A Hybrid Deep Learning and Machine Learning-Based Approach to Classify Defects in Hot Rolled Steel Strips for Smart Manufacturing

Smart manufacturing is a process that optimizes factory performance and production quality by utilizing various technologies including the Internet of Things(IoT)and artificial intelligence(AI).Quality control is an important part of today’s smart manufacturing process,effectively reducing costs and enhancing operational efficiency.As technology in the industry becomes more advanced,identifying and classifying defects has become an essential element in ensuring the quality of products during the manufacturing process.In this study,we introduce a CNN model for classifying defects on hot-rolled steel strip surfaces using hybrid deep learning techniques,incorporating a global average pooling(GAP)layer and a machine learning-based SVM classifier,with the aim of enhancing accuracy.Initially,features are extracted by the VGG19 convolutional block.Then,after processing through the GAP layer,the extracted features are fed to the SVM classifier for classification.For this purpose,we collected images from publicly available datasets,including the Xsteel surface defect dataset(XSDD)and the NEU surface defect(NEU-CLS)datasets,and we employed offline data augmentation techniques to balance and increase the size of the datasets.The outcome of experiments shows that the proposed methodology achieves the highest metrics score,with 99.79%accuracy,99.80%precision,99.79%recall,and a 99.79%F1-score for the NEU-CLS dataset.Similarly,it achieves 99.64%accuracy,99.65%precision,99.63%recall,and a 99.64%F1-score for the XSDD dataset.A comparison of the proposed methodology to the most recent study showed that it achieved superior results as compared to the other studies.

Properties of radiation defects and threshold energy of displacement in zirconium hydride obtained by new deep-learning potential

Zirconium hydride(ZrH_(2)) is an ideal neutron moderator material. However, radiation effect significantly changes its properties, which affect its behavior and the lifespan of the reactor. The threshold energy of displacement is an important quantity of the number of radiation defects produced, which helps us to predict the evolution of radiation defects in ZrH_(2).Molecular dynamics(MD) and ab initio molecular dynamics(AIMD) are two main methods of calculating the threshold energy of displacement. The MD simulations with empirical potentials often cannot accurately depict the transitional states that lattice atoms must surpass to reach an interstitial state. Additionally, the AIMD method is unable to perform largescale calculation, which poses a computational challenge beyond the simulation range of density functional theory. Machine learning potentials are renowned for their high accuracy and efficiency, making them an increasingly preferred choice for molecular dynamics simulations. In this work, we develop an accurate potential energy model for the ZrH_(2) system by using the deep-potential(DP) method. The DP model has a high degree of agreement with first-principles calculations for the typical defect energy and mechanical properties of the ZrH_(2) system, including the basic bulk properties, formation energy of point defects, as well as diffusion behavior of hydrogen and zirconium. By integrating the DP model with Ziegler–Biersack–Littmark(ZBL) potential, we can predict the threshold energy of displacement of zirconium and hydrogen in ε-ZrH_(2).

Method for Detecting Industrial Defects in Intelligent Manufacturing Using Deep Learning

With the advent of Industry 4.0,marked by a surge in intelligent manufacturing,advanced sensors embedded in smart factories now enable extensive data collection on equipment operation.The analysis of such data is pivotal for ensuring production safety,a critical factor in monitoring the health status of manufacturing apparatus.Conventional defect detection techniques,typically limited to specific scenarios,often require manual feature extraction,leading to inefficiencies and limited versatility in the overall process.Our research presents an intelligent defect detection methodology that leverages deep learning techniques to automate feature extraction and defect localization processes.Our proposed approach encompasses a suite of components:the high-level feature learning block(HLFLB),the multi-scale feature learning block(MSFLB),and a dynamic adaptive fusion block(DAFB),working in tandem to extract meticulously and synergistically aggregate defect-related characteristics across various scales and hierarchical levels.We have conducted validation of the proposed method using datasets derived from gearbox and bearing assessments.The empirical outcomes underscore the superior defect detection capability of our approach.It demonstrates consistently high performance across diverse datasets and possesses the accuracy required to categorize defects,taking into account their specific locations and the extent of damage,proving the method’s effectiveness and reliability in identifying defects in industrial components.

