CN121214035A - A machine vision-based online detection method and system for ingot plating defects - Google Patents
A machine vision-based online detection method and system for ingot plating defectsInfo
- Publication number
- CN121214035A CN121214035A CN202511388791.4A CN202511388791A CN121214035A CN 121214035 A CN121214035 A CN 121214035A CN 202511388791 A CN202511388791 A CN 202511388791A CN 121214035 A CN121214035 A CN 121214035A
- Authority
- CN
- China
- Prior art keywords
- defect
- polarization
- features
- ingot
- machine vision
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
- Investigating Materials By The Use Of Optical Means Adapted For Particular Applications (AREA)
Abstract
本发明涉及图像处理技术领域，具体为一种基于机器视觉的摘锭镀层缺陷在线检测方法及系统，通过获取多视角图像，预处理后对图像数据流进行特征提取，并与预设缺陷字典进行匹配或聚类，初步识别潜在缺陷区域并触发精细化多角度图像采集，应用融合偏振感知卷积核的多任务CNN模型，联合优化像素级分割损失和边界预测损失，输出缺陷区域的像素级语义分割和精确边界，计算重构误差进行缺陷分类。最后，系统还具备缺陷根因分析与工艺流程调整功能，实现缺陷的追溯和生产优化。该系统整合了先进的视觉技术、深度学习和多模态信息融合，实现了摘锭镀层缺陷的在线、精准、自动化检测。
This invention relates to the field of image processing technology, specifically to an online detection method and system for plating defects on ingots based on machine vision. The system acquires multi-view images, preprocesses them, extracts features from the image data stream, and matches or clusters them with a pre-defined defect dictionary to initially identify potential defect regions and trigger refined multi-angle image acquisition. It then applies a multi-task CNN model with polarization-aware convolutional kernels to jointly optimize pixel-level segmentation loss and boundary prediction loss, outputting pixel-level semantic segmentation and precise boundaries of the defect region. The system calculates reconstruction error for defect classification. Finally, the system also features defect root cause analysis and process flow adjustment functions, enabling defect traceability and production optimization. This system integrates advanced vision technology, deep learning, and multimodal information fusion to achieve online, accurate, and automated detection of plating defects on ingots.
Description
Technical Field
The invention relates to the technical field of image processing, in particular to a method and a system for on-line detection of spindle plating defects based on machine vision.
Background
With the rapid development of industrial automation and intelligent manufacturing, the demand for online detection of product quality is increasing. Especially in complex manufacturing processes such as spindle plating, the traditional manual detection method is low in efficiency, is easily influenced by subjective factors and is difficult to ensure consistency.
Due to the inherent specular or mixed reflection characteristics of the coating surface, high light, glare or multiple reflections often exist in the captured image, which easily affect the accuracy of defect discrimination, and the true physical structure of the spindle coating is difficult to reflect. For example, a tiny crack may be masked or obscured by reflection interference, and even distortion of the shape may be a serious problem for the accuracy of machine identification of defects. The existence of complex layered structures in the spindle coating results in defects that may be located within the interior of the multilayer material. These layers may not only interfere with the detection process, but may even introduce false information, leading to false or missing alarms, which is a critical issue facing current machine vision detection.
Therefore, an on-line detection method and system for the defects of the spindle plating layer based on machine vision are provided.
Disclosure of Invention
The invention provides a method and a system for on-line detection of defects of a spindle coating based on machine vision, which realize automatic identification, classification and traceability of defects on the surface of the coating by combining machine vision, deep learning and multi-mode image processing technology, improve the quality of products and optimize the production process, and comprise the following steps:
Acquiring multi-view continuous images of the spindle plating layer under different illumination conditions and polarization states by using a camera array with a polarization filter;
The method comprises the steps of carrying out feature extraction on an image data stream, utilizing the extracted features to match and cluster with a preset dictionary, and identifying potential defect areas in a spindle plating layer;
Synchronously extracting conventional visual features and polarization optical features from multi-mode input data, designing a double-branch output structure, jointly optimizing segmentation loss and boundary prediction loss, outputting a segmentation result and a boundary positioning map of a defect region, and combining the reconstruction errors to obtain defect classification;
The root cause of the defect is traced back, and the process flow is adjusted in time.
