基于深度学习的织物缺陷在线检测算法

织物缺陷在线检测是纺织行业面临的重大难题，针对当前织物缺陷检测中存在的误检率高、漏检率高、实时性不强等问题，提出了一种基于深度学习的织物缺陷在线检测算法。首先基于GoogLeNet网络架构，并参考其他分类模型的经典算法，搭建出适用于实际生产环境的织物缺陷分类模型；其次利用质检人员标注的不同种类织物图片组建织物缺陷数据库，并用该数据库对织物缺陷分类模型进行训练；最后对高清相机在织物验布机上采集的图片进行分割，并将分割后的小图以批量的方式传入训练好的分类模型，实现对每张小图的分类，以此来检测缺陷并确定其位置。对该模型在织物缺陷数据库上进行了验证。实验结果表明：织物缺陷分类模型平均每张小图的测试时间为0.37 ms，平均测试时间比GoogLeNet减少了67%，比ResNet-50减少了93%；同时模型在测试集上的正确率达到99.99%。说明其准确率与实时性均满足实际工业需求。     

Designing a two-rank acceptance sampling plan for quality inspection of geospatial data products

To address the disadvantages of classical sampling plans designed for traditional industrial products, we originally propose a two-rank acceptance sampling plan (TRASP) for the inspection of geospatial data outputs based on the acceptance quality level (AQL). The first rank sampling plan is to inspect the lot consisting of map sheets, and the second is to inspect the lot consisting of features in an individual map sheet. The TRASP design is formulated as an optimization problem with respect to sample size and acceptance number, which covers two lot size cases. The first case is for a small lot size with nonconformities being modeled by a hypergeometric distribution function, and the second is for a larger lot size with nonconformities being modeled by a Poisson distribution function. The proposed TRASP is illustrated through two empirical case studies. Our analysis demonstrates that: (1) the proposed TRASP provides a general approach for quality inspection of geospatial data outputs consisting of non-uniform items and (2) the proposed acceptance sampling plan based on TRASP performs better than other classical sampling plans. It overcomes the drawbacks of percent sampling, i.e., “strictness for large lot size, toleration for small lot size,” and those of a national standard used specifically for industrial outputs, i.e., “lots with different sizes corresponding to the same sampling plan.”     

改进注意力机制的型钢表面微小缺陷检测方法

针对型钢表面缺陷种类多样、微小缺陷占比较大导致的检测效率低、检测精度差的问题,提出了一种基于双重多尺度注意力机制的表面缺陷检测方法DMSA-YOLOv3,实现型钢表面多尺度缺陷快速精确检测。构建了基于通道和空间的双重多尺度注意力模型DMSA,对不同尺度特征进行筛选融合,强化小尺度缺陷的特征权重;改进了YOLOv3模型,使用深度可分离卷积对DarkNet53特征提取主干网络实现轻量化处理,提高检测速度,并构建多尺度长距离上下文特征提取层,使用4种不同扩张率的并行空洞卷积替代全局池化,提高模型对小尺寸缺陷的特征提取能力;构建了融合DMSA模型和改进YOLOv3模型的DMSA-YOLOv3缺陷检测模型,并应用于型钢表面多尺度缺陷检测。实验结果表明：DMSA-YOLOv3模型具有97.6%的多类别平均检测精度和55.3?frame/s的检测速度,与YOLOv3模型相比分别提升了4.7个百分点和24.5?frame/s;最小可检出20×20像素（约10×10?mm2）缺陷,与YOLOv3模型相比提高了6.25倍,有效提升了型钢表面缺陷的检测精度与检测速度。     

