Mould global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning, computing equipment and readable storage medium

Abstract

The invention provides a mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning, a multi-sensor detection platform is applied, and the method comprises the following steps: loading a mold; path planning: calculating the CAD model of the to-be-detected mold through a genetic algorithm to generate scanning path data acquisition; scanning and collecting multi-modal data of multiple angles of the to-be-detected mold; a dynamic calibration step; s5, a data fusion step: fusing the multi-modal data to generate a global model of the to-be-tested mold; and S6, defect identification: extracting geometric features and density features of the global model, and comparing an extraction result with the CAD model to generate a deviation chromatogram.

Description

Technical Field

[0001] The present invention relates to the field of precision industrial detection technology, and in particular to a mold global intelligent detection method, computing device and readable storage medium based on multi-sensor dynamic calibration and semantic optimization learning. Background Art

[0002] A mold (or die) is a forming tool used in the mass production of industrial products. Through a cavity or structure of a specific shape, it processes raw materials (such as metal, plastic, rubber, glass, etc.) into parts or products with a fixed shape, size, and surface quality. There are many different types of molds with a wide range of applications, covering almost all manufacturing sectors.

[0003] Mold inspection is a key step in mold manufacturing, use, and maintenance, directly affecting product quality and production efficiency. Due to the complex structure of the mold cavity and high precision requirements, the following core issues and solutions are often encountered during the inspection process:

[0004] Data fragmentation problem: Traditional single sensors (such as optical scanners and industrial CT) have a limited detection range and cannot cover the entire structure of complex molds, resulting in a blind spot rate of more than 25%. It is particularly difficult to detect hidden areas such as deep cavities and narrow flow channels.

[0005] Accumulated stitching errors: Multi-view data stitching relies on the traditional ICP (Iterative Closest Point) algorithm, resulting in a global error of more than 0.1mm, affecting detection accuracy.

[0006] Low defect recognition efficiency: Manual interpretation is time-consuming and the missed detection rate exceeds 15%. The existing AI model has insufficient generalization ability for small defects (<50μm) and has a high misjudgment rate.

[0007] In summary, existing technologies cannot effectively solve problems such as data fragmentation, splicing errors, and low defect recognition efficiency in mold inspection. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a mold full-area intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning, so as to achieve full-area coverage detection of molds, enhance defect recognition capability and improve detection efficiency.

[0009] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0010] A global intelligent mold detection method based on multi-sensor dynamic calibration and semantic optimization learning, using a multi-sensor detection platform, includes the following steps:

[0011] S1 mold loading step: Place the mold to be inspected on the multi-sensor inspection platform, and the laser tracker on the multi-sensor inspection platform locates the mold reference point in real time;

[0012] S2 path planning step: Genetic algorithm is used to calculate the CAD model of the mold to be tested and generate a scanning path covering the entire CAD model of the mold to be tested;

[0013] S3 data acquisition step: using multiple sensors on the multi-sensor detection platform to scan and collect multimodal data of multiple angles of the mold to be tested through the scanning path;

[0014] S4 dynamic calibration step: using the laser tracker as a reference, establishing a global coordinate system on the multi-sensor detection platform and uniformly mapping multimodal data from multiple angles to the global coordinate system;

[0015] S5 data fusion step: using a multi-view semantic splicing algorithm to fuse the multimodal data of multiple angles of the mold to be tested mapped to the global coordinate system to generate three-dimensional point cloud data of the global model of the mold to be tested;

[0016] S6 defect identification step: Use a multi-task Bayesian deep learning model to extract the geometric and density features of the global model, compare the extraction results with the CAD model, and generate a deviation chromatogram.

[0017] Preferably, the multiple sensors include at least a blue light scanner, an industrial CT and a micro optical probe.