Cross-Project Software Defect Prediction Based on SMOTE and Deep Canonical Correlation Analysis

Cross-Project Defect Prediction(CPDP)is a method that utilizes historical data from other source projects to train predictive models for defect prediction in the target project.However,existing CPDP methods only consider linear correlations between features(indicators)of the source and target projects.These models are not capable of evaluating non-linear correlations between features when they exist,for example,when there are differences in data distributions between the source and target projects.As a result,the performance of such CPDP models is compromised.In this paper,this paper proposes a novel CPDP method based on Synthetic Minority Oversampling Technique(SMOTE)and Deep Canonical Correlation Analysis(DCCA),referred to as S-DCCA.Canonical Correlation Analysis(CCA)is employed to address the issue of non-linear correlations between features of the source and target projects.S-DCCA extends CCA by incorporating the MlpNet model for feature extraction from the dataset.The redundant features are then eliminated by maximizing the correlated feature subset using the CCA loss function.Finally,cross-project defect prediction is achieved through the application of the SMOTE data sampling technique.Area Under Curve(AUC)and F1 scores(F1)are used as evaluation metrics.This paper conducted experiments on 27 projects from four public datasets to validate the proposed method.The results demonstrate that,on average,our method outperforms all baseline approaches by at least 1.2%in AUC and 5.5%in F1 score.This indicates that the proposed method exhibits favorable performance characteristics.

Intelligent diagnosis of atrial septal defect in children using echocardiography with deep learning

Background Atrial septal defect(ASD)is one of the most common congenital heart diseases.The diagnosis of ASD via transthoracic echocardiography is subjective and time-consuming.Methods The objective of this study was to evaluate the feasibility and accuracy of automatic detection of ASD in children based on color Doppler echocardiographic static images using end-to-end convolutional neural networks.The proposed depthwise separable convolution model identifies ASDs with static color Doppler images in a standard view.Among the standard views,we selected two echocardiographic views,i.e.,the subcostal sagittal view of the atrium septum and the low parasternal four-chamber view.The developed ASD detection system was validated using a training set consisting of 396 echocardiographic images corresponding to 198 cases.Additionally,an independent test dataset of 112 images corresponding to 56 cases was used,including 101 cases with ASDs and 153 cases with normal hearts.Results The average area under the receiver operating characteristic curve,recall,precision,specificity,F1-score,and accuracy of the proposed ASD detection model were 91.99,80.00,82.22,87.50,79.57,and 83.04,respectively.Conclusions The proposed model can accurately and automatically identify ASD,providing a strong foundation for the intelligent diagnosis of congenital heart diseases.

Track Defects Recognition Based on Axle-Box Vibration Acceleration and Deep- Learning Techniques

As an important component of load transfer,various fatigue damages occur in the track as the rail service life and train traffic increase gradually,such as rail corrugation,rail joint damage,uneven thermite welds,rail squats fas-tener defects,etc.Real-time recognition of track defects plays a vital role in ensuring the safe and stable operation of rail transit.In this paper,an intelligent and innovative method is proposed to detect the track defects by using axle-box vibration acceleration and deep learning network,and the coexistence of the above-mentioned typical track defects in the track system is considered.Firstly,the dynamic relationship between the track defects(using the example of the fastening defects)and the axle-box vibration acceleration(ABVA)is investigated using the dynamic vehicle-track model.Then,a simulation model for the coupled dynamics of the vehicle and track with different track defects is established,and the wavelet power spectrum(WPS)analysis is performed for the vibra-tion acceleration signals of the axle box to extract the characteristic response.Lastly,using wavelet spectrum photos as input,an automatic detection technique based on the deep convolution neural network(DCNN)is sug-gested to realize the real-time intelligent detection and identification of various track problems.Thefindings demonstrate that the suggested approach achieves a 96.72%classification accuracy.

Surface Defect Detection and Evaluation Method of Large Wind Turbine Blades Based on an Improved Deeplabv3+Deep Learning Model

