Aluminum Alloy is the main structure material in aerospace industry. On-line defect detection for Aluminum alloy in pulsed gas tungsten arc welding (GTAW) is still challenging, especially for increasing application of robotics. This paper presents an intelligent methodology for real-time evaluation of weld penetration defects based on arc audible sound sensing for aluminum alloy in robotic pulsed GTAW. The generation mechanism of arc sound was investigated using correlation analysis, high-speed camera observing and frequency spectrum analysis before denoising of arc sound. Two feature selection approaches based on Fisher distance and principle component analysis (PCA) were developed to select frequency components related to seam defects, and then, their performance were qualitatively and quantitatively analyzed. Finally, a new classification model integrating support vector machine with grid search optimization and cross-validation (SVM-GSCV) was proposed to identify under penetration, normal penetration and burning through. The proposed methodologies were verified to be effective with high accuracy and robustness. This paper can provide some guidance for condition monitoring of additive manufacturing (AM) or process industry.

In engineering practice, mechanical equipment is usually in polytropic working conditions, where the data distribu-tion of training set and test set is inconsistent, resulting in insuffi-cient generalization ability of the intelligent diagnosis model. Sim-ultaneously, different tasks often need to be modeled separately. Domain adaptation, as one of the research contents of transfer learning, has certain advantages in solving the problem of incon-sistent feature distribution. This paper designs and establishes a Domain Adaptation Framework based on Multi-scale Mixed Do-main Feature (DA-MMDF) for cross-domain intelligent fault di-agnosis of rolling bearings under polytropic working conditions. The proposed method first uses the multi-scale mixed domain fea-ture extractor to obtain features from the collected data, which constructs a complete feature space through Variational Mode De-composition (VMD) and mixed domain feature extraction to fully mine the state information and intrinsic attributes of the vibration signal. Secondly, the dimensionality reduction and optimization of features are achieved through extreme gradient promotion, and meaningful and sensitive features are selected according to the im-portance of features to eliminate redundant information. The op-timized important features are combined with the manifold em-bedded distribution alignment method to realize the distribution alignment of data in different fields and cross-domain diagnosis. In order to verify the effectiveness of the proposed approach, the rolling bearing data sets gathered from the laboratories are em-ployed and analyzed. The analysis result confirms that DA-MMDF is able to achieve effective transfer diagnosis between polytropic working conditions. Compared with traditional intelligent fault di-agnosis methods and domain adaptation methods, the method pro-posed in this article achieved state-of-the-art performances.

Anomaly detection is the cornerstone for the health management of rolling element bearings. The unsupervised learning model for anomaly detection driven only by normal data has received increasing attention in recent years. In this article, an innovative deep-learning-based model, namely, memory residual regression autoencoder (MRRAE), is developed to improve the accuracy of anomaly detection in bearing condition monitoring. The memory module and autoregressive estimator are applied to calculate the probability density distribution of the latent memory residual representation. The reconstruction errors and surprisal values of the proposed model are used to detect the abnormal condition of bearing. To verify the superiority of the proposed method in anomaly detection, two sets of run-to-failure experimental data set gathered from the laboratories are studied and analyzed. The result demonstrates that the proposed MRRAE model achieves superior performance compared with several conventional and deep-learning-based anomaly detection methods. Furthermore, the proposed method pays close attention to the special structure of bearing vibration signal and provides a new way for explaining the decision-making processes of deep neural networks.

Strict industry demands necessitate better penetration classification in gas tungsten arc welding (GTAW) of aluminum alloys. An innovative classification method, namely time-frequency spectrogram based convolutional neural network (TF-CNN) has been developed in this paper. The logarithmic time-frequency spectrograms are used to interpret the raw arc sound data in the time-frequency domain and served as the input of an optimized convolutional neural network (CNN) model for welding quality classification. The L-ReLU activation function and RMSprop optimizer are utilized to improve the property of CNN. Effective arc sound features are extracted automatically by the CNN model. The proposed method is experimentally verified to recognize four penetration states with an average accuracy of 98.2%, which significantly outperformed four typical comparison models. The generation mechanism and characteristics of arc sound are discussed as is the adaptive feature extraction of TF-CNN. These studies expand the application scope of CNN to intelligent welding and achieve notable effects.