[0018] Preferably, the S2 path planning step includes the following steps:

[0019] Step S2-1: Input the mold CAD model and extract the boundaries and features of the key structures of the mold CAD model to be tested through a boundary recognition algorithm;

[0020] Step S2-2: Using a genetic algorithm to scan the CAD model of the mold to be tested, and with the goal of maximizing the coverage of the key components and minimizing the scanning time, generating an optimal scanning path for multi-sensor collaboration.

[0021] Preferably, the key structure includes at least an inner cavity, a parting surface, and a flow channel;

[0022] The criterion for maximizing the coverage is that the target blind area rate of key components is less than 1%.

[0023] Preferably, the data collection step in step S3 includes the following collection methods:

[0024] Blue light scanner acquisition: Continuously scan the mold surface. Each frame of the point cloud contains approximately several three-dimensional coordinate points. The robot arm moves to achieve seamless splicing for scanning and acquisition.

[0025] Industrial CT acquisition: non-destructive scanning of the mold to be tested to generate three-dimensional tomographic images;

[0026] Micro-optical probe acquisition: The deep cavity surface morphology is collected through laser confocal technology, and scanning and acquisition are performed in conjunction with the six-degree-of-freedom movement of the robotic arm.

[0027] Preferably, the S4 dynamic calibration includes the following steps:

[0028] S41. Reference Coordinate System Construction: Using the laser tracker as a reference, a global coordinate system is established on the multi-sensor detection platform. The target ball on the motion platform is tracked by laser beams to obtain the sensor's position and status in real time.

[0029] S42. Dynamic sensor calibration: The sensor is fixed to a multi-sensor detection platform and a standard calibration component is scanned in multiple postures. The standard calibration component is used for sensor calibration and coordinate system calibration. The transformation matrix from the sensor coordinate system to the global coordinate system is solved using the least squares method and iterative optimization:

[0030] X 全局 =R·X 传感器 +T, where R is the rotation matrix and T is the translation vector.

[0031] S43. Deviation elimination and compensation: Use a laser tracker to correct sensor motion errors in real time, and introduce a temperature compensation model to correct thermal deformation.

[0032] Preferably, the S5 data fusion comprises the following steps:

[0033] S51. Coordinate alignment: Based on the multi-sensor dynamic calibration results, the multi-modal data of multiple angles of the mold to be tested are uniformly mapped to the global coordinate system;

[0034] S52. Semantic Feature Enhancement: Using an improved PointNet++ network, we perform semantic segmentation on the multimodal data collected from multiple angles of the mold under test. We extract key component feature points and assign semantic labels to them. We then select semantically consistent feature points for registration.

[0035] S52. Layered registration strategy:

[0036] (1) Coarse registration: Initial alignment is performed based on the global geometric distribution of semantic feature points to reduce the positional differences of multimodal data at multiple angles of the mold to be tested;

[0037] (2) Precise registration: Improve the ICP algorithm and use a defined semantic weight function to optimize the matching of corresponding points.

[0038] Define the semantic weight function w i =α·s i +β·ci , where s i is the semantic confidence of the point, c i is the curvature value of the point.

[0039] S53. Error Optimization: The Huber loss function is iteratively optimized to reduce the influence of outliers and achieve a global stitching error of <0.05 mm. Finally, all registered multimodal data are fused to generate a 3D point cloud of the global model of the mold to be tested.

[0040] Preferably, the S6 defect identification comprises the following steps:

[0041] S61 Multi-task Bayesian Model: The fused 3D point cloud data and CT image of the global model of the mold to be tested are input into the multi-task Bayesian model. The geometric features of the point cloud and the density features of the CT image of the mold to be tested are extracted through the 3D ResNet-50 network.

[0042] S62 thermal deformation compensation: Based on the thermal expansion coefficient of the mold material, combined with real-time temperature data, dynamic correction of the test results obtained by the S61;

[0043] S63. Result visualization and report generation: Compare the 3D point cloud data of the global model of the mold to be tested generated in the S5 data fusion step with the CAD model to generate a deviation chromatogram.