The accumulation of defects on wind turbine blade surfaces can lead to irreversible damage,impacting the aero-dynamic performance of the blades.To address the challenge of detecting and quantifying surface defects on wind turbine blades,a blade surface defect detection and quantification method based on an improved Deeplabv3+deep learning model is proposed.Firstly,an improved method for wind turbine blade surface defect detection,utilizing Mobilenetv2 as the backbone feature extraction network,is proposed based on an original Deeplabv3+deep learning model to address the issue of limited robustness.Secondly,through integrating the concept of pre-trained weights from transfer learning and implementing a freeze training strategy,significant improvements have been made to enhance both the training speed and model training accuracy of this deep learning model.Finally,based on segmented blade surface defect images,a method for quantifying blade defects is proposed.This method combines image stitching algorithms to achieve overall quantification and risk assessment of the entire blade.Test results show that the improved Deeplabv3+deep learning model reduces training time by approximately 43.03%compared to the original model,while achieving mAP and MIoU values of 96.87%and 96.93%,respectively.Moreover,it demonstrates robustness in detecting different surface defects on blades across different back-grounds.The application of a blade surface defect quantification method enables the precise quantification of dif-ferent defects and facilitates the assessment of risk levels associated with defect measurements across the entire blade.This method enables non-contact,long-distance,high-precision detection and quantification of surface defects on the blades,providing a reference for assessing surface defects on wind turbine blades.

Identification of XLPE cable insulation defects based on deep learning

The insulation aging of cross-linked polyethylene(XLPE)cables is the main reason for the reduction in cable life.There is currently a lack of rapid and effective methods for detecting cable insulation defects in power-related sectors.To this end,this paper presents a method for identifying insulation defects in XLPE cables based on deep learning algorithms.First,the principle of the harmonic method for detecting cable insulation defects is introduced.Second,the ANSYS software is used to simulate the cable insulation layer containing bubbles,protrusions,and water tree defects,and the effects of each type of defect on the magnetic field strength and eddy loss current of the cable insulation layer are analyzed.Then,a total of 10 characteristic quantities of the total harmonic content and 2nd to 10th harmonic currents are constructed to establish a database of cable insulation defects.Finally,the deep learning algorithm,long short-term memory(LSTM),is used to accurately identify the types of insulation defects in cables.The results indicate that the LSTM algorithm can effectively diagnose and identify insulation defects in cables with an accuracy of 95.83%.

Deep Learning-Based Model for Defect Detection and Localization on Photovoltaic Panels

The Problem of Photovoltaic(PV)defects detection and classification has been well studied.Several techniques exist in identifying the defects and localizing them in PV panels that use various features,but suffer to achieve higher performance.An efficient Real-Time Multi Variant Deep learning Model(RMVDM)is presented in this article to handle this issue.The method considers different defects like a spotlight,crack,dust,and micro-cracks to detect the defects as well as loca-lizes the defects.The image data set given has been preprocessed by applying the Region-Based Histogram Approximation(RHA)algorithm.The preprocessed images are applied with Gray Scale Quantization Algorithm(GSQA)to extract the features.Extracted features are trained with a Multi Variant Deep learning model where the model trained with a number of layers belongs to different classes of neurons.Each class neuron has been designed to measure Defect Class Support(DCS).At the test phase,the input image has been applied with different operations,and the features extracted passed through the model trained.The output layer returns a number of DCS values using which the method identifies the class of defect and localizes the defect in the image.Further,the method uses the Higher-Order Texture Localization(HOTL)technique in localizing the defect.The pro-posed model produces efficient results with around 97%in defect detection and localization with higher accuracy and less time complexity.

Deep Learning Based Underground Sewer Defect Classification Using a Modified RegNet

The sewer system plays an important role in protecting rainfall and treating urban wastewater.Due to the harsh internal environment and complex structure of the sewer,it is difficult to monitor the sewer system.Researchers are developing different methods,such as the Internet of Things and Artificial Intelligence,to monitor and detect the faults in the sewer system.Deep learning is a promising artificial intelligence technology that can effectively identify and classify different sewer system defects.However,the existing deep learning based solution does not provide high accuracy prediction and the defect class considered for classification is very small,which can affect the robustness of the model in the constraint environment.As a result,this paper proposes a sewer condition monitoring framework based on deep learning,which can effectively detect and evaluate defects in sewer pipelines with high accuracy.We also introduce a large dataset of sewer defects with 20 different defect classes found in the sewer pipeline.This study modified the original RegNet model by modifying the squeeze excitation(SE)block and adding the dropout layer and Leaky Rectified Linear Units(LeakyReLU)activation function in the Block structure of RegNet model.This study explored different deep learning methods such as RegNet,ResNet50,very deep convolutional networks(VGG),and GoogleNet to train on the sewer defect dataset.The experimental results indicate that the proposed system framework based on the modified-RegNet(RegNet+)model achieves the highest accuracy of 99.5 compared with the commonly used deep learning models.The proposed model provides a robust deep learning model that can effectively classify 20 different sewer defects and be utilized in real-world sewer condition monitoring applications.

