Non-ferrous metal ingot detection method and equipment based on three-dimensional vision

Abstract

The invention discloses a non-ferrous metal ingot detection method and equipment based on three-dimensional vision. Belongs to the field of nonferrous metal detection. The method comprises the steps of construction of a three-dimensional pixel model with laser projection structured light projection and line laser scanning camera shooting, adaptive filtering preprocessing, establishment of a three-dimensional point cloud model through deep learning, feature information extraction, deep learning algorithm identification and classification, production process problem analysis and optimization and the like in sequence. Therefore, high-precision, multi-angle and all-directional detection of the non-ferrous metal ingot is realized. The equipment comprises a multi-axis collaborative three-dimensional visual inspection subsystem consisting of a mounting bracket, an X-Y-Z servo motion platform, a line laser scanning camera and the like. According to the method, the quality detection precision and defect detection rate of the non-ferrous metal ingot can be effectively improved, detailed product quality analysis is provided, and the method plays an important role in improving the product quality, optimizing the production process and reducing the production cost.

Description

Technical Field

[0001] The present invention belongs to the field of non-ferrous metal detection, and more specifically relates to a non-ferrous metal ingot detection method and device based on three-dimensional vision. Background Art

[0002] In the non-ferrous metal processing industry, the detection of product quality is particularly crucial. Especially for non-ferrous metal ingots produced by the continuous casting and rolling process, the monitoring and analysis of defects are essential for improving product quality and optimizing the production process. However, traditional manual detection methods are not only inefficient but also difficult to achieve high-precision and high-stability detection effects. This is mainly because when the human eye conducts product quality detection, it is easily affected by factors such as environmental light and visual fatigue, resulting in large subjectivity and instability in the detection results.

[0003] With the rapid development of computer technology, machine vision, and artificial intelligence technology, automated detection technology has gradually been applied to the quality detection of non-ferrous metal ingots. However, existing automated detection systems still have some problems. For example, under complex working conditions such as high temperature, strong oxide scale, and multi-scale texture interference, the defect detection rate of existing detection systems fluctuates greatly and the stability is poor. In addition, existing systems are difficult to effectively detect defects at the micron level, such as welds, cracks, etc.

[0004] To solve these problems, industry experts and researchers have begun to explore new detection methods, such as detection methods based on three-dimensional vision. This method is based on three-dimensional scanning technology and combines deep learning algorithms to perform high-precision detection of non-ferrous metal ingots from multiple angles and in all directions. In addition, by constructing a three-dimensional point cloud model, detailed feature information on the product surface can be further extracted for more in-depth and accurate product quality analysis. However, the research and practical application of the non-ferrous metal ingot detection method based on three-dimensional vision are still in the initial stage and need further optimization and improvement.

[0005] Industry data shows that the comprehensive qualified rate of zinc ingot production lines using traditional methods only reaches 92.4%, showing a significant gap from the 99% qualified rate required by intelligent manufacturing. Especially under complex working conditions such as high temperature, strong oxide scale, and multi-scale texture interference, the fluctuation range of the defect detection rate of existing systems exceeds ±20%, seriously restricting the large-scale production of high-end non-ferrous metal products. Therefore, it is necessary to construct a new detection system integrating three-dimensional shape perception, adaptive optical compensation, and lightweight deep learning to achieve online precise discrimination of micron-level defects. Summary of the Invention

[0006] The main technical problems to be solved by the present invention are how to improve the quality detection accuracy of non-ferrous metal ingots, enhance the defect detection rate by adopting a three-dimensional vision-based detection method, and maintain stable detection performance under complex working conditions such as high temperature, strong oxide scale, and multi-scale texture interference. At the same time, it is also necessary to solve how to effectively identify and judge micron-level defects, and accurately classify and record the high-precision detection results for further analysis of production process problems and optimization and improvement.