基于深度学习的型钢表面多形态微小缺陷检测算法

为了解决由于型钢表面缺陷形态多样、微小缺陷众多所带来的检测效率低与检测精度差的问题,提出一种基于可变形卷积与多尺度-密集特征金字塔的型钢表面缺陷检测算法——Steel-YOLOv3。首先,使用可变形卷积代替Darknet53网络部分残差单元的卷积层,从而强化特征提取网络对型钢表面多类型缺陷的特征学习能力;其次,设计了多尺度-密集特征金字塔模块：在原有YOLOv3算法的3层预测尺度上增加1层更浅层的预测尺度,再对多尺度特征图进行跨层密集连接,从而增强对密集微小缺陷的表征能力;最后,针对型钢缺陷尺寸分布特点,使用K-means维度聚类方法优化先验框尺寸并将先验框平均分配到4个对应预测尺度上。实验结果表明：Steel-YOLOv3算法具有89.24%的检测平均精度均值（mAP）,与Faster R-CNN（Faster Region-based Convolutional Neural Network）、SSD（Single Shot MultiBox Detector）、YOLOv3和YOLOv5算法相比分别提高了3.51%、26.46%、12.63%和5.71%,且所提算法显著提升了微小剥落缺陷的检出率。另外,所提算法的每秒检测图像数量达到25.62张,满足实时检测的要求,可实际应用于型钢表面缺陷的在线检测。     

板带钢表面质量实时监测体系研究

板带钢市场竞争的日益激烈,传统的检测方法正逐渐退出应用舞台,更稳定高效表的面检测解决方案成为国内外钢铁企业的找寻目标.针对实际应用的需要,从表面缺陷采集系统,缺陷处理和识别系统、质量智能评价与推理系统等3个方面首次系统地阐述了板带钢表面质量实时监测体系,以期推动目前我国板带钢表面缺陷的在线监测研究水平.     

Process knowledge based multi-class support vector classification (PK-MSVM) approach for surface defects in hot rolling

Random surface defects occur during the hot bar rolling of steels and are identified either by manual or by automated inspection techniques. Manual inspection techniques are purely based on the process knowledge of the inspector such as the location, type and kind of defects, and the primary sources of these defects. The automated techniques, to identify and classify the defects, rely on machine vision technologies and image processing algorithms based on support vector machines, wavelets, image processing and statistical inference. Both these approaches have their own advantages and limitations. To improve the accuracy of classification of these defects a process knowledge based support vector classification scheme is proposed (called PK-MSVM) which combines feature extraction task of automated inspection with the process knowledge. The defect observation data from the imaging sensor is transformed to include this process knowledge. Three attributes of the defects – length to width ratio, longitudinal location and transverse location- are used for this transformation are they are closely related to the thermo-mechanics of the rolling process. Different formulations of the multi-class support vector machines (MSVMs) are compared for this classification with or without process knowledge based transformation: one-against-one, one-against-all and Hastie’s algorithm of multi class SVM. It is found that the new approach (PK-MSVM) performs better than traditional MSVM for all the three formulations. For the best case, the performance sees a jump of more than 100%. Thus incorporating process knowledge in identification and classification does increase the reliability of inspection considerably.     

钢板表面缺陷在线视觉检测系统研究

本文设计了一套智能无损检测系统以实现对钢板表面缺陷的在线检测.检测系统由新型LED光源,明暗域结合成像光学系统、高速高分辨率线阵CCD器件、FPGA嵌入式处理系统和缺陷自动分类子系统等组成.系统对钢板表面的气泡、夹杂、结疤、划伤和压痕等主要缺陷进行无损检测,并基于Bayes决策理论,实现缺陷的自动分类功能.本系统在实验室检测指标为:宽度最大为1 800 mm,运行速度不大于1.5 m/s,振动幅度小于1 mm,横纵向检测分辨率为0.8 mm×0.8 mm,尺寸检测误差不大于1 mm.     

Sim-YOLOv5s: A method for detecting defects on the end face of lithium battery steel shells

The detection of lithium battery shell defects is an important aspect of lithium battery production. The presence of pits, R-angle injuries, hard printing, and other defects on the end face of lithium battery shells severely affects the production safety and usage safety of lithium battery products. In this study, we propose an effective defect-detection model, called Sim-YOLOv5s, for lithium battery steel shells. In this model, we propose a fast spatial pooling pyramid structure, SimSPPF, to speed up the model and embed the attention mechanism convolutional block attention module in the backbone. Contextual information can be aggregated over a large perceptual field by using the new upsampling operator, which has a larger field of perception. A cross-layer connection operation is performed to fuse shallow feature information with deep feature information. The experimental results show that the proposed Sim-YOLOv5s model has a better overall performance with a mean average precision of 88.3%, which is 6.9% better than that of YOLOv5s. Therefore, the proposed Sim-YOLOv5s can lay the foundation for the industrial implementation of real-time inspection of lithium battery products.     