Visualization for Explanation of Deep Learning-Based Defect Detection Model Using Class Activation Map

Recently,convolutional neural network(CNN)-based visual inspec-tion has been developed to detect defects on building surfaces automatically.The CNN model demonstrates remarkable accuracy in image data analysis;however,the predicted results have uncertainty in providing accurate informa-tion to users because of the“black box”problem in the deep learning model.Therefore,this study proposes a visual explanation method to overcome the uncertainty limitation of CNN-based defect identification.The visual repre-sentative gradient-weights class activation mapping(Grad-CAM)method is adopted to provide visually explainable information.A visualizing evaluation index is proposed to quantitatively analyze visual representations;this index reflects a rough estimate of the concordance rate between the visualized heat map and intended defects.In addition,an ablation study,adopting three-branch combinations with the VGG16,is implemented to identify perfor-mance variations by visualizing predicted results.Experiments reveal that the proposed model,combined with hybrid pooling,batch normalization,and multi-attention modules,achieves the best performance with an accuracy of 97.77%,corresponding to an improvement of 2.49%compared with the baseline model.Consequently,this study demonstrates that reliable results from an automatic defect classification model can be provided to an inspector through the visual representation of the predicted results using CNN models.

Defect Detection Model Using Time Series Data Augmentation and Transformation

Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal dependence,and noise.Therefore,methodologies for data augmentation and conversion of time series data into images for analysis have been studied.This paper proposes a fault detection model that uses time series data augmentation and transformation to address the problems of data imbalance,temporal dependence,and robustness to noise.The method of data augmentation is set as the addition of noise.It involves adding Gaussian noise,with the noise level set to 0.002,to maximize the generalization performance of the model.In addition,we use the Markov Transition Field(MTF)method to effectively visualize the dynamic transitions of the data while converting the time series data into images.It enables the identification of patterns in time series data and assists in capturing the sequential dependencies of the data.For anomaly detection,the PatchCore model is applied to show excellent performance,and the detected anomaly areas are represented as heat maps.It allows for the detection of anomalies,and by applying an anomaly map to the original image,it is possible to capture the areas where anomalies occur.The performance evaluation shows that both F1-score and Accuracy are high when time series data is converted to images.Additionally,when processed as images rather than as time series data,there was a significant reduction in both the size of the data and the training time.The proposed method can provide an important springboard for research in the field of anomaly detection using time series data.Besides,it helps solve problems such as analyzing complex patterns in data lightweight.

Faster Metallic Surface Defect Detection Using Deep Learning with Channel Shuffling

Deep learning has been constantly improving in recent years,and a significant number of researchers have devoted themselves to the research of defect detection algorithms.Detection and recognition of small and complex targets is still a problem that needs to be solved.The authors of this research would like to present an improved defect detection model for detecting small and complex defect targets in steel surfaces.During steel strip production,mechanical forces and environmental factors cause surface defects of the steel strip.Therefore,the detection of such defects is key to the production of high-quality products.Moreover,surface defects of the steel strip cause great economic losses to the high-tech industry.So far,few studies have explored methods of identifying the defects,and most of the currently available algorithms are not sufficiently effective.Therefore,this study presents an improved real-time metallic surface defect detection model based on You Only Look Once(YOLOv5)specially designed for small networks.For the smaller features of the target,the conventional part is replaced with a depthwise convolution and channel shuffle mechanism.Then assigning weights to Feature Pyramid Networks(FPN)output features and fusing them,increases feature propagation and the network’s characterization ability.The experimental results reveal that the improved proposed model outperforms other comparable models in terms of accuracy and detection time.The precision of the proposed model achieved by mAP@0.5 is 77.5%on the Northeastern University,Dataset(NEU-DET)and 70.18%on the GC10-DET datasets.