[0044] The present invention also provides a computing device, comprising:

[0045] one or more processors;

[0046] A storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a method for global intelligent detection of molds based on multi-sensor dynamic calibration and semantic optimization learning as described above.

[0047] The present invention also provides a computer-readable storage medium, which stores a program. When the program is executed by a processor, it implements the method described in the above-mentioned mold full-domain intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning.

[0048] Beneficial effects of the invention: The present invention collects data through multiple sensors and uses genetic algorithms to optimize the scanning path, achieving a mold inspection scanning blind area rate of less than 1%, supporting blind angle detection from micron-level flow channels to 5-meter-level molds;

[0049] High-precision detection: Using multi-sensor dynamic calibration and semantic stitching technology to achieve a global error of <0.05mm, a defect detection rate of >99%, and a false positive rate of <1%;

[0050] Supports online automated testing and thermal deformation compensation, shortening testing time by more than 80%, greatly improving testing efficiency;

[0051] Adaptability and compatibility: supports seamless switching of molds of different sizes and can be integrated into the production line to achieve real-time quality control;

[0052] Cost optimization: Reduce manual intervention and mold repair times, and extend mold service life. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning of the present invention. DETAILED DESCRIPTION

[0054] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0055] like Figure 1 As shown, the present invention provides a mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning, which applies a multi-sensor detection platform. The multi-sensor detection platform includes multiple sensors, reference equipment, and motion and control units, wherein the multiple sensors include:

[0056] Main sensor: High-precision blue light scanner (GOM ATOS Core 350, accuracy ±0.005mm, scanning range 350mm×260mm), used for rapid acquisition of the mold surface topography to be tested;

[0057] Auxiliary sensor: A micro-optical probe (Keyence VK-X200K, resolution 0.1μm), mounted on a six-degree-of-freedom robotic arm (repeatability ±0.02mm), is used for detailed scanning of deep cavities and narrow flow channels (micropores with a diameter of ≤2mm) in the mold to be tested.

[0058] Industrial CT system: high-resolution X-ray tomography scanner (such as Nikon XT H 225ST, accuracy ±0.01mm, maximum scanning workpiece size φ500mm×500mm) for non-destructive 3D reconstruction of closed flow channels.

[0059] In addition, the benchmark equipment includes: a laser tracker (Leica AT960, ranging accuracy ±(10μm+1.5μm / m)), which cooperates with the global target sphere array to establish a global coordinate system.

[0060] Motion and control unit: A dual-station servo motion platform is used to support automatic loading and unloading of molds and rotation / translation motion (travel range: X / Y / Z axis ±2m, rotation accuracy ±5arcsec). The sensor and the robotic arm communicate in real time via the EtherCAT bus, enabling sub-millimeter motion control.

[0061] Based on the above multi-sensor detection platform, the method of the present invention includes the following steps:

[0062] S1 mold loading step: Place the mold to be inspected on the multi-sensor inspection platform, and the laser tracker on the multi-sensor inspection platform locates the mold reference point in real time;

[0063] S2 path planning step: Genetic algorithm is used to calculate the CAD model of the mold to be tested and generate a scanning path covering the entire CAD model of the mold to be tested;

[0064] Specifically, the specific steps of path planning are as follows:

[0065] Step S21: Input the CAD model of the mold to be tested, and extract the boundaries and features of the key structures of the CAD model of the mold to be tested through the boundary recognition algorithm. The key structures here include at least the inner cavity, parting surface, runner and other components of the mold to be tested. The boundary recognition algorithm locates the edge or contour by detecting discontinuities in the data (such as sudden changes in pixel values ​​and gradient changes), which is a prior art algorithm.

[0066] Step S22: Using a genetic algorithm to scan the CAD model of the mold to be tested, and with the goal of maximizing the coverage of the key components and minimizing the scanning time, generate an optimal scanning path for multi-sensor collaboration. Here, the target blind area rate for maximizing the coverage is <1%.