Preferably, the step of preprocessing the multi-view continuous image specifically includes dynamically fusing and enhancing the multi-view continuous image, identifying local characteristics of the image, wherein the local characteristics include gradient amplitude, local contrast and texture complexity, and dynamically adjusting fusion weights of the multi-view continuous image according to the local characteristics to generate an image data stream.
The method comprises the steps of carrying out feature extraction on an image data stream, wherein the feature extraction comprises the steps of constructing a dictionary of features aiming at a preset defect mode, carrying out feature extraction on the image data stream, wherein the feature extraction comprises scale space extremum detection, key point positioning, dominant direction distribution and generation of a high-dimensional descriptor, carrying out matching and clustering on the extracted features and the preset dictionary, identifying potential defect areas in a spindle coating, reconstructing a surface normal by using a Lanbert reflection model, and obtaining depth information by an integration method to generate a feature vector to be detected.
Preferably, the step of performing defect detection by using a CNN model with fused polarization-aware convolution kernel specifically comprises the steps of adopting a CNN structure, embedding a polarization-sensitive processing layer in a convolution kernel of a CNN architecture and processing polarization information to obtain the CNN model with fused polarization-aware convolution kernel, recognizing abnormality by learning a normal mode and characteristics of a defect-free surface, calculating a reconstruction error after a feature vector to be detected is reconstructed when performing defect detection, and marking the reconstruction error higher than a preset threshold as an abnormal region;
And aiming at the abnormal region, performing defect classification by using a lightweight convolutional neural network, wherein the defect classification comprises uneven plating thickness, missing, surface scratch, pit, flaking, foaming, abnormal chromatic aberration and oxidized spots, and training a binary neural network for each defect classification independently to work together with a CNN model fused with a polarization perception convolutional kernel.
Preferably, the defect detection specifically adopts a multi-task deep learning architecture, and comprises a shared feature extraction backbone network configured to extract common feature representations from the feature vectors to be detected, and a plurality of task heads for bounding box positioning and pixel-level semantic segmentation, wherein the task heads are used for executing respective identification tasks based on the shared features, synchronously generating bounding box geometric coordinates and geometric outlines of the defect areas in a single calculation process, determining accurate geometric shapes, sizes and spatial distributions of defects by analyzing pixel-level semantic segmentation results, and identifying defect classifications according to the accurate geometric shapes, sizes and spatial distributions.
Preferably, the multi-mode input data comprises an intensity image, a polarization degree image, a polarization angle image, a surface depth image, a surface normal image and penetrability imaging data, wherein the penetrability imaging data comprises near infrared, terahertz and ultrasonic data, and the penetrability imaging data is integrated through a customized input layer and a characteristic fusion module.
Preferably, the step of fusing the segmentation result and the boundary locating map of the defect area comprises the steps of enabling a CNN model of a polarization perception convolution kernel to be provided with an independent feature encoder, synchronously extracting conventional visual features and polarization optical features from multi-mode input data, including two paths of output branches, wherein the main branches generate a pixel-level semantic segmentation probability map of the defect area, assist in branch prediction of probability distribution of defect boundaries, capturing fine-grained outline features of defects through joint optimization of segmentation loss and boundary prediction loss, and synchronously outputting the defect segmentation result and the boundary locating map.
An on-line detection system for defects of spindle plating based on machine vision, comprising:
the data acquisition and preprocessing module is used for acquiring multi-view continuous images of the spindle plating layer under different illumination conditions and polarization states by using a camera array with a polarization filter;
the potential defect area identification module is used for carrying out feature extraction on the image data stream, matching and clustering the extracted features with a preset dictionary and identifying potential defect areas in the spindle plating layer;
the refined feature vector generation module is used for triggering refined multi-angle image acquisition aiming at the potential defect area, reconstructing surface normal and depth information of the spindle plating layer by utilizing a self-calibration luminosity three-dimensional algorithm, and obtaining a feature vector to be detected;
The defect detection and analysis module is used for performing defect detection by applying a CNN model fused with a polarization perception convolution kernel, inputting a feature vector to be detected into the CNN model for reconstruction, and calculating a reconstruction error;
and a defect root cause analysis and optimization module for tracing the root cause of the defect and timely adjusting the process flow.