YOLO-RLC:An Advanced Target-Detection Algorithm for Surface Defects of Printed Circuit Boards Based on YOLOv5

Printed circuit boards(PCBs)provide stable connections between electronic components.However,defective printed circuit boards may cause the entire equipment system to malfunction,resulting in incalculable losses.Therefore,it is crucial to detect defective printed circuit boards during the generation process.Traditional detection methods have low accuracy in detecting subtle defects in complex background environments.In order to improve the detection accuracy of surface defects on industrial printed circuit boards,this paper proposes a residual large kernel network based on YOLOv5(You Only Look Once version 5)for PCBs surface defect detection,called YOLO-RLC(You Only Look Once-Residual Large Kernel).Build a deep large kernel backbone to expand the effective field of view,capture global informationmore efficiently,and use 1×1 convolutions to balance the depth of the model,improving feature extraction efficiency through reparameterization methods.The neck network introduces a bidirectional weighted feature fusion network,combined with a brand-new noise filter and feature enhancement extractor,to eliminate noise information generated by information fusion and recalibrate information from different channels to improve the quality of deep features.Simplify the aspect ratio of the bounding box to alleviate the issue of specificity values.After training and testing on the PCB defect dataset,our method achieved an average accuracy of 97.3%(mAP50)after multiple experiments,which is 4.1%higher than YOLOv5-S,with an average accuracy of 97.6%and an Frames Per Second of 76.7.The comparative analysis also proves the superior performance and feasibility of YOLO-RLC in PCB defect detection.

DSN-BR-Based Online Inspection Method and Application for Surface Defects of Pharmaceutical Products in Aluminum-Plastic Blister Packages

Ensuring high product quality is of paramount importance in pharmaceutical drug manufacturing,as it is subject to rigorous regulatory practices.This study presents a research focused on the development of an on-line detection method and system for identifying surface defects in pharmaceutical products packaged in aluminum-plastic blisters.Firstly,the aluminum-plastic blister packages exhibit multi-scale features and inter-class indistinction.To address this,the deep semantic network with boundary refinement(DSN-BR)model is proposed,which leverages semantic segmentation domain knowledge,to accurately segment the defects in pixel level.Additionally,a specialized image acquisition module that minimizes the impact of ambient light is established,ensuring high-quality image capture.Finally,the image acquisition module,image detection module,and data management module are designed to construct a comprehensive online surface defect detection system.To validate the effectiveness of our approach,we employ a real dataset for instance verification on the implemented system.The experimental results substantiate the outstanding performance of the DSN-BR,achieving the mean intersection over union(MIoU)of 90.5%.Furthermore,the proposed system achieves an inference speed of up to 14.12 f/s,while attaining an F1-Score of 98.25%.These results demonstrate that the system meets the actual needs of the enterprise and provides theoretical and methodological support for intelligent inspection of product surface quality.By standardizing the control process of pharmaceutical manufacturing and improving the management capability of the manufacturing process,our approach holds significant market application prospects.

A Composite Transformer-Based Multi-Stage Defect Detection Architecture for Sewer Pipes

Urban sewer pipes are a vital infrastructure in modern cities,and their defects must be detected in time to prevent potential malfunctioning.In recent years,to relieve the manual efforts by human experts,models based on deep learning have been introduced to automatically identify potential defects.However,these models are insufficient in terms of dataset complexity,model versatility and performance.Our work addresses these issues with amulti-stage defect detection architecture using a composite backbone Swin Transformer.Themodel based on this architecture is trained using a more comprehensive dataset containingmore classes of defects.By ablation studies on the modules of combined backbone Swin Transformer,multi-stage detector,test-time data augmentation and model fusion,it is revealed that they all contribute to the improvement of detection accuracy from different aspects.The model incorporating all these modules achieves the mean Average Precision(mAP)of 78.6% at an Intersection over Union(IoU)threshold of 0.5.This represents an improvement of 14.1% over the ResNet50 Faster Region-based Convolutional Neural Network(R-CNN)model and a 6.7% improvement over You Only Look Once version 6(YOLOv6)-large,the highest in the YOLO methods.In addition,for other defect detection models for sewer pipes,although direct comparison with themis infeasible due to the unavailability of their private datasets,our results are obtained from a more comprehensive dataset and have superior generalization capabilities.

Defects evolution in n-type 4H-SiC induced by electron irradiation and annealing

Radiation damage produced in 4H-SiC by electrons of different doses is presented by using multiple characterization techniques. Raman spectra results indicate that SiC crystal structures are essentially impervious to 10 Me V electron irradiation with doses up to 3000 kGy. However, irradiation indeed leads to the generation of various defects, which are evaluated through photoluminescence(PL) and deep level transient spectroscopy(DLTS). The PL spectra feature a prominent broad band centered at 500 nm, accompanied by several smaller peaks ranging from 660 to 808 nm. The intensity of each PL peak demonstrates a linear correlation with the irradiation dose, indicating a proportional increase in defect concentration during irradiation. The DLTS spectra reveal several thermally unstable and stable defects that exhibit similarities at low irradiation doses.Notably, after irradiating at the higher dose of 1000 kGy, a new stable defect labeled as R_(2)(Ec-0.51 eV) appeared after annealing at 800 K. Furthermore, the impact of irradiation-induced defects on SiC junction barrier Schottky diodes is discussed. It is observed that high-dose electron irradiation converts SiC n-epilayers to semi-insulating layers. However, subjecting the samples to a temperature of only 800 K results in a significant reduction in resistance due to the annealing out of unstable defects.