[0067] Specifically, the genetic algorithm is a global optimization algorithm that simulates biological natural selection and genetic mechanisms. It searches for the optimal solution in the solution space through operations such as encoding, selection, crossover, and mutation. Specifically, in this invention, the genetic algorithm is used to convert the inspection requirements of the mold CAD model into multi-sensor scanning path planning. With the goals of maximizing key component coverage (blind spot rate <1%) and minimizing scanning time, the optimal scanning path for multi-sensor collaboration is generated through iterative optimization.

[0068] The optimal scanning path for multi-sensor collaboration is generated by encoding a sequence of sensor poses (position and state). The fitness function combines path length, overlap, and blind spot coverage. The sensor pose sequence uses real-valued encoding, directly representing the spatial pose of the sensor as continuous numerical values, avoiding the accuracy loss and decoding complexity of binary encoding.

[0069] S3 data acquisition step: using multiple sensors on the multi-sensor detection platform to scan and collect multimodal data of multiple angles of the mold to be tested through the scanning path;

[0070] Specifically, the acquisition method of the present invention adopts a multi-sensor collaborative acquisition method:

[0071] Blue light scanner: Continuously scans the mold surface at a frequency of 10Hz. Each frame of the point cloud contains approximately 1 million three-dimensional coordinate points, and seamless stitching within a range of 2m×2m is achieved through the movement of the robotic arm.

[0072] Industrial CT: The mold was scanned non-destructively with the X-ray tube voltage set to 180 kV, the current to 200 μA, and the exposure time to 2 s / layer, generating a 3D tomographic image of 512 × 512 × 512 voxels.

[0073] Micro optical probe: It uses laser confocal technology to collect deep cavity surface morphology, with a lateral resolution of 0.5μm and a longitudinal resolution of 0.1μm. Combined with the six-degree-of-freedom movement of the robotic arm, it covers the blind spots of traditional optical scanning.

[0074] Perform real-time data calibration on the collected data:

[0075] The laser tracker monitors the robot arm's posture (position and status) in real time, updates the sensor coordinate system conversion matrix every 50ms, and compensates for motion errors; temperature sensors (accuracy of ±0.1°C) are distributed on multiple sensor detection platforms to correct the impact of ambient temperature on sensor and mold thermal deformation in real time.

[0076] Specifically, a sensor adaptation strategy is also adopted in the data collection process of the mold to be tested:

[0077] For the open area of ​​the mold to be tested, in this embodiment, the open area is selected as the mold surface, and a blue light scanner is preferably used to perform surface scanning at a speed of 50 mm / s, with a point cloud sampling interval of 0.1 mm.

[0078] For deep mold cavities and narrow flow channels (such as microholes with a diameter of ≤2mm), the system switches to a micro-optical probe and uses a robotic arm to perform spiral scanning at a speed of 0.1mm / s with a step size of 0.05mm to ensure detail resolution.

[0079] The industrial CT performs spiral scanning on the closed flow channel of the mold to be tested, with the layer thickness set to 0.1mm and the reconstruction accuracy reaching ±0.01mm.

[0080] S4 dynamic calibration step: using the laser tracker as a reference, a global coordinate system is established on the multi-sensor detection platform and multimodal data of multiple angles are uniformly mapped to the global coordinate system.

[0081] Specifically, the S4 dynamic calibration step includes the following steps:

[0082] S41. Reference Coordinate System Construction: Using the laser tracker as a reference, a global coordinate system is established on the multi-sensor detection platform. (A global coordinate system is a fixed reference frame used to uniformly describe the spatial relationships of objects, locations, or events throughout a system or scene. It serves as the basis for all other local coordinate systems, ensuring that the positions and orientations of different objects can be accurately expressed and calculated in the same space.) Using laser beams to track a target ball on the motion platform, the sensor position and status are determined in real time.