改进的卷积神经网络行人检测方法

提出基于YOLOV3和DenseNet相结合的轻量化行人检测算法。加入HSV图像处理模块强化行人特征,利用卷积神经网络提取行人特征,通过k均值聚类算法筛选预测框,借鉴特征金字塔的思想做高低层特征融合和预测,利用Dense Block结构对网络轻量化进行完善,在国际广泛使用的行人数据集上进行一系列实验。实验结果表明,检测速度比现有的优秀目标检测模型YOLOV3提升了8倍,模型大小为YOLOV3的1/107,所提方法在测试集上的实时性和准确率都有所提高。     

A Neural Network Degradation Model for Computing and Updating Residual Life Distributions

The ability to accurately estimate the residual life of partially degraded components is arguably the most challenging problem in prognostic condition monitoring. This paper focuses on the development of a neural network-based degradation model that utilizes condition-based sensory signals to compute and continuously update residual life distributions of partially degraded components. Initial predicted failure times are estimated through trained neural networks using real-time sensory signals. These estimates are used to derive a prior failure time distribution for the component that is being monitored. Subsequent failure time estimates are then utilized to update the prior distributions using a Bayesian approach. The proposed methodology is tested using real world vibration-based degradation signals from rolling contact thrust bearings. The proposed methodology performed favorably when compared to other reliability-based and statistical-based benchmarks.     

Knowledge discovery in steel bar rolling mills using scheduling data and automated inspection

There are various surface defects which occur during the hot rolling of steels. It is difficult to correctly identify and control these defects due to the different inspection techniques on different materials and sizes. Also, the statistical data analysis techniques typically used like the principal component analysis, factor analysis etc. require a lot of plant data and are computationally very intensive. Before a detailed analysis of the actual cause of the defects can be done, it is necessary to separate the defects as those coming from the continuous casting or the rolling mill. Once this is done, analysis on the individual components can then be completed to find the root cause. To accomplish both these analysis, Bayesian hierarchical modeling is done on the automated inspection of the bars to form a causal relationship of the defects to the process equipments. Variance reduction model is used at the top of the analysis and regression models are used in the next level.     

Critical point defect detection for small‐pixel thin‐film transistor array applied to medical display

Thin‐film transistor (TFT) array testing technique has been used, which provides defect detection capability to control the yield of the TFT process. In the past, several defect inspection technologies have been developed and applied for the TFT array testing. When the TFT array pixel size is getting smaller and the resolution is higher, they also encounter the performance limitation on detecting the critical defect in this small‐pixel TFT array and facing a limited testing requirement. For medical display applications, the display pixels on an array panel are getting smaller and smaller; therefore, defect detection is getting more important and critical for managing yield with quality. In this study, a novel approach for defect detection was proposed. Here, the proposed voltage imaging technique is used for the TFT array test, and it provides better small‐pixel TFT array defect detection capability. The experimental results show that by using the voltage imaging technique, detecting critical point defect of TFT array can be effectively improved. And the detected small‐pixel size of TFT array panels can be smaller than 55 µm of an advanced medical display.     

An automatic surface inspection system for flat rolled steel

The paper describes a feasibility study of on-line classification of visible defects in flat rolled steel, as the steel emerges from the rolling mill. A visual signal developed by a television camera is processed by a software simulation of the proposed hardware system which stores information from the scans and alerts a small computer when a possible flaw has been detected. A pattern recognition algorithm, executable by the computer, makes the final detection and classification decision before the next coil is rolled. A detailed simulation of the hardware using photographs of steel defects is presented.     