Wafer Defect Map Pattern Recognition Based on Improved ResNet

The defect detection of wafers is an important part of semiconductor manufacturing.The wafer defect map formed from the defects can be used to trace back the problems in the production process and make improvements in the yield of wafer manufacturing.Therefore,for the pattern recognition of wafer defects,this paper uses an improved ResNet convolutional neural network for automatic pattern recognition of seven common wafer defects.On the basis of the original ResNet,the squeeze-and-excitation(SE)attention mechanism is embedded into the network,through which the feature extraction ability of the network can be improved,key features can be found,and useless features can be suppressed.In addition,the residual structure is improved,and the depth separable convolution is added to replace the traditional convolution to reduce the computational and parametric quantities of the network.In addition,the network structure is improved and the activation function is changed.Comprehensive experiments show that the precision of the improved ResNet in this paper reaches 98.5%,while the number of parameters is greatly reduced compared with the original model,and has well results compared with the common convolutional neural network.Comprehensively,the method in this paper can be very good for pattern recognition of common wafer defect types,and has certain application value.

Printed Circuit Board (PCB) Surface Micro Defect Detection Model Based on Residual Network with Novel Attention Mechanism

Printed Circuit Board(PCB)surface tiny defect detection is a difficult task in the integrated circuit industry,especially since the detection of tiny defects on PCB boards with large-size complex circuits has become one of the bottlenecks.To improve the performance of PCB surface tiny defects detection,a PCB tiny defects detection model based on an improved attention residual network(YOLOX-AttResNet)is proposed.First,the unsupervised clustering performance of the K-means algorithm is exploited to optimize the channel weights for subsequent operations by feeding the feature mapping into the SENet(Squeeze and Excitation Network)attention network;then the improved K-means-SENet network is fused with the directly mapped edges of the traditional ResNet network to form an augmented residual network(AttResNet);and finally,the AttResNet module is substituted for the traditional ResNet structure in the backbone feature extraction network of mainstream excellent detection models,thus improving the ability to extract small features from the backbone of the target detection network.The results of ablation experiments on a PCB surface defect dataset show that AttResNet is a reliable and efficient module.In Torify the performance of AttResNet for detecting small defects in large-size complex circuit images,a series of comparison experiments are further performed.The results show that the AttResNet module combines well with the five best existing target detection frameworks(YOLOv3,YOLOX,Faster R-CNN,TDD-Net,Cascade R-CNN),and all the combined new models have improved detection accuracy compared to the original model,which suggests that the AttResNet module proposed in this paper can help the detection model to extract target features.Among them,the YOLOX-AttResNet model proposed in this paper performs the best,with the highest accuracy of 98.45% and the detection speed of 36 FPS(Frames Per Second),which meets the accuracy and real-time requirements for the detection of tiny defects on PCB surfaces.This study can provide some new ideas for other real-time online detection tasks of tiny targets with high-resolution images.

Industry-Oriented Detection Method of PCBA Defects Using Semantic Segmentation Models

Automated optical inspection(AOI)is a significant process in printed circuit board assembly(PCBA)production lines which aims to detect tiny defects in PCBAs.Existing AOI equipment has several deficiencies including low throughput,large computation cost,high latency,and poor flexibility,which limits the efficiency of online PCBA inspection.In this paper,a novel PCBA defect detection method based on a lightweight deep convolution neural network is proposed.In this method,the semantic segmentation model is combined with a rule-based defect recognition algorithm to build up a defect detection frame-work.To improve the performance of the model,extensive real PCBA images are collected from production lines as datasets.Some optimization methods have been applied in the model according to production demand and enable integration in lightweight computing devices.Experiment results show that the production line using our method realizes a throughput more than three times higher than traditional methods.Our method can be integrated into a lightweight inference system and pro-mote the flexibility of AOI.The proposed method builds up a general paradigm and excellent example for model design and optimization oriented towards industrial requirements.