[0083] S42. Sensor dynamic calibration: Fix the sensor to the multi-sensor detection platform, perform multi-pose scanning on the standard calibration parts, and solve the transformation matrix from the sensor coordinate system to the global coordinate system through the least squares method and iterative optimization:

[0084] X 全局 =R·X 传感器 +T, where R is the rotation matrix and T is the translation vector.

[0085] Specifically, during the calibration phase, the robotic arm drives the sensor to scan a standard calibration part (such as a ball paddle) in five different postures (pitch angle ±15°, rotation angle 0° / 90° / 180°) to obtain the point cloud data of the standard part in the sensor coordinate system (point cloud data is a collection of data points defined in three-dimensional space. These points usually represent the geometric shape of the object surface or space), and solve the transformation matrix by comparing it with the theoretical coordinates in the global coordinate system.

[0086] Specifically, the standard calibration part is a physical model with known high-precision geometric features, which is used for sensor calibration and coordinate system calibration, including but not limited to:

[0087] Ball plate calibration component: There are multiple high-precision spheres distributed on the surface (diameter tolerance ±0.001mm), and the distance between the sphere centers is known with an accuracy of ±0.002mm;

[0088] Cylindrical array calibration component: Contains regularly arranged high-precision cylindrical holes, with the theoretical values ​​of the cylinder diameter and hole spacing known (accuracy ±0.005mm);

[0089] Plane grid calibration component: The surface is etched with micron-level precision grid lines, which are used to calibrate the plane geometry measurement accuracy of the sensor.

[0090] S43. Bias Elimination and Compensation: A laser tracker is used to correct sensor motion errors in real time, and a temperature compensation model is introduced to correct thermal deformation. The calibration error is converged to within ±0.01 mm through iterative optimization using the Levenberg-Marquardt algorithm, ensuring global measurement accuracy ≤±0.01 mm.

[0091] And the accuracy of the global coordinate system is ±0.001mm.

[0092] S5 data fusion step: using a multi-view semantic splicing algorithm to fuse the multimodal data of multiple angles of the mold to be tested mapped to the global coordinate system to generate three-dimensional point cloud data of the global model of the mold to be tested;

[0093] Preferably, the S5 data fusion step includes the following steps:

[0094] S51. Coordinate alignment: Based on the multi-sensor dynamic calibration results, the multi-modal data of multiple angles of the mold to be tested are uniformly mapped to the global coordinate system;

[0095] S52. Semantic Feature Enhancement: Use the PointNet++ network to perform semantic segmentation on multimodal data mapped to the global coordinate system, extract key component feature points and assign semantic labels, and select semantically consistent feature points for registration.

[0096] Specifically, the multimodal data here is the multi-view point cloud data of the mold to be tested collected by multiple sensors, which is semantically segmented. The feature points of key components such as the inner cavity surface, parting line, and runner edge of the mold to be tested are extracted and assigned semantic labels (such as "sealing surface" and "runner wall"). Only feature points with consistent semantic labels are selected for alignment.

[0097] The registration refers to the process of aligning multi-view point clouds in a spatial coordinate system, specifically including S53. a layered registration step.

[0098] The semantic segmentation method is: enhance the local geometric feature expression through hierarchical feature learning (SAM module). The network is pre-trained on the ImageNet-3D dataset and fine-tuned for the mold detection task to perform semantic segmentation, with a semantic segmentation accuracy of 98.7%.

[0099] S53. Layered registration strategy:

[0100] (1) Coarse registration: Initial alignment is performed based on the global geometric distribution of semantic feature points to reduce the positional differences in the multi-view point cloud data of the mold to be tested;

[0101] Specifically, PCA alignment: PCA principal components are calculated for each viewpoint data cloud, the principal component direction is used as the initial alignment reference, and the first principal component direction is used as the global coordinate axis to achieve preliminary alignment of the point cloud posture.

[0102] 4PCS algorithm: Randomly select 4 point groups from semantic feature points, quickly estimate the rotation and translation parameters between point clouds through geometric consistency matching (such as distance and angle constraints), and generate a coarse registration matrix.