Compared with the prior art, the invention has the beneficial effects that:
1. The camera array provided with the polarizing filter is used for acquiring multi-view and multi-polarization state images, the polarization information can effectively distinguish surface defects from normal areas, and the camera array has higher sensitivity to fine defects and defects of different types, such as scratches, pits and chromatic aberration. The deep learning anti-reflection network can effectively separate reflection interference, ensure the purity of input image data and further improve the accuracy of subsequent feature extraction and defect identification.
2. The CNN architecture integrating the polarization perception convolution kernel and the application of the self-calibration luminosity three-dimensional algorithm enable the system to acquire defect characteristics from multiple dimensions such as geometric and optical characteristics, so that accurate identification and classification of various complex defects are realized, and high detection precision can be maintained even in an online high-speed detection scene.
3. The method can synchronously generate the geometric coordinates of the boundary frame and the geometric outlines of the pixel level of the defect area, marks the credible probability on each pixel point, judges whether the detected defect area belongs to or not in a combined way, and can more accurately determine the position, the shape, the size and the spatial distribution of the defect through a multi-task deep learning architecture and a multi-mode input data segmentation network. This refined defect localization capability provides reliable data support for subsequent process optimization and quality control.
Drawings
FIG. 1 is a flow chart showing the steps of a method for on-line detecting defects of a spindle plating layer based on machine vision;
FIG. 2 is a schematic diagram of a system for on-line detecting defects of a spindle plating layer based on machine vision according to the present invention;
fig. 3 is a flowchart of a multi-task deep learning architecture according to a first embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1 to 3, the invention provides a method and a system for on-line detecting defects of a spindle plating layer based on machine vision, and the technical scheme is as follows, referring to fig. 1, a flow chart of steps of the method for on-line detecting defects of the spindle plating layer based on machine vision provided by the invention is provided, which comprises the following steps:
Acquiring multi-view continuous images of a spindle plating layer under different illumination conditions and polarization states by using a camera array with a polarization filter;
Extracting features of the image data stream, including scale space extremum detection, key point positioning, dominant direction distribution and generation of a high-dimensional descriptor; triggering fine multi-angle image acquisition aiming at the potential defect area, and reconstructing surface normal and depth information of the spindle plating layer by using a self-calibration luminosity three-dimensional algorithm to obtain a feature vector to be detected;
And embedding a polarization feature extraction special layer in a network to execute defect detection by applying a CNN architecture fused with a polarization perception convolution kernel, wherein the defect detection is to input a feature vector to be detected into a pre-training model for reconstruction, calculate errors between an original vector and a reconstructed vector, classify defects according to the errors, trace back root causes of defect generation and provide support for process optimization.
Referring to fig. 2, a schematic structure diagram of a machine vision-based on-line detection system for defects of a spindle plating layer provided by the invention comprises a data acquisition and preprocessing module, a potential defect area recognition module, a refined feature vector generation module, a defect detection and analysis module and a defect root cause analysis and optimization module.
Example 1
The embodiment aims at elaborating a machine vision online detection system combining polarization imaging, deep learning and multi-task learning, and realizing high-precision, real-time identification, classification and tracing of defects of a spindle plating layer.
An array of at least 4-6 industrial-level high-resolution cameras is first deployed. The cameras surround the spindle coating conveyor above in an annular or fan-shaped arrangement, ensuring that the coating surface is captured by one camera from different angles, for example, every 60 degrees or 45 degrees. And a switchable polarizing filter, such as a linear polarizing plate, is arranged in front of each camera lens, can be rapidly rotated to four angles of 0 DEG, 45 DEG, 90 DEG and 135 DEG, and can be rapidly switched through a precise mechanical device or a liquid crystal adjustable polarizing filter.
The light array formed by orderly arranging the LED light sources in a programmable control mode is used for providing at least three different illumination conditions, such as uniform diffuse light, specific angle directional light and high contrast oblique light, and is used for simulating various illumination scenes possibly encountered in actual production, the illumination scenes have specific influence on the acquisition result of a plating layer, the brightness of the LED light sources is adjustable, and the brightness adjustment can be completed in response to an instruction after a brightness adjustment instruction is sent out.