基于神经网络的冷轧带钢表面缺陷检测

带钢表面缺陷是影响带钢质量的重要因素,对带钢进行表面缺陷检测对提高带钢质量具有重要意义。传统人工检测的方法往往不能得到令人满意的检测结果。为此,提出了采用基于前馈神经网络(FFN)的方法对在线带钢的表面缺陷进行检测,检测结果令人满意,表明了该方法的有效性。     

中厚板热轧生产调度优化方法

中厚板热轧生产调度, 是一个有优先约束、等待时间和缓冲容量有限的单机调度问题. 用AON (Activity-on-node)网络对问题进行描述, 提出并证明了面向单机调度问题的AON网络平衡定理, 根据平衡定理, 建立了以轧机利用率最大为优化目标的非线性约束优化数学模型, 并利用优化软件LINGO进行求解. 计算实例表明, 所提出的数学优化方法, 与现有的启发式方法相比, 能够获得更好的优化目标, 所得到的生产调度方案, 生产节奏稳定, 更有利于组织生产.     

Automated visual inspection of rolled metal surfaces

A prototype for an automated visual on-line metal strip inspection system is described. The system is capable of both detecting and classifying surface defects in copper alloy strips, and it has been installed for evaluation in a production line in a rolling mill. The image acquisition part of the system is based on a CCD line scan camera and condensing bright field illuminators. The inspection algorithms are based on morphological preprocessing and combined statistical and structural defect recognition. The image processing hardware consists of commercial modules. An analysis of this implementation is presented. A similar inspection principle has also been successfully applied to steel strip inspection.     

A machine vision inspector for beer bottle

A machine-vision-based beer bottle inspector is presented. The mechanical structure and electric control system are illustrated in detail. A method based on the histogram of edge points is applied for real-time determination of inspection area. For defect detection of bottle wall and bottle bottom, an algorithm based on local statistical characteristics is proposed. In bottle finish inspection, two artificial neural networks are used for low-level inspection and high-level judgment, respectively. A prototype was developed and experimental results demonstrate the feasibility of the inspector. Inspections performed by the prototype have proved the effectiveness and value of proposed algorithms in automatic real-time inspection.     

A Radial Basis Function network training algorithm using a non-symmetric partition of the input space - Application to a Model Predictive Control configuration

This work presents the non-symmetric fuzzy means algorithm which is a new methodology for training Radial Basis Function neural network models. The method is based on a non-symmetric fuzzy partition of the space of input variables which results to networks with smaller structures and better approximation capabilities compared to other state-of-the-art training procedures. The lower modeling error and the smaller size of the produced models become particularly important when they are used in online applications. This is demonstrated by integrating the model produced by the proposed algorithm in a Model Predictive Control configuration, resulting in better control performance and shorter computational times.     

一种带钢表面缺陷识别与分类的研究——基于混合加权特征和RBF网络的方法

本论文中,采用灰度直方图特征、灰度共生矩阵特征和小波变换特征的提取方法,三种特征方法的结合能够很好的实现分类的目的。在提取特征向量的基础上,本研究基于MATLAB6.5环境下的神经网络工具箱,采用了兼顾识别速度与分类准确性的RBF神经网络分类器对带钢表面缺陷进行识别与分类,此算法可以作为高速生产线的带钢表面缺陷的实时检测优选方案。     

Re-YOLOv5:一种基于结构重参数化的钢材缺陷检测方法

钢材在生产的过程中很容易产生裂纹、斑点等缺陷,而目前对于所产生缺陷的检测技术还不是很成熟。为了实现对工业钢材生产过程中所产生的钢材缺陷进行实时鲁棒检测,以YOLOv5为基础,引入了结构重参数化方法,建立了Re-YOLOv5工业钢材缺陷检测模型。在该模型中,将YOLOv5的Neck层与Head层合并为Head层,用作预测,并且加入RepVGG模块和卷积层,输出预测结果。Backbone用作特征提取,可以在改善模型推理速度的同时提高检测准确率。同时,采用改进后的空间金字塔池化模块SPP^*对候选框进行分类和修正,以获取多尺度特征信息,并引入了有助于模型加深的CCBL模块。在公开的NEU-DET钢材缺陷图片数据集上进行测试,提出的模型的检测精度可达77.8%,与基线模型YOLOv5s相比,实现了6%的精度提升,且单幅图片的推理时间仅为8.9 ms,满足工业生产实时性需求。此外,该模型所占内存较小,便于部署到工业设备中。