequence, including: A priority queue is created, and an element in the priority queue contains a currently arrived drilling task point identifier, a drilling task point access sequence that has passed, and a corresponding total motion cost; A first element in the priority queue with a minimum total motion cost is popped out as a current expansion element; It is judged whether the access sequence of the current expansion element contains all executable drilling task points, if yes, the current expansion element is marked as a candidate solution and the subsequent expansion step is skipped, and if not, the subsequent expansion step is performed; Adjacent drilling task points directly connected with the drilling task point identifier of the current expansion element and not appearing in the access sequence of the current expansion element are searched in the collision-free connection relationship; For each adjacent drilling task point, a motion cost of a feasible motion path from the drilling task point of the current expansion element to the adjacent drilling task point is obtained, and the motion cost is added to the total motion cost of the current expansion element to obtain an expanded total motion cost; The step of popping out the first element in the priority queue is repeatedly performed until the current expansion element popped out is marked as a candidate solution; An optimal candidate solution with a minimum expanded total motion cost is selected from all candidate solutions, and a path combination corresponding to the optimal candidate solution is taken as the optimal drilling trajectory sequence.

6. A deep learning based track panel drilling parameter intelligent optimization system for implementing the method of any one of claims 1-5, characterized in that, Including: The first unit is configured to obtain three-dimensional scanning data of a track plate and preset drilling process data, perform feature region identification on the three-dimensional scanning data by using a deep learning semantic segmentation network, and determine a reference surface region and a hole site region of the track plate according to a region probability distribution corresponding to each feature. The second unit is configured to establish a local coordinate system of the track plate according to the reference surface region, perform geometric center calculation and normal vector estimation on the hole site region in the local coordinate system, determine spatial position data of each hole site, calculate a position deviation vector by comparing the spatial position data with standard hole site data in the drilling process data, and construct a pose correction field. The third unit is configured to correct drilling path parameters in the drilling process data point by point based on the pose correction field, and obtain a drilling task point sequence adapted to an actual state of the track plate. The fourth unit is configured to generate an optimal drilling trajectory sequence satisfying all constraint conditions by a state space search algorithm, taking the drilling task point sequence as a state node, and taking joint space motion range, tool posture reachability and collision safety constraints of the drilling device as state transition conditions. The fifth unit is configured to control the drilling device to perform the orbital plate drilling operation according to the optimal drilling trajectory sequence.

7. An electronic device, comprising: The computer program product comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 5.

Images