Further, all cameras and light sources are accurately synchronized by an industrial controller, and when the spindle plating passes through a detection area, the photoelectric sensor triggers image acquisition, so that clear continuous images without smear can be obtained under the condition that the spindle moves at a high speed. The image acquisition frequency is set at 100-200 frames/second, forming a continuous set of processable images.
Each spindle was assumed to be 50cm in length and the detection zone width 10cm. And 6 cameras are deployed in total for a single camera at 100 frames/second, and images corresponding to polarization state switching and illumination switching are acquired every second.
A dataset is constructed containing the reflected interference image and its corresponding non-reflected real scene image. The data are collected through an actual production line, in particular by comparing the reflection with the no reflection, and are realized through special shading or polarization setting or are obtained by simulating different reflection conditions by using rendering software. During training, model learning converts an input image with reflection into a "clean" non-reflected image. In consideration of the real-time requirement of online detection, the network model needs pruning and quantization, reduces the calculation loss, and is deployed on a high-performance GPU server, so that the antireflection processing time of each image is ensured to be within 10-20 milliseconds.
The preprocessing step comprises the steps of carrying out dynamic fusion and enhancement, carrying out intensity normalization on the multi-view image, carrying out weighted fusion according to the characteristics of local gradient, contrast, texture complexity and the like, and finally carrying out self-adaptive histogram equalization to optimize the image quality.
Further, the local gradient amplitude, the local contrast and the texture complexity of each frame of image are accurately calculated and quantized, and the fusion weights of different illumination images are dynamically adjusted to form a high-quality image. And then based on the local characteristics, dynamically adjusting the fusion weight of each view angle image to obtain a composite image with higher quality, and further generating an optimized image data stream which can be used as the input of a model.
Through image acquisition of multiple visual angles, multiple polarizations and multiple illuminations, the information of the spindle plating layer is comprehensively captured, and the problem of missing detection in the traditional method is effectively avoided. Meanwhile, the deep learning antireflection network can intelligently eliminate the interference of the high reflection surface, and ensure the definition of the image. By combining the dynamic fusion and enhancement technology, the system can effectively highlight defect characteristics and inhibit background noise, and provides high-quality data for subsequent accurate identification.
And the novel defects with large differences from the preset defect dictionary and the pixel-level segmentation boundary fuzzy regions are intelligently identified and preferentially screened through samples with low model prediction confidence, and are submitted to an expert feedback loop. And accurately labeling the high-value samples and confirming the defect types on a labeling interface, and synchronously updating the defect dictionary. These new annotation data are then used to incrementally train the deep learning model, with incremental learning to avoid "catastrophic forgetfulness". The self-adaption, continuous iteration and optimization of the model performance are realized, and the new appearance, unknown or edge defect type can be effectively treated.
Further, image features are extracted by scale space extremum detection, keypoint localization, dominant direction assignment, and generation of high-dimensional descriptors. These features are then used to match and cluster with a pre-set defect dictionary to identify potential defect areas.
The system compares the high-dimensional descriptors of each key point extracted from the current image with a pre-established defect feature dictionary which stores typical feature descriptors of various known defect types, the descriptors are given in text form, and the similarity between the descriptors is calculated to find out the region highly matched with the defect features depicted in the dictionary.
For those keypoints that successfully match the defect feature, the system performs a cluster analysis by which the overall profile and extent of the defect can be identified from the discrete keypoint information. The system classifies key points which are close to each other in space position and have similar characteristics into the same group, so that a complete and continuous potential defect area is outlined, namely the position of the defect is roughly determined, then the current defect area is triggered to acquire a refined multi-angle image, and then the CNN model is utilized to perform refined identification.
An image is acquired using at least three light sources. The camera response is calibrated by photographing a known shape reference, optimally solving for the light source direction and brightness. And reconstructing a surface normal line by using a lambert reflection model according to different illumination intensities and light source parameters, and obtaining depth information by an integration method to generate a feature vector to be detected.
The method aims at generating a rough defect area and enriching multi-angle image data in the area, and after determining a rough range, the outline boundary is finely identified, so that the optimal input is provided for the subsequent feature extraction, the recall rate of the subsequent defect identification can be improved through dynamic fusion, and the detection effect of fine scratches and pits is particularly improved obviously.