[0103] Semantic labels are used to filter out feature points such as parting lines and runner edges, and the 4PCS (4Points CongruentSets) algorithm is used for rapid coarse registration, reducing the initial position error between point clouds to within 1mm.

[0104] (2) Precise registration: Improve the ICP algorithm and use the semantic weight function to optimize the matching of corresponding points. Define the semantic weight function w i =α·s i +β·c i , where s i is the semantic confidence of the point (output by PointNet++), c i The curvature value of the point is given a higher weight (α = 0.7, β = 0.3) for key areas such as the sealing surface. The semantic weight and normal vector constraint are used to optimize the matching of corresponding points. The rotation matrix R and translation vector T are iteratively calculated. The optimization is stopped when the root mean square error (RMSE) between the point cloud data is less than 0.05mm, so that the global stitching error is less than 0.05mm.

[0105] S53. Error Optimization: We iteratively optimize the Huber loss function to reduce the influence of outliers, achieving a global stitching error of less than 0.05 mm. Ultimately, we fuse all registered multimodal data to generate a 3D point cloud of the global model of the mold under test. Global stitching error refers to the spatial positional deviation caused by factors such as coordinate alignment and registration algorithms during the fusion process of 3D point cloud data collected by multiple sensors from different angles.

[0106] Specifically, a point cloud normal vector constraint is introduced (the point cloud normal vector is a unit vector perpendicular to the local surface, estimated by fitting a plane through PCA analysis of neighborhood points. Its direction needs to be uniformly adjusted, and it plays a core role in tasks such as reconstruction, rendering, and segmentation. When searching for corresponding points, points with consistent normal vector directions are prioritized to improve the robustness of the registration.

[0107] The Huber loss function is used to replace the traditional mean square error to reduce the influence of outliers on the registration results, and the iteration is performed until the stitching error is <0.05mm.

[0108] S6 defect recognition step: Use the multi-task Bayesian deep learning model to extract the geometric features and density features of the global model and calculate the detection results (inner cavity surface deviation, parting surface gap, runner size deviation).

[0109] Preferably, the S6 defect identification step includes the following steps:

[0110] S61 Multi-task Bayesian Model: The input of the multi-task Bayesian network is the 3D point cloud data of the global model of the mold to be tested generated by the data fusion step in step S5 and the 3D tomographic image collected by industrial CT. The geometric features (such as curvature and roughness) of the 3D point cloud data of the mold to be tested and the density features of the industrial CT image are extracted. Finally, the results such as the inner cavity surface deviation, parting surface gap, and runner size deviation of the mold to be tested are output. The core design of the multi-task Bayesian model is a shared feature extraction layer + task-specific output layer. The specific structure is as follows:

[0111] (1) Shared feature extraction layer: Using the 3D ResNet-50 network, it is responsible for extracting common features such as mold geometric contours and material density distribution from the three-dimensional point cloud data and CT images of the mold to be tested.

[0112] (2) Task-specific output layer: For the three subtasks (inner cavity surface deviation prediction, parting surface gap calculation, and runner size measurement), independent fully connected layers are designed to output the dedicated prediction results of each task. The connection layer of each subtask contains the following structure:

[0113] Input layer: receives the output feature vector from the shared feature extraction layer (3D ResNet-50), with dimension D (determined by the global average pooling layer of ResNet-50);

[0114] Hidden layer: This layer has 1 layer and H = D / 2 neurons. Its purpose is to reduce the amount of computation and avoid overfitting by reducing dimensionality. This layer uses the ReLU activation function and introduces nonlinear transformations to capture complex relationships between features.

[0115] Bayesian uncertainty layer: Monte Carlo dropout is used to simulate the uncertainty distribution of model parameters by randomly inactivating neurons.

