In actual welding scenarios, an effective path planner is needed to find a collision-free path in the configuration space for the welding manipulator with obstacles around. However, as a state-of-the-art method, the sampling-based planner only satisfies the probability completeness and its computational complexity is sensitive with state dimension. In this paper, we propose a path planner for welding manipulators based on deep reinforcement learning for solving path planning problems in high-dimensional continuous state and action spaces. Compared with the sampling-based method, it is more robust and is less sensitive with state dimension. In detail, to improve the learning efficiency, we introduce the inverse kinematics module to provide prior knowledge while a gain module is also designed to avoid the local optimal policy, we integrate them into the training algorithm. To evaluate our proposed planning algorithm in multiple dimensions, we conducted multiple sets of path planning experiments for welding manipulators. The results show that our method not only improves the convergence performance but also is superior in terms of optimality and robustness of planning compared with most other planning algorithms.

The remaining useful life (RUL) prediction is of significance to the health management of bearings. Recently, deep learning has been widely investigated for bearing RUL prediction due to its great success in sequence learning. However, the improvement of the prediction accuracy of existing deep learning algorithms heavily relies on feature engineering such as handcrafted feature generation and time–frequency transformation, which increase the complexity and difficulty of the actual deployment. In this paper, a novel spatial attention-based convolutional transformer (SAConvFormer) is proposed to establish an accurate bearing RUL prediction model based on raw vibration data without prior knowledge or feature engineering. In this algorithm, firstly, a convolutional neural network enhanced by a spatial attention mechanism is proposed to squeeze the feature maps and extract the local and global features from raw bearing vibration data effectively. Then, the extracted senior features are fed into a transformer network to further explore the sequential patterns relevant to the bearing RUL. An experimental study using the XJTU-SY rolling bearings dataset revealed the merits of the proposed deep learning algorithm in terms of root-mean-square-error (RMSE) and mean-absolute-error (MAE) in comparison with other state-of-the-art algorithms.

Emotions are complex, and people vary greatly in their accuracy in recognizing their own emotions and those of others. With advances in computer science and neuroscience, there is a desire to use automated techniques to help people identify emotions. Bio-electrical signals have been proven effective for emotion detection, but the acquisition of conventional electrocardiogram (ECG) and EEG requires medical-specific equipment, which is very expensive, uncomfortable, and inconvenient due to the large number of electrodes and the hair-covered scalp. In this article, a novel emotion recognition method based on the feature fusion of single-lead EEG and ECG signals is proposed, using the long short term memory (LSTM)-MLP-based model and the CNN-based model for feature fusion and classification, respectively, with fivefold cross-validation for validation. The ECG and EEG signals of 15 participants were collected in five states: 1) happy; 2) relaxed; 3) calm; 4) sad; and 5) afraid, each of which was stimulated using the participants’ own proposed music. Various time-domain features, frequency-domain features, and nonlinear features were extracted from the ECG and EEG signals. Experimental results demonstrate that the accuracy of emotion recognition and classification of signals captured by the proposed device can reach 92.08% using the CNN model. While using the LSTM-MLP feature fusion model, the accuracy figure can be improved to 95.07%. The results of the ablation experiment indicate that the feature fusion approach does improve the accuracy of recognition. It is demonstrated that the proposed device and emotional recognition approach are effective and feasible.

Recent advance of deep learning has seen remarkable progress in compound fault diagnosis modelling for industrial robots. Nevertheless, the data scarcity of compound fault samples jeopardizes the modelling performance of deep learning algorithms. Meta learning has become an effective tool in few-shot fault diagnosis modelling. However, due to the training instability of meta learning, it is challenging to deploy advanced networks such as Transformers as the base learner due to the extremely large model size. Therefore, this study proposes a lightweight convolutional Transformers (LCT) network enhanced meta learning (Meta-LCT) method to achieve accurate compound fault diagnosis with limited compound fault samples. Specifically, the LCT is firstly designed by taking advantage of linear spatial reduction (LSR) attention and spatial pooling mechanism to achieve high computational efficiency. LCT is adopted as the base learner in the Meta-SGD algorithm, and then the meta-training is performed based on the single fault data. Subsequently, the limited compound fault samples are used in the meta testing stage to obtain a compound fault diagnosis model. An experimental study based on the real-world compound fault dataset of industrial robots is presented. The experimental results indicate that the proposed Meta-LCT can achieve the compound fault diagnosis accuracy of 81.1% when only 40 data samples in each compound fault category are available.