Further, referring to fig. 3, a flowchart of a multi-task deep learning architecture according to a first embodiment of the present invention is shown. At the early layers of CNN, custom layers dedicated to processing and fusing polarization features are embedded, responsible for converting raw polarization measurements into more physically meaningful features, such as polarization degree maps and polarization angle maps, and fusing them with intensity image features. Specifically, a polarization fusion convolution block is designed, and feature maps from different modes are received and weighted and fused.
High quality, defect free spindle plating images were collected as training data sets. These data are used for training using self-encoder or stream-based generative models. The training goal is to allow the model to reconstruct exactly the non-defective input image or feature vector. The model learns the intrinsic representation, texture law and polarization response characteristics of the normal plating surface by optimizing the reconstruction loss.
Further, in the detection stage, after the spindle plating image to be detected is subjected to data acquisition and preprocessing, the feature vector to be detected is input into a pre-trained model. The system calculates the pixel level difference, i.e. reconstruction error, between the original input vector and the model reconstruction vector. When the reconstruction error of a certain pixel or region is higher than a preset threshold value, the pixel or region is marked as an abnormal region. The threshold is set to the average of the normal reconstruction errors plus 3-5 standard deviations. For example, by testing on 1000 non-defective samples, the average value of normal reconstruction errors was found to be 0.02 and the standard deviation was found to be 0.005. Then the threshold may be set to 0.035. Any pixel has a reconstruction error exceeding 0.035, i.e. is determined to be potentially anomalous.
Further, for the marked abnormal region, a plurality of lightweight convolutional neural networks are utilized to carry out refined defect classification. A strategy is employed to train one binary neural network separately for each preset defect type. The defect types include uneven plating thickness, missing, surface scratches, pits, flaking, blistering, abnormal color difference, and oxidized spots. For each defect type, thousands of images containing that type of defect are collected and input into a corresponding binary classification network for training. The training uses a combined loss function including bounding box loss, objective loss, and segmentation loss;
Further, the combined loss function further comprises a unified IoU loss function, the unified IoU loss function optimizes positioning by dynamically adjusting a boundary box scaling mechanism and a bidirectional weight distribution strategy, and the overlapping degree and the shape similarity of the prediction box and the real box are measured. This enables the system to output extremely accurate defect geometry information, including shape, size, and spatial distribution, providing more scientific and reliable data support for subsequent process optimization and quality control.
The modular design can avoid the problem of class imbalance, each classifier can focus on identifying the micro-features of a specific type, and when a new defect type appears, only one new binary classifier is required to be trained, and the whole large-scale network is not required to be retrained.
The core of the multi-task deep learning architecture is a shared feature extraction backbone network. The network is configured to extract common and semantically rich feature representations from the feature vectors to be detected. The backbone network is selected, the light weight and high efficiency characteristics of the backbone network are considered, the backbone network can be pre-trained on a massive image data set, general visual characteristics are obtained, and then fine adjustment is carried out by utilizing the defect data of the spindle plating layer, so that the backbone network is suitable for the characteristics of specific fields.
Further, over the output of the shared backbone network, a plurality of task-specific headers for bounding box localization and pixel-level semantic segmentation are connected. These headers perform the respective recognition tasks independently but cooperatively based on the shared representation of the features. The bounding box localization header is responsible for outputting the bounding box geometry of each detected defect, which predicts the location and size of the bounding box by regression loss, while judging whether the region does contain a defect by classification loss. The pixel-level semantic segmentation head outputs the pixel-level geometric outline of the defective region, i.e., a binary mask of the same resolution as the input image, and is trained by pixel-level classification loss.
A key advantage of this multitasking architecture is that it enables simultaneous generation of bounding box geometric coordinates of the detected defect and pixel-level geometric contours of the defect region in a single unified calculation process. Performing one forward propagation gives a rough location and precise shape of the defect. The position and the appearance of the defect are described, and the geometric shape, the size and the spatial distribution of the defect can be accurately determined through analysis of pixel-level semantic segmentation results, so that the defect type can be identified.