[0007] To achieve the above object, the present invention is implemented by the following technical solutions: The method includes:

[0008] Place the ingot on the X-Y-Z servo motion platform, and use a laser to project structured light onto the surface of the ingot;

[0009] Use a line laser scanning camera to take pictures of the ingot from multiple angles, convert the deformation information generated by the structured light on the surface of the ingot into pixel data, and construct a three-dimensional pixel model of the ingot;

[0010] Adopt an adaptive filtering algorithm to preprocess the three-dimensional pixel data; use an improved convolutional neural network algorithm to perform deep learning on the preprocessed three-dimensional pixel data to establish a three-dimensional point cloud model of the ingot;

[0011] Analyze each point in the three-dimensional point cloud model to extract the characteristic information on the surface of the ingot;

[0012] Use a deep learning algorithm to identify and classify the extracted features, and judge whether there are defects in the ingot and the type of defects;

[0013] Record the defect information in the process quality database, and find out the production process problems leading to the ingot defects by analyzing big data, and optimize them.

[0014] In one solution, the use of a laser to project structured light onto the surface of the ingot includes: automatically transporting the metal ingot to be detected to the working area of the X-Y-Z servo motion platform, and real-time monitoring the position of the ingot through a photoelectric sensor array. When the edge of the ingot triggers the light curtain signal of the infrared opposed sensor, the conveyor line immediately starts the hydraulic buffer device for precise stop to ensure that the deviation between the center of the ingot and the origin of the scanning coordinate system is less than ±0.1 mm.

[0015] In one solution, the construction of the three-dimensional pixel model of the ingot includes: using a line laser scanning camera to scan the ingot from different angles, converting the deformation information generated by the structured light on the surface of the ingot into pixel data, and constructing a three-dimensional pixel model of the ingot;

[0016] High-frequency reciprocating scanning, dynamically following the ingot conveying speed, integrating encoder pulse signals and laser rangefinder data to achieve multi-axis collaborative spatio-temporal synchronization control;

[0017] Analyze the phase information from the fringe deformation images captured during the scanning period to characterize the surface height information; subsequently, use the ICP algorithm for point cloud registration.

[0018] In one solution, establishing the three-dimensional point cloud model of the ingot includes: performing adaptive noise filtering on the original point cloud: aiming at the high specular reflection characteristics of the metal surface, designing an outlier removal algorithm based on local curvature statistics. For each point, calculate the eigenvalue of the covariance matrix within its radius neighborhood, define the anisotropy index and curvature, establish a dynamic threshold function, and determine and remove noise points when the neighborhood point density is lower than T

[0019] Subsequently, use the improved PointNet++ network for feature learning: divide the point cloud into local spherical domains with a radius of 2 mm, and extract hierarchical features through a multi-layer perceptron.

[0020] In one solution, extracting the feature information of the ingot surface includes:

[0021] Establish a multi-scale feature descriptor, determine the local coordinate system through the eigenvector of the covariance matrix, where is the neighborhood centroid; the color feature realizes non-linear enhancement through HSV space conversion, and define the chromaticity difference metric as, where is the gamma correction coefficient to compensate for the color deviation caused by specular reflection on the metal surface;

[0022] The texture feature uses an improved 3D local binary pattern algorithm: construct a spherical neighborhood with a radius centered on, project the neighborhood points onto the local tangent plane to generate a depth map, calculate the rotation-invariant LBP value, and construct a statistical histogram containing 32 azimuth bins;

[0023] The shape feature is quantified by differential geometric quantities, calculate the principal curvature and its direction: use the moving least squares method (MLS) to fit the local quadratic surface, then the Gaussian curvature, the mean curvature;

[0024] The attention mechanism is used for dynamic weighting in the feature fusion stage.

[0025] In one solution, judging whether there are defects in the ingot and the type of defects includes: running a defect classification network based on multi-granularity spatio-temporal attention. This network receives the voxelized feature field as input, and through the cooperation of global and local channels, to achieve comprehensive recognition of macroscopic morphology and microscopic details;

[0026] Then integrate the multi-granularity features through the feature distillation module, and use a dual-branch structure to judge the type of defects;

[0027] The auxiliary branch is responsible for constructing the prototype memory bank to enhance the classification performance;

[0028] In the inference stage, the system automatically adjusts the resolution according to the confidence level to achieve more refined feature extraction and accurate defect determination.

[0029] In one solution, recording the defect information into the process quality database includes: recording the defect information of the ingot into the blockchain-based process quality database to generate a digital twin containing three-dimensional point cloud fingerprints, defect maps, and process parameters;

[0030] The production process is optimized in real time through a double-loop feedback mechanism: the inner loop docks the defect classification results with the production control parameters in real time; the outer loop searches for the optimal process plan through federated learning and Bayesian optimization.

[0031] On the other hand, a non-ferrous metal ingot detection device based on three-dimensional vision, the device is applicable to the method, and the device includes: a mounting bracket, an X-Y-Z servo motion platform, a line laser scanning camera, and a turbine air-cooled fan;

[0032] After the ingot conveyor line brings the product into the scanning range, the multi-axis collaborative three-dimensional vision detection subsystem starts to work, and a high-rigidity linear module is used to build an X-Y-Z all-round servo motion platform. Among them, the X-axis uses an absolute encoder to achieve ±0.05mm dynamic following synchronization with the ingot conveyor line, the Y-axis is equipped with a line laser scanning camera to perform 4KHz high-frequency reciprocating scanning, and the sub-micron three-dimensional point cloud reconstruction is completed through the projection of fringe structured light.

[0033] Advantages of the present invention:

[0034] The present invention adopts a three-dimensional vision-based detection method, which can perform high-precision detection of non-ferrous metal ingots from multiple angles and in all directions, greatly improving the defect detection rate and detection accuracy. Especially for micron-level defects such as welds and cracks, effective detection can also be carried out.

[0035] By establishing a three-dimensional pixel model and a three-dimensional point cloud model of the ingot, the present invention can extract detailed feature information on the product surface, provide more in-depth and accurate product quality analysis, and provide strong support for optimizing the production process.

[0036] The present invention uses deep learning algorithms to identify and classify the extracted features, assisting engineers in judging whether there are defects in the ingot and the types of defects, improving the accuracy and speed of process decision-making.

[0037] By providing a non-ferrous metal ingot detection method and device based on three-dimensional vision, the present invention realizes efficient, high-precision, and comprehensive automated detection of non-ferrous metal ingots, and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flowchart of the method of the present invention;

[0039] Figure 2 is a flowchart of extracting feature information of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0040] For the convenience of understanding the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. The typical embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0041] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. For the convenience of understanding the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. The typical embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0042] As Figure 1 shown, the specific implementation steps of a non-ferrous metal ingot detection method based on three-dimensional vision are as follows:

[0043] Step 1: Place the ingot on the X-Y-Z servo motion platform, and turn on the line laser scanning camera to project structured light onto the surface of the ingot using the laser.

[0044] In the specific implementation process of projecting structured light, first, the ingot conveyor line automatically transports the metal ingot to be detected to the working area of the X-Y-Z servo motion platform. The position of the ingot is monitored in real time through the optoelectronic sensor array. When the edge of the ingot triggers the light curtain signal of the infrared opposed sensor, the conveyor line immediately activates the hydraulic buffer device for precise stop, ensuring that the deviation between the center of the ingot and the origin of the scanning coordinate system is less than ±0.1 mm. At this time, the servo drive system activates the three-axis linkage mechanism. The X-axis establishes real-time position feedback with the conveyor sprocket through an absolute encoder and uses the PID closed-loop control algorithm to dynamically compensate for the displacement error caused by the vibration of the conveyor belt. The line laser scanning camera mounted on the Y-axis guide rail then starts the preheating program. The built-in semiconductor laser completes wavelength stabilization (450 nm blue light band) within 3 ms, shapes the laser beam into a linear structured light with a thickness of 0.05 mm and a divergence angle of 0.5 mrad through the collimating lens group, and performs horizontal scanning at a frequency of 4 kHz by the galvanometer system. During the projection process, the control system synchronously adjusts the Z-axis lead screw lifting mechanism, automatically matches the best projection angle according to the preset ingot type database (usually set as an incident angle of 45° ± 5°), and at the same time, the turbine air-cooled module starts double-cycle heat dissipation to maintain the temperature of the laser at a constant temperature of 25°C ± 0.5°C to ensure the stability of the light intensity. When the structured light covers the entire surface of the ingot, the system triggers a nanosecond-level synchronization signal through the FPGA chip, enabling the high-speed CMOS image sensor to start capturing the fringe deformation pattern modulated by the surface topography of the ingot and transmitting it to the preprocessing unit for light intensity normalization processing to eliminate the brightness distortion caused by the difference in the reflectivity of the metal surface. The entire projection process is completed within 800 ms. During this period, the motion platform continuously fine-tunes its attitude to ensure that the coverage density of the structured light fringes in the edge area of the ingot is not less than 120 lines / mm, providing high-quality original optical data for subsequent 3D reconstruction.

[0045] Step 2: Use the line laser scanning camera to take pictures of the ingot from multiple angles, convert the deformation information generated by the structured light on the surface of the ingot into pixel data, and construct a three-dimensional pixel model of the ingot.

[0046] After the structured light projection is completed, the system enters the 3D data acquisition stage. The line laser scanning camera performs high-frequency reciprocating scanning at a speed of 0.8 m / s under the drive of the Y-axis servo motor. At the same time, the X-axis platform performs dynamic following motion according to the ingot conveying speed (typical value 0.3 m / s), and realizes the spatio-temporal synchronous control of multi-axis collaboration by fusing the encoder pulse signal (resolution 1 μm) and the laser rangefinder data (sampling rate 10 kHz) through the Kalman filter. When the galvanometer system scans at a frequency of 4 kHz, the CMOS sensor captures a fringe deformation image of 1024×768 pixels in each scanning cycle, and analyzes the phase information through the improved Gray code combined with the phase shift method: Let the light intensity distribution of the projected sine fringe be I p(x,y) = A(x,y) + B(x,y)cos[2πf0x + φ(x,y)]

[0047] After the surface modulation of the ingot, the expression of the deformation stripe is I d (x,y) = A'(x,y) + B'(x,y)cos[2πf0x + φ(x,y) + Δφ(x,y)], where the phase difference characterizes the surface height h(x,y) information (BC is the projection baseline, and AC is the imaging distance). For each spatial point (x,y), its three-dimensional coordinates (X,Y,Z) are solved through the binocular vision geometric model:

[0048]

[0049] In the formula, b is the baseline distance between the two cameras (calibration value 175mm), d = u l - u r is the parallax, f is the equivalent focal length, (u0,v0) is the principal point coordinates, and T is the translation amount of the coordinate system. The system acquires 8 groups of multi-view data at a step of 15° within a 120° viewing angle range, and realizes point cloud registration through the ICP (Iterative Closest Point) algorithm, with the registration error controlled within ±0.03mm. The finally constructed three-dimensional voxel model is stored using an octree data structure, with a spatial resolution of 50μm / voxel, and each voxel contains RGB reflectivity, normal vector and confidence weight where σ z is the standard deviation of height measurement. This process is accelerated by GPU computing and completes the fusion processing of 500,000 point cloud data within 1.2 seconds, providing a three-dimensional digital twin with sub-millimeter accuracy for subsequent defect analysis.

[0050] Step 3: Establish a point cloud model, and use an adaptive filtering algorithm to preprocess the three-dimensional voxel data to eliminate noise interference and improve the accuracy of the point cloud model. Use an improved convolutional neural network algorithm to perform deep learning on the preprocessed three-dimensional voxel data to establish a three-dimensional point cloud model of the ingot.

[0051] After the acquisition of three-dimensional voxel data is completed, the system starts the point cloud modeling process. First, perform adaptive noise filtering on the original point cloud : For the high specular reflection characteristics of the metal surface, design an outlier removal algorithm based on local curvature statistics. For each point p i , calculate its radius r = 2.5σ z in the neighborhood (σ zFor the covariance matrix eigenvalue λ1≥λ2≥λ3 of the standard deviation of height measurement in Step 2, define the anisotropy index α = (λ1 - λ3) / λ1 and the curvature κ = 2(λ2 + λ3) / (λ1 + λ2 + λ3), and establish the dynamic threshold function T = 0.6exp(-3α) + 0.15κ 0.5

[0052] , when the neighborhood point density is lower than T, it is determined as a noise point and removed. Subsequently, an improved PointNet++ network is used for feature learning: the point cloud is divided into local spherical domains with a radius of 2 mm, and hierarchical features are extracted through a multi-layer perceptron (MLP). The feature update formula for the l-th layer of the network is:

[0053]

[0054] where is the k-nearest neighbor (k = 32) of point i, W (l) is the learnable weight matrix, and σ is the LeakyReLU activation function. For the possible tiny defects on the ingot surface, a channel attention module is embedded in the network to define the feature channel weight to enhance the response of key features. For missing area repair, a conditional generative adversarial network (CGAN) is used for data completion. The loss function of the generator G includes a Chamfer distance term and a normal vector consistency constraint:

[0055]

[0056] where λ = 0.5 is the balance coefficient, is the point cloud normal vector. Finally, the point cloud model is optimized by a non-linear optimization algorithm to establish the energy equation where f(p i ) is the geometric constraint term, μ = 0.2 is the regularization coefficient, and the Levenberg-Marquardt algorithm is used for iterative solution until the residual decrease rate is less than 10 -5 . This process runs on an NVIDIA A100 GPU with CUDA 11.4, accelerates the calculation through an octree spatial index, outputs a topology-optimized point cloud model with adjustable density (default 0.1 mm resolution), and adopts a progressive encryption sampling strategy (Farthest Point Sampling) to automatically increase the sampling density to 0.05 mm in the curvature mutation area to ensure that the surface continuity and defect edge sharpness meet the requirements of ASTM E3-11 standard.

[0057] Step 4: Extract feature information, analyze each point in the three-dimensional point cloud model through the self-developed ingot defect feature extraction algorithm, and extract the feature information of the ingot surface, such as color, texture, shape, etc.

[0058] As shown Figure 2 in the figure, after constructing a high-precision point cloud model, the system executes a multi-modal feature extraction process.

[0059] S401. For the surface features of the ingot, first establish a multi-scale feature descriptor: construct a local reference frame (LRF) within the k-neighborhood (k = 64) of point p i , determine the local coordinate system through the eigenvectors of the covariance matrix , where is the centroid of the neighborhood. The color feature is non-linearly enhanced through HSV space conversion, and the chromaticity difference metric is defined as , where γ = 2.2 is the gamma correction coefficient to compensate for the color deviation caused by specular reflection on the metal surface.

[0060]

[0060] S402. The texture feature adopts an improved 3D local binary pattern (LBP) algorithm: construct a spherical neighborhood with a radius r = 0.5 mm centered at p i , project the neighborhood points onto the local tangent plane to generate a depth map

[0061] Calculate the rotation-invariant LBP value:

[0062]

[0063] , where ROR represents the cyclic shift operation and u is the uniform pattern constraint. At the same time, introduce the normal direction histogram (NDH) to describe the macroscopic morphology, map the point cloud normal vector to the spherical coordinate system, construct a statistical histogram containing 32 azimuth bins, and the weight of each bin , where σ n = 0.15 controls the attenuation of the normal vector similarity.

[0064] The shape feature is quantified by differential geometric quantities, calculate the principal curvatures k1, k2 and their directions: use the moving least squares (MLS) method to fit the local quadratic surface z = ax 2 + bxy + cy 2 + dx + ey, then the Gaussian curvature and the mean curvature To capture the features of micro defects, design a multi-scale shape index Calculate the weighted sum at three scales of 0.1 mm, 0.3 mm, and 0.5 mm The weight is determined by the noise level of each scale.

[0065] S403. In the feature fusion stage, use the attention mechanism for dynamic weighting, define the feature vector Learning Feature Importance Weights through a Gated Recurrent Unit (GRU):

[0066]

[0067] where $W$ g , $W$ h are learnable parameters, representing the Hadamard product. After the final feature set is reduced in dimension by t-SNE, a feature field is constructed in the Hilbert space where the bandwidth $l = 0.2$ mm and the coefficient $\alpha$ i is solved through optimization. This algorithm realizes real-time calculation on an FPGA acceleration card, can process the feature extraction of $2.8\times10^6$ points per second, and the correlation coefficient between the generated feature map and the metallographic inspection result reaches more than 0.97, meeting the ASTM E112-13 standard for grain size analysis.

[0068] Step 5: Defect Identification and Classification. Use a deep learning algorithm to identify and classify the extracted features, and judge whether there are defects in the ingot and the types of defects. This algorithm can intelligently learn and remember the defect features of various types, making defect identification and classification more accurate and fast.

[0069] After completing the multi-modal feature extraction, the system starts a defect classification network based on multi-granularity spatio-temporal attention (MGSA-Net). The network input is the feature field $\varPhi(x, y, z)$ generated in Step 4. The continuous space is discretized into cubic cells of $0.1$ mm 3 through voxelization. The feature vector within each voxel where is a spatial weighting kernel ($\sigma$ p = 0.03 mm). The network architecture adopts a dual-channel design: the global channel extracts macroscopic morphological features through three-dimensional sparse convolution, and the update formula for the convolution kernel parameters is:

[0070]

[0071] where is the set of valid voxels, suppressing the gradient of the smooth region; the local channel constructs a dynamic graph attention network (DGAT), constructs a spherical graph structure with a radius $R = 3$ mm centered on each defect candidate region, and the attention coefficient calculation between nodes $u$ and $v$ is:

[0072]

[0073] Among them, d = 256 is the feature dimension, and λ = 0.7 controls the spatial attenuation intensity. A multi-granularity feature distillation module is introduced in the middle layer of the network. Feature pyramids are extracted at three scales of [0.5mm, 2mm, 5mm] through deformable pooling, and a gated fusion mechanism is adopted:

[0074]

[0075] The classification head is designed as a two-branch structure: the main branch outputs the probability of the defect category through a fully connected layer The auxiliary branch constructs a prototype memory bank (k is the number of defect categories), and calculates the contrast loss to enhance the inter-class separability:

[0076]

[0077] Among them, τ = 0.1 is the temperature coefficient. The total loss function is the weighted sum L of the cross-entropy loss and the contrast loss total = L ce + 0.3L cont , and an online hard example mining strategy is introduced. Samples with classification confidence in the interval [0.4, 0.6] are imposed with a 3-fold loss weight. The curriculum learning strategy is adopted in the training stage. First, the features at the 5mm scale are mainly used for training for 100 epochs, and then the finer-scale features are gradually introduced. During inference, an adaptive resolution switching mechanism is deployed. When the regional classification confidence is lower than the threshold θ = 0.9, the module in step four is automatically called to re-extract local features at a resolution of 0.05mm to form a closed-loop optimization. This system realizes real-time classification of 126,000 voxels per second on NVIDIA A100, and the F1-score for seven types of defects such as pores, cracks, and slag inclusions reaches 98.7%. The defect prototype library is continuously updated through the federated learning framework to meet the online detection requirements of ISO 4967 standards.

[0078] Step Six: Process Optimization and Quality Tracking. Record the defect information in the process quality database. By analyzing big data, find out the production process problems that cause ingot defects and optimize them. At the same time, the recorded ingot information can be used for quality tracking, facilitating the quality traceability and control of products.

[0079] After completing the intelligent defect identification, the system constructs a process quality tracking system based on blockchain, and each ingot generates a unique digital twin containing three-dimensional point cloud fingerprints, defect maps, and process parameters. This twin is stored in a distributed database through a lightweight hash tree structure, where key process parameters (such as casting temperature, cooling rate, alloy composition) and defect characteristics form a time-series correlation chain, and an improved Merkle-Patricia tree is used to achieve high-speed writing of 3,000 records per second. The quality traceability engine is built with a causal inference module. When detecting subcutaneous porosity defects, it automatically correlates the historical curve of the electromagnetic stirring frequency of the melting furnace with the vibration spectrum data of the mold, and mines multi-order process deviations with a latency of more than 72 hours through a time-domain convolutional network.

[0080] The process optimization system adopts a double-loop feedback mechanism: the inner loop is docked with the defect classification results and production control parameters in real time. When the detection rate of microcracks exceeds the dynamic threshold, it triggers the PID controller to fine-tune the water distribution strategy in the secondary cooling zone of the continuous casting machine; the outer loop converges cross-production line data through federated learning, and uses the Bayesian optimization algorithm to search for Pareto optimal solutions in the process parameter hyperspace (constrained domain of temperature field ±15°C, casting speed ±0.2 m / min, electromagnetic stirring intensity ±10%), and generates a process improvement plan containing 23 control nodes. The quality dashboard integrates a three-dimensional thermal field simulation module, visually displays the predicted changes in the dendritic growth direction inside the ingot after process adjustment, and calculates the defect elimination probability using the Monte Carlo method.

[0081] For quality traceability requirements, a traceability system based on spatio-temporal semantic retrieval is developed, which supports the reverse tracking mode of "searching for processes by defects". Input any point cloud defect characteristics, and the system matches similar historical cases through a graph neural network and automatically generates a multi-dimensional traceability report containing metallurgical mechanism analysis. When there is a quality dispute, the zero-knowledge proof protocol can verify the process compliance of a specific ingot without disclosing business secrets. This system has achieved millisecond-level data synchronization with MES and ERP systems. In an aluminum alloy ingot production line with an annual output of 500,000 tons, the scrap rate caused by process defects has been reduced by 62%, and the quality traceability efficiency has been increased by 40 times, meeting the special quality requirements of the IATF 16949 automotive industry.

[0082] As Figure 2 shown, a non-ferrous metal ingot detection device based on three-dimensional vision includes a mounting bracket, an X-Y-Z servo motion platform, a line laser scanning camera, a turbine air-cooled fan, etc.

[0083] After the ingot conveyor line brings the product into the scanning range, the multi-axis collaborative 3D vision inspection subsystem starts to work. It builds an X-Y-Z all-round servo motion platform with a high-rigidity linear module. Among them, the X-axis uses an absolute encoder to achieve dynamic following synchronization with the ingot conveyor line with an accuracy of ±0.05 mm. The Y-axis is equipped with a line laser scanning camera to perform 4KHz high-frequency reciprocating scanning, and completes sub-micron-level 3D point cloud reconstruction through stripe structured light projection. The industrial-grade edge computing unit integrates an anti-thermal noise optimization algorithm to complete defect feature extraction and deep learning classification, and sends hierarchical control instructions to the PLC through the Profinet bus to trigger the pneumatic rejection mechanism to act. For defective products, a process quality database is constructed to optimize production process parameters. It realizes the closed-loop management of "detection-sorting-traceability-optimization".

[0084] Through the technological breakthroughs of 3D perception + intelligent control + extreme environment tolerance, this solution systematically solves the quality inspection problems in the non-ferrous metal smelting industry, realizes the transformation from "experience-driven" to "data-driven", and provides a full-stack solution for intelligent manufacturing upgrading.

[0085] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0086] It should be understood that the detailed description of the technical solutions of the present invention with the help of the preferred embodiments above is illustrative rather than restrictive. Those of ordinary skill in the art can modify the technical solutions recorded in each embodiment on the basis of reading the specification of the present invention, or perform equivalent replacement on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A detection method for non-ferrous metal ingots based on three-dimensional vision, characterized in that: The described method includes: Placing the ingot on an X-Y-Z servo motion platform and using a laser to project structured light onto the surface of the ingot; Using a line laser scanning camera to take pictures of the ingot from multiple angles, converting the deformation information generated by the structured light on the ingot surface into pixel data, and constructing a three-dimensional pixel model of the ingot; Preprocessing the three-dimensional pixel data using an adaptive filtering algorithm; using an improved convolutional neural network algorithm to perform deep learning on the preprocessed three-dimensional pixel data to establish a three-dimensional point cloud model of the ingot; Analyzing each point in the three-dimensional point cloud model to extract the feature information on the ingot surface; Using a deep learning algorithm to identify and classify the extracted features, and judging whether there are defects in the ingot and the type of defects; Recording the defect information into the process quality database, analyzing big data to find out the production process problems leading to ingot defects, and optimizing them.

2. The non-ferrous metal ingot detection method based on three-dimensional vision according to claim 1, wherein: The described using a laser to project structured light onto the surface of the ingot includes: Automatically transporting the metal ingot to be detected to the working area of the X-Y-Z servo motion platform, and real-time monitoring the position of the ingot through an optoelectronic sensor array. When the edge of the ingot triggers the light curtain signal of the infrared opposed sensor, the conveyor line immediately starts the hydraulic buffer device for precise stopping, ensuring that the deviation between the center of the ingot and the origin of the scanning coordinate system is less than ±0.1 mm.

3. The non-ferrous metal ingot detection method based on 3D vision according to claim 1, wherein: The described constructing a three-dimensional pixel model of the ingot includes: By using a line laser scanning camera to scan the ingot from different angles, converting the deformation information generated by the structured light on the ingot surface into pixel data, and constructing a three-dimensional pixel model of the ingot; High-frequency reciprocating scanning, dynamically following the ingot conveying speed, fusing the encoder pulse signal and the laser rangefinder data to achieve multi-axis collaborative spatio-temporal synchronous control; Analyzing the phase information from the fringe deformation images captured in the scanning cycle to characterize the surface height information; subsequently, using the ICP algorithm for point cloud registration.

4. A method for detecting non-ferrous metal ingots based on three-dimensional vision according to claim 1, characterized in that: The described establishing a three-dimensional point cloud model of the ingot includes: Performing adaptive noise filtering on the original point cloud: Aiming at the high specular reflection characteristics of the metal surface, designing an outlier removal algorithm based on local curvature statistics. For each point, calculate the eigenvalue of the covariance matrix within its radius neighborhood, define the anisotropy index and curvature, establish a dynamic threshold function, and determine it as a noise point and remove it when the neighborhood point density is lower than T Subsequently, using an improved PointNet++ network for feature learning: Dividing the point cloud into local spherical domains with a radius of 2 mm, and extracting hierarchical features through a multi-layer perceptron.

5. A non-ferrous metal ingot detection method based on three-dimensional vision according to claim 1, characterized in that: The described extracting the feature information on the ingot surface includes: Establishing a multi-scale feature descriptor, determining the local coordinate system through the eigenvector of the covariance matrix, where is the neighborhood centroid; the color feature is non-linearly enhanced through HSV space conversion, and the chromaticity difference metric is defined as, where is the gamma correction coefficient to compensate for the color deviation caused by the specular reflection on the metal surface; The texture feature uses an improved 3D local binary pattern algorithm: Constructing a spherical neighborhood with a radius centered on, projecting the neighborhood points onto the local tangent plane to generate a depth map, calculating the rotation-invariant LBP value, and constructing a statistical histogram containing 32 azimuth bins; The shape features are quantified by differential geometric quantities, and the principal curvatures and their directions are calculated: The local quadratic surface is fitted using the moving least squares (MLS) method, then the Gaussian curvature and the mean curvature; The attention mechanism is adopted for dynamic weighting in the feature fusion stage.

6. A method for detecting non-ferrous metal ingots based on three-dimensional vision according to claim 1, characterized in that: The judgment on whether there are defects in the ingot and the defect types includes: running a defect classification network based on multi-granularity spatio-temporal attention; receiving the voxelized feature field as input, and collaborating through both global and local channels; Then, the multi-granularity features are integrated through a feature distillation module, and a dual-branch structure is adopted to judge the defect types; Among them, the auxiliary branch is responsible for constructing the prototype memory bank to enhance the classification performance; In the inference stage, the system automatically adjusts the resolution according to the confidence level to achieve more refined feature extraction and accurate defect determination.

7. A non-ferrous metal ingot detection method based on three-dimensional vision according to claim 1, characterized in that: The recording of the defect information into the process quality database includes: recording the defect information of the ingot into the blockchain-based process quality database to generate a digital twin containing the three-dimensional point cloud fingerprint, the defect map, and the process parameters; The production process is optimized in real time through a double-loop feedback mechanism: The inner loop docks the defect classification results and the production control parameters in real time; the outer loop searches for the optimal process plan through federated learning and Bayesian optimization.

8. A non-ferrous metal ingot detection device based on three-dimensional vision, the device being applicable to the method according to any one of claims 1-7, characterized in that: The device includes: a mounting bracket, an X-Y-Z servo motion platform, a line laser scanning camera, and a turbo air-cooled fan; After the ingot conveyor line brings the product into the scanning range, the multi-axis collaborative three-dimensional vision detection subsystem starts to work. A high-rigidity linear module is used to build an X-Y-Z omnidirectional servo motion platform. Among them, the X-axis uses an absolute encoder to achieve dynamic following synchronization with the ingot conveyor line with an accuracy of ±0.05 mm. The Y-axis is equipped with a line laser scanning camera to perform 4KHz high-frequency reciprocating scanning, and the sub-micron three-dimensional point cloud reconstruction is completed through the projection of the fringe structured light.

Images