Joining of carbon fibre reinforced polymer (CFRP) composites and aluminium alloys – A review Part A Applied science and manufacturing

Automated defect inspection system for metal surfaces based on deep learning and data augmentation

Welding quality directly affects the welding structure’s service performance and life. Hence, the effective monitoring welding defects is essential to ensure the quality of the weld structure. Owing to the non-uniformity of the shape, position and size of welding defects, it is a complicated task to analyze and evaluate the acquired welding defects images manually. Fortunately, deep learning has been successfully applied to image analysis and target recognition. However, the use of deep learning to identify welding defects is time-consuming and less accurate due to the lack of adequate training data samples, which easily cause redundancy into the classifier. In this situation, we proposed a new transfer learning model based on MobileNet as a welding defect feature extractor. By using the ImageNet dataset (non-welding defect data) to pre-train a MobileNet model, migrate the MobileNet model to the welding defects classification field. This article suggested a new TL-MobileNet structure by adding a new Full Connection layer (FC-128) and a Softmax classifier into a traditional model called MobileNet. The entire training process of TL-MobileNet model has been successfully optimized by the DropBlock technology and Global average pooling (GAP) method. They can effectively accelerate the convergence rate and improve the classification network generalization. By testing the proposed TL-MobileNet on the welding defects dataset, it turned out our model prediction accuracy has arrived at 97.69%. The experimental results show that in several aspects, TL-MobileNet have better performance than other transfer learning models and traditional neural network methods.

/pdf/a-new-image-recognition-and-classification-method-combining-kcoqpwkb5g.pdf

A New Image Recognition and Classification Method Combining Transfer Learning Algorithm and MobileNet Model for Welding Defects

Manufacturing quality prediction model, as an effective measure to monitor the quality in advance, has been developed using various data-driven techniques. However, multi-parameter in multi-stage of the modern manufacturing industry brings about the curse of dimensionality, leading to the difficulties for feature extraction, learning and quality modeling. To address this issue, three dimension reduction techniques are investigated in this paper, i.e., principal component analysis (PCA), locally linear embedding (LLE), and isometric mapping (Isomap). Specifically, the PCA is a linear dimension reduction technique, the LLE is a nonlinear reduction technique with local perspective, and the Isomap is a nonlinear reduction technique from global perspective. After getting the low-dimensional information from the PCA, the LLE, and the Isomap methods respectively, a support vector machine (SVM) is utilized for modeling. To reveal the effectiveness of the dimension reduction techniques and compare the difference of the three dimension reduction techniques, two experimental manufacturing data are collected from a competition about manufacturing quality control in Tianchi Data Lab of China. The comparison experiments indicate that the dimension reduction techniques have capacity for improving the SVM modeling performance indeed, and the Isomap–SVM model with the nonlinear global dimension reduction outperforms all the candidate models in terms of qualitative and quantitative analysis.

A comparison of dimension reduction techniques for support vector machine modeling of multi-parameter manufacturing quality prediction

Real-time penetration state monitoring using convolutional neural network for laser welding of tailor rolled blanks

Welding defects detection based on deep learning with multiple optical sensors during disk laser welding of thick plates

In order to explore the relationship between the welding process and welded quality, a multiple sensor fusion system was built to obtain the photodiode and visible light information during the welding. Features of keyhole, plasma and spatters were extracted from five sensors, including two photodiode sensors, one spectrometer sensor, one ultraviolet and visible light sensing camera and one auxiliary illumination sensing camera, 15 features were analyzed by normalization and principle component analysis, and principle component numbers was chosen as input parameters of support vector machine classification, Three weld quality types were defined according to the weld seam width and weld depth. The overall accuracy of training data was 98%, and the overall accuracy of testing data was 91%, respectively. Experimental results showed that the estimation on welding status was accurate and effective, thus providing an experimental example of monitoring high-power disk laser welding quality.

Prediction of high power laser welding status based on PCA and SVM classification of multiple sensors

Dissimilar joining of 304 stainless steel to PMMA is accomplished by a Nd: YAG pulsed laser. The light and process were represented by laser peak power, welding speed, defocus, pulse duration, pulse frequency and Ar gas flow rate, as these experiment parameters have significant effects on joints strength, weld width, joints quality and appearance. The welding parameters were optimized by an orthogonal experiment design, then the validation and optimization experiments were carried out to obtain the best combination of process parameters. A maximum weld strength of the joints of 974 N was achieved (the joints efficiency reached 70% of the base PMMA), much larger than ever reported. Bead profiles such as bead width and mechanical strength were investigated, and five types of 304 stainless steel-PMMA welded joints and three kinds of fracture model have been proposed.

Optimization of weld strength for laser welding of steel to PMMA using Taguchi design method

The spectrometer is applied in this paper to get the in-process information of welding defects during the high-power disk laser welding process. The high-dimensional signal captured by the spectrometer is fed into a data-driven framework based on stacked auto-encoder (SAE) to automatically extract salient features for performing real-time welding defects detection. The particle swarm optimization (PSO) algorithm is employed to optimize the proposed data-driven framework by acquiring the global optimal parameters to strengthen its capability in extracting the representative features from the original high-dimensional signal. The extracted features are classified by a softmax classifier to get the real-time identifications of the welding defects. The proposed framework is compared with the conventional shallow artificial intelligent methods, such as back-propagation (BP) neural network and support vector machine (SVM), and reveals better performance. The actual welding experiments under different welding parameters are implemented to validate the detection accuracy of our proposed data-driven framework. This paper provides an effective framework for the detection of the high-power disk laser welding status in real-time.

Data-Driven Detection of Laser Welding Defects Based on Real-Time Spectrometer Signals

This paper researches the magneto-optical (MO) imaging law of weld cracks under alternating magnetic field excitation. Weld surface and subsurface cracks are detected by a MO sensor, and the relationship between the MO images’ characteristics and the magnetic field strength is analyzed based on the Faraday MO effect. A magnetic dipole model is proposed to study the magnetic field distribution over the weld crack. A finite element analysis (FEA) model of the weld crack is established, and the relationship between the magnetic flux leakage signal and the crack width is analyzed, which is useful for identifying cracks either on the surface or on the subsurface of the weld. A MO imaging nondestructive testing (NDT) experiment is carried out to detect weld cracks under alternating magnetic field excitation, and the difference among weld cracks is obtained by analyzing the gray values of the weld cracks’ MO images. Research results show that the magnetic flux leakage signals of weld surface and subsurface cracks can be clearly distinguished, the magnetic field intensity of the surface cracks is larger than that of the subsurface cracks at the same width, and the MO image of the weld cracks can reflect the intensity of the magnetic field through varied brightness, that is, the gray value of the MO image can match the corresponding magnetic field intensity.

Nanfeng Zhang

Papers

Welding defects detection based on deep learning with multiple optical sensors during disk laser welding of thick plates

Prediction of high power laser welding status based on PCA and SVM classification of multiple sensors

Optimization of weld strength for laser welding of steel to PMMA using Taguchi design method

Data-Driven Detection of Laser Welding Defects Based on Real-Time Spectrometer Signals

Weld cracks nondestructive testing based on magneto-optical imaging under alternating magnetic field excitation