Further, the multi-modal input data includes intensity images which are conventional gray scale images, brightness information reflecting surfaces, polarization degree images reflecting surface roughness, material scattering characteristics, polarization angle images reflecting surface normal directions or texture anisotropism, three-dimensional height information reconstructed by photometric stereo algorithm for surface depth maps, height fluctuation directly reflecting surfaces, surface normal images reflecting surface normal directions of each pixel point, extremely strong recognition capability for microscopic geometric features, and penetrating imaging data such as near infrared, terahertz or ultrasonic data, which are used for detecting defects inside or on subsurface of a plating layer.
The customized input layer respectively performs characteristic extraction on different mode data such as intensity, polarization, depth, normal line and the like, and designs special branches on penetrating imaging data such as near infrared and terahertz. The feature fusion module adopts a multi-level attention fusion mechanism, and dynamically learns and fuses the high-level semantic features of each mode through splicing and cross-mode attention.
For each mode of data, independent input branches such as intensity images, polarization degree images and polarization angle images are designed and can be directly input as independent channels of CNNs, and depth images, normal images and penetrability imaging data are subjected to preliminary feature extraction through respective convolution layers.
The heterogeneous multi-mode data are integrated through the customized input layer and the characteristic fusion module, so that information of different modes can be effectively utilized and mutually supplemented. The feature fusion module enables the network to self-adaptively learn the dependency relationship among different modal features and dynamically allocate weights.
Features are learned and extracted from multimodal input data by separate feature encoders. Specifically, the decoder divides two independent heads, namely, a first head is responsible for outputting a probability map of a defect area, which is a gray scale map with the same size as an input image, the value of each pixel represents the probability that the pixel belongs to the defect area, the probability is 1, the pixel belongs to the defect area, and a second head is specially used for outputting a probability map of a defect boundary, the probability is larger than 0.8, the boundary of the defect area is highlighted, namely, the boundary line of the defect and a normal area is highlighted according to the probability, and the boundary line of the defect area is highlighted.
Further, to ensure that the network can accomplish both tasks at the same time with high quality, the model is trained by jointly optimizing the pixel level separation loss and the boundary prediction loss. Pixel level classification loss is used to measure the difference between the predicted segmentation mask and the real mask, ensuring proper classification at the pixel level. The boundary prediction loss is used for measuring the similarity between the prediction boundary and the real boundary and encouraging the prediction boundary to be close to the real boundary. The total loss function is a weighted sum of the two loss terms.
With this joint optimization, the model is encouraged during training to capture and output the subdivision profile of the defect while distinguishing between the defect and normal areas, thereby providing extremely accurate geometric information.
The integrated system is able to correlate the detailed information for each defect, including type, location, size, time of occurrence, with real-time process parameters of the production line, raw material lot, etc. When defects are found, potential problems in a production link can be rapidly located, an active learning mechanism is introduced, and images with uncertain models or easy errors are intelligently selected for manual labeling, so that rapid iteration and improvement of model performance are realized with less manpower investment.
Integrating a complete system, automating data transmission and processing flow among all modules of the system, judging surface scratches when detecting slender linear abnormality, abrupt change of normal direction and length-width ratio of >5:1, outputting scratches A,1.2mm multiplied by 0.05mm, coordinates (X1, Y1) - (X2, Y2), judging pits when circular low-contrast area, normal radial change, area <0.5mm 2, output pits B, diameter 0.2 mm/depth 0.03mm, coordinates (X3, Y3), judging plating non-uniformity when detecting abnormal polarization degree/thickness and irregular plaque, outputting areas 5mm 2, thickness fluctuation 0.01mm and coordinates (X4, Y4), and other defect examples are peeling D (2 mm 2, irregular, X5, Y5), and bubbles E (0.1 mm, circular, X6, Y6)
The integrated system uploads the defect data to the central data platform in real time through the MQTT, and the defect data are aligned and associated with the data such as real-time process parameters, raw material batches and the like through time stamps. The system builds a root cause analysis model driven by data, deduces the causal relation between the process parameter change and the defects by using the Grangel causal test, and automatically generates a process optimization suggestion.