[0116] Output layer: Fully connected using one neuron, with linear activation as the activation function. The output of the inner cavity surface deviation prediction subtask is the deviation values ​​of the three spatial coordinate axes; the output of the parting surface gap calculation subtask is the maximum gap value and uniformity standard deviation; the output of the runner dimension measurement subtask is the runner cross-sectional dimensions (depth, width, fillet radius) and surface roughness (Ra value).

[0117] (3) The loss function adopts a weighted sum multi-task loss function: λ i Represents the weight coefficient, L i Represents the loss function of each subtask. The loss function of each subtask is as follows:

[0118] The inner cavity surface deviation prediction subtask uses the mean square error loss function, which is expressed as:

[0119]

[0120] Where N represents the number of points in the three-dimensional point cloud of the inner cavity surface, is the three-dimensional deviation vector predicted by the model, Indicates the actual deviation value between the CAD model and the point cloud.

[0121] For the parting surface gap calculation subtask, the loss function uses mean square error + log mean square error:

[0122]

[0123] Among them, M represents the number of points in the three-dimensional point cloud of the parting surface, represents the prediction gap value, It is the actual measured gap value after virtual mold closing.

[0124] For the flow channel size measurement subtask, the loss function uses Huber loss:

[0125]

[0126] Where P is the number of flow channel size parameters, is the predicted size parameter; is the actual size measured by industrial CT; δ is the robust threshold.

[0127] S62. Thermal deformation compensation:

[0128] Based on the thermal expansion coefficient of the mold material and in combination with real-time temperature data, the test results obtained in S61 are dynamically corrected. For example, when the mold temperature is 20°C higher than the reference temperature, a compensation of 0.002mm is added to the parting surface gap measurement value.

[0129] S63. Results Visualization and Report Generation: Compares the 3D point cloud data of the full-domain mold model generated in the S5 data fusion step with the CAD model, generating a deviation spectrum (with a color scale accuracy of 0.001mm) to visually display out-of-tolerance areas. Automatically outputs an inspection report containing key indicators such as the internal cavity surface pass rate, maximum parting surface clearance, and runner dimension Cpk value. Export is supported in PDF and XML formats.

[0130] Through actual experiments, the mold inspection effect of this embodiment is as follows: in a certain automobile mold inspection, this system successfully identified a 30μm runner corner depression that was missed by traditional methods. The inspection time was only 28 minutes, which was 88.3% shorter than the traditional method (240 minutes).

[0131] The accuracy statistics of repeated detection 10 times showed that the standard deviation of the inner cavity surface deviation measurement was 0.008mm, and the standard deviation of the parting surface gap measurement was 0.002mm, verifying the high repeatability of the system.

[0132] The present invention also provides a computing device, comprising:

[0133] one or more processors;

[0134] A storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a method for global intelligent detection of molds based on multi-sensor dynamic calibration and semantic optimization learning as described above.

[0135] The present invention also provides a computer-readable storage medium, which stores a program. When the program is executed by a processor, it implements the method described in the above-mentioned mold full-domain intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning.

[0136] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning, characterized by: A multi-sensor detection platform was applied, and the method included the following steps: S1 mold loading step: Place the mold to be inspected on the multi-sensor inspection platform, and the laser tracker on the multi-sensor inspection platform locates the mold reference point in real time; S2 path planning step: Genetic algorithm is used to calculate the CAD model of the mold to be tested and generate a scanning path covering the entire CAD model of the mold to be tested; S3 data acquisition step: using multiple sensors on the multi-sensor detection platform to scan and collect multimodal data of multiple angles of the mold to be tested through the scanning path; S4 dynamic calibration step: using the laser tracker as a reference, establishing a global coordinate system on the multi-sensor detection platform and uniformly mapping multimodal data from multiple angles to the global coordinate system; S5 data fusion step: using a multi-view semantic splicing algorithm to fuse the multimodal data of multiple angles of the mold to be tested mapped to the global coordinate system to generate three-dimensional point cloud data of the global model of the mold to be tested; S6 defect identification step: Use a multi-task Bayesian deep learning model to extract the geometric and density features of the global model, compare the extraction results with the CAD model, and generate a deviation chromatogram.

2. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The multiple sensors include at least a blue light scanner, an industrial CT and a micro optical probe.

3. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The S2 path planning step includes the following steps: Step S2-1: Input the mold CAD model and extract the boundaries and features of the key structures of the mold CAD model to be tested through a boundary recognition algorithm; Step S2-2: Using a genetic algorithm to scan the CAD model of the mold to be tested, and with the goal of maximizing the coverage of the key components and minimizing the scanning time, generating an optimal scanning path for multi-sensor collaboration.

4. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 3 is characterized in that: The key structure at least includes an inner cavity, a parting surface and a flow channel.

5. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The S3 data collection step includes the following collection methods: Blue light scanner acquisition: Continuously scan the mold surface. Each frame of the point cloud contains approximately several three-dimensional coordinate points. The robot arm moves to achieve seamless splicing for scanning and acquisition. Industrial CT acquisition: non-destructive scanning of the mold to be tested to generate three-dimensional tomographic images; Micro-optical probe acquisition: The deep cavity surface morphology is collected through laser confocal technology, and scanning and acquisition are performed in conjunction with the six-degree-of-freedom movement of the robotic arm.

6. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The S4 dynamic calibration includes the following steps: S41. Reference Coordinate System Construction: Using the laser tracker as a reference, a global coordinate system is established on the multi-sensor detection platform. The target ball on the motion platform is tracked by laser beams to obtain the sensor's position and status in real time. S42. Dynamic sensor calibration: The sensor is fixed to a multi-sensor detection platform and a standard calibration component is scanned in multiple postures. The standard calibration component is used for sensor calibration and coordinate system calibration. The transformation matrix from the sensor coordinate system to the global coordinate system is solved using the least squares method and iterative optimization: X 全局 =R·X 传感器 +T, where: R is the rotation matrix and T is the translation vector; S43. Deviation elimination and compensation: Use a laser tracker to correct sensor motion errors in real time, and introduce a temperature compensation model to correct thermal deformation.

7. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The S5 data fusion includes the following steps: S51. Coordinate alignment: Based on the multi-sensor dynamic calibration results, the multi-modal data of multiple angles of the mold to be tested are uniformly mapped to the global coordinate system; S52. Semantic Feature Enhancement: Using an improved PointNet++ network, we perform semantic segmentation on the multimodal data collected from multiple angles of the mold under test. We extract key component feature points and assign semantic labels to them. We then select semantically consistent feature points for registration. S52. Layered registration strategy: (1) Coarse registration: Initial alignment is performed based on the global geometric distribution of semantic feature points to reduce the positional differences of multimodal data at multiple angles of the mold to be tested; (2) Precise registration: Improve the ICP algorithm and use the defined semantic weight function to optimize the matching of corresponding points. Define the semantic weight function w i =α·s i +β·c i , where s i is the semantic confidence of the point, c i is the curvature value of the point; S53. Error Optimization: By iteratively optimizing the Huber loss function, a global stitching error of <0.05 mm is achieved. All registered multimodal data are fused to generate a 3D point cloud of the global model of the mold to be tested.

8. The mold global intelligent detection method based on multi-sensor dynamic calibration and semantic optimization learning according to claim 1, characterized in that: The S6 defect identification includes the following steps: S61 Multi-task Bayesian Model: The fused 3D point cloud data and CT image of the global model of the mold to be tested are input into the multi-task Bayesian model. The geometric features of the point cloud and the density features of the CT image of the mold to be tested are extracted through the 3D ResNet-50 network. S62 thermal deformation compensation: Based on the thermal expansion coefficient of the mold material, combined with real-time temperature data, dynamic correction of the test results obtained by the S61; S63. Result visualization and report generation: Compare the 3D point cloud data of the global model of the mold to be tested generated in the S5 data fusion step with the CAD model to generate a deviation spectrum.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images