The highly-automated process shortens the detection and analysis process which may need minutes or even hours to the second level or even the sub-second level, greatly improves the overall efficiency and response speed of the production line, eliminates the delay of manual intervention and data conversion, realizes the full-process high-speed operation from image acquisition to defect identification, classification and reporting, and ensures that the spindle plating layer can be monitored in real time and accurately in high-speed production.
The embodiment realizes the omnibearing automatic detection of the spindle plating defects, from data acquisition to preprocessing, potential defect identification to refined feature extraction and finally intelligent defect detection, classification and root cause analysis, and provides a comprehensive and prospective solution. The solution of the embodiment improves the quality control level of the spindle plating product, reduces the labor cost, provides powerful technical support for the construction of intelligent factories, and finally promotes the continuous optimization of the production process and the promotion of the industrial competitiveness.
Example two
The challenges of the embodiment lie in that the surface of the spindle plating layer has geometric characteristics of non-plane and complex curved surfaces, and the defect types are more refined, including micro or subsurface defects such as micro cracks, internal oxidation, abnormal crystals, multi-layer peeling and the like, which are difficult to identify through a single mode, and the specific steps of the embodiment are as follows:
The system is used for deploying a high-resolution, 1200 ten thousand-pixel and 500-fps global shutter industrial camera, and capturing images of intensity, polarization degree and polarization angle by matching with a polarization filter. The integrated high-precision laser triangulation sensor has Z-axis repetition precision reaching +/-2 microns and can identify micro pits with depth greater than 10 microns. Part of the cameras are equipped with multispectral imaging units for detecting internal oxidation or subsurface delamination defects. The system adopts an active light source array with programmable, multi-angle, multiband and variable polarization directions, and can dynamically adjust illumination within 10 milliseconds to ensure optimal imaging.
Further, after preprocessing, accurate registration is carried out on the original data of each mode under a unified three-dimensional coordinate system, and the cross-mode registration error is controlled within 2 pixels. The core adopts a three-dimensional convolutional neural network to perform feature extraction and fusion, and an inter-integrated cross-modal attention mechanism improves the signal to noise ratio after processing by up to 25%. The system introduces time sequence analysis capability, and multiple view images shot at different time points are connected into a group of images to capture dynamic defects of the coating, so that the root cause of the defects can be found more clearly. The advanced deep learning de-artifacting algorithm can reduce the false detection rate caused by reflection artifacts to below 0.1%.
Based on the fused three-dimensional multi-modal characteristics, the system realizes accurate defect segmentation by using an advanced three-dimensional semantic segmentation network. The recall rate of the integral defect detection reaches more than 98.5%, and the micrometer-level crack detection precision can reach 0.05 mm. The system can identify and classify various complex defects, and the classification accuracy of the key defect types reaches more than 97%. The identified defects are subjected to high-precision three-dimensional quantification, and for example, the depth measurement error of the micro pits can be controlled to be within +/-5 percent. And identifying and adopting a multitask learning paradigm, and synchronously completing three-dimensional boundary box positioning, pixel/voxel level segmentation and fine classification. The system deeply correlates defect data with real-time process parameters, successfully correlates more than 80% of typical defects with at least one key process parameter, and deduces or determines the cause of the defects through historical data.
When the defects are classified and positioned, the system correlates the data, including type, position, size, occurrence time stamp, etc. with real-time production parameters of the production line, such as plating solution temperature, current density, deposition time, raw material batch, equipment running state, etc., correlates the defects with the process flow, and accurately controls the problem source.
Further, the time series data is analyzed, and a large amount of defect data and associated production parameters thereof are summarized. For example, if a pit of a specific type is found to occur in a large amount within a certain period of time, and the associated data shows that the bath temperature is abnormally increased for that period of time, it can be preliminarily inferred that the temperature abnormality is the root cause of the pit generation.
Detailed defect reports and trend charts are generated, and changes of defect types, occurrence frequencies, severity degree along with time and production parameters are intuitively displayed. And the analysis result is fed back to production line operators and engineers in real time, so that specific process optimization suggestions are provided. For example, the system will show that the coating surface scratch rate increases significantly when the bath temperature exceeds a set threshold, and will also give a clear maintenance cue that "recommended inspection of the temperature control system". "
The embodiment expands the application of the three-dimensional convolutional neural network, forms the rapid positioning for tiny and difficultly perceived defects, frames the area where the defects are located, simultaneously adds the closed loop feedback mechanism which is the key for realizing intelligent manufacturing and continuous quality improvement, realizes rapid detection and decision making by the system, ensures that the automatic rejection rate of the serious defects reaches 99%, improves the recognition capability by 5-10% by continuously increasing learning, improves the qualification rate of products by 2-3 percentage points, reduces the rejection rate by online defect recognition and process flow optimization measures, and fully verifies the efficiency improvement of the system in the aspects of improving quality control and optimizing production.
Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (8)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511388791.4A CN121214035A (en) | 2025-09-26 | 2025-09-26 | A machine vision-based online detection method and system for ingot plating defects |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511388791.4A CN121214035A (en) | 2025-09-26 | 2025-09-26 | A machine vision-based online detection method and system for ingot plating defects |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN121214035A true CN121214035A (en) | 2025-12-26 |
Family
ID=98120562
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202511388791.4A Pending CN121214035A (en) | 2025-09-26 | 2025-09-26 | A machine vision-based online detection method and system for ingot plating defects |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN121214035A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121661426A (en) * | 2026-02-05 | 2026-03-13 | 深圳圣融达科技有限公司 | An AI-based visual inspection method and system for detecting appearance defects in intelligent capacitor manufacturing. |
| CN121684965A (en) * | 2026-02-09 | 2026-03-17 | 福建众辰精密机芯有限公司 | Traceability Management System and Method for Watch Movement Components |
-
2025
- 2025-09-26 CN CN202511388791.4A patent/CN121214035A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN121661426A (en) * | 2026-02-05 | 2026-03-13 | 深圳圣融达科技有限公司 | An AI-based visual inspection method and system for detecting appearance defects in intelligent capacitor manufacturing. |
| CN121684965A (en) * | 2026-02-09 | 2026-03-17 | 福建众辰精密机芯有限公司 | Traceability Management System and Method for Watch Movement Components |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA3062044C (en) | System and method for work piece inspection | |
| CN121214035A (en) | A machine vision-based online detection method and system for ingot plating defects | |
| CN111612737B (en) | Artificial board surface flaw detection device and detection method | |
| JP2021515885A (en) | Methods, devices, systems and programs for setting lighting conditions and storage media | |
| CN119831966A (en) | PCB production line defect detection system and method based on image processing | |
| CN120833319B (en) | Mobile phone silica gel shell appearance detection system based on optical image sensor | |
| CN109840900A (en) | A kind of line detection system for failure and detection method applied to intelligence manufacture workshop | |
| CN120235842A (en) | Intelligent detection method, device and equipment for plastic containers | |
| CN119904423B (en) | Method for detecting surface of automobile part based on large model | |
| Banus et al. | A deep-learning based solution to automatically control closure and seal of pizza packages | |
| CN120496055B (en) | A Vision-Based Method and System for Quality Inspection of Paper Tray Production | |
| CN120655596A (en) | Earphone appearance shell appearance defect detection method based on visual detection | |
| CN120823155A (en) | A method and system for intelligent detection of shell manufacturing defects | |
| CN120450205A (en) | Rapid batch online detection method and system for micro bearings | |
| CN120339701B (en) | A surface mount inspection system for semiconductor packaging | |
| CN121415147A (en) | Online Defect Detection System for Hardware Tools Based on AI Image Recognition | |
| CN116843615B (en) | An intelligent full inspection method for lead frames based on flexible optical paths | |
| CN117969553A (en) | An online visual inspection system for the appearance of TPV knitted composite tubes | |
| CN119574572B (en) | Intelligent test method and equipment for mobile phone accessory production | |
| CN120195178B (en) | An online detection system for diamond polishing and grinding | |
| Zhang | Dual-channel capsule shape detection system based on computer vision and optical imaging technologies | |
| CN118980694B (en) | A method for online detection of turbine volute machining defects | |
| CN121231495A (en) | Image Recognition-Based Defect Detection Method and System for Die-Cured Products | |
| CN120877011A (en) | Future factory high-precision die detection method and system based on vision | |
| CN120847112A (en) | An online detection system and method for surface defects of LED display screens |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |