Although the development of Internet of Things (IoT) provides a significant boost for the applications of deep learning algorithms, it is generally hard to fully implement the deep learning algorithms by IoT devices due to their limited calculation capacity. The problem could be alleviated by deploying the deep learning algorithms with edge computing. Herein, in this article, we propose a kind of distributed long short-term memory (DLSTM) neural networks and deploy them on the IoT environment to handle the large-scale spatiotemporal correlation regression tasks. Specifically, the presented DLSTM neural networks adopt the collaborative computing architecture with the terminals, edges and cloud, in order to realize the lightweight deep learning on the IoT devices and improve the learning efficiency. The generalization ability of LSTM neural networks is promoted through introducing the distributed memory cells to implement the information sharing between different edge servers and employing the attention mechanism in LSTM neural networks. Meanwhile, the deep fully connected networks are deployed among the cloud to extract the spatiotemporal correlations in the variety of data from different time and space regions, which enhances the transferability of LSTM neural networks. Numerical experiment shows that the proposed DLSTM neural networks reduce 36% model parameters size, 32% memory consumption among the cloud, and more than half of the prediction errors compared with traditional LSTM neural networks. Besides, the favourable transferability of the present DLSTM neural networks is also verified via numerical experiment.

In the context of Industry 4.0, autonomous and data-driven manufacturing processes are advancing rapidly, with wire arc additive manufacturing (WAAM) emerging as a promising technique for producing large-scale metal components. Ensuring quality control and part traceability in WAAM remains an area of active research, as existing process monitoring systems often require operator intervention and are tailored to specific machine setups and camera configurations, limiting adaptability across industrial environments. This study addresses these challenges by developing an angle-invariant melt pool analysis pipeline capable of recognising bead features in wire-based directed energy deposition from monitoring images captured using various camera qualities, positions, and angles. A new benchmark dataset, WAAM-ViD, is also introduced to support future research. The proposed pipeline integrates two deep learning models: DeepLabv3, fine-tuned through active learning for precise melt pool segmentation (Dice similarity coefficient of 95.90%), and WAAM-ViDNet, a regression-based multimodal model that predicts melt pool width using the segmented images and camera calibration data, achieving 88.71% accuracy. The results demonstrate the pipeline’s effectiveness in enabling real-time process monitoring and control in WAAM, representing a step toward fully autonomous and adaptable additive manufacturing systems.

Part quality monitoring and control in wire-based directed energy deposition additive manufacturing (w-DEDAM) processes has been garnering continuous interest from both the academic and industrial sectors. However, maintaining a consistent layer height and ensuring that the wall height aligns closely with the design, as depicted in computer-aided design (CAD) models, pose significant challenges. These challenges arise due to the uncertainties associated with the manufacturing process and the working environment, particularly with extended processing times. To achieve these goals in an industrial scenario, the deposition geometry must be measured with precision and efficiency throughout the part-building process. Moreover, it is essential to comprehend the changes in the interlayer deposition height based on various process parameters. This paper first examines the behaviour of interlayer deposition height when process parameters change within different wall regions, with a particular focus on the transition areas. In addition, this paper explores the potential of geometry monitoring information in implementing interlayer wall height compensation during w-DEDAM part-building. The in-process layer height was monitored using a coherent range-resolved interferometry (RRI) sensor, and the accuracy and efficiency of this measurement were carefully studied. Leveraging this information and understanding of deposition geometry, the control points of the process parameters were identified. Subsequently, appropriate and varied process parameters were applied to each wall region to gradually compensate for wall height. The wall height discrepancies were generally compensated for in two to three layers.

In fire emergency management, a delayed execution will cause a significant number of casualties. Conventional fire drills typically only identify a certain percentage of evacuation bottlenecks after the building has been constructed, which is hard to improve. This paper proposes an innovative framework to validate fire emergency evacuation at the early design stage. According to the experience and knowledge of fire emergency evacuation design, the proposed framework also introduces a seamless two-way information channel to embed fire emergency evacuation simulations into a BIM-based design environment. Several critical factors for fire evacuation have been reviewed in relevant domain knowledge, which is used to build virtual characters to test in experimental scenarios. The results are analyzed to validate fire emergency evacuation factors, and the feedback knowledge is stored as a knowledge model for further applications.

Process control and quality assurance of wire + arc additive manufacturing (WAAM) and automated welding rely heavily on in-process monitoring videos to quantify variables such as melt pool geometry, location and size of droplet transfer, arc characteristics, etc. To enable feedback control based upon this information, an automatic and robust segmentation method for monitoring of videos and images is required. However, video segmentation in WAAM and welding is challenging due to constantly fluctuating arc brightness, which varies with deposition and welding configurations. Additionally, conventional computer vision algorithms based on greyscale value and gradient lack flexibility and robustness in this scenario. Deep learning offers a promising approach to WAAM video segmentation; however, the prohibitive time and cost associated with creating a well-labelled, suitably sized dataset have hindered its widespread adoption. The emergence of large computer vision models, however, has provided new solutions. In this study a semi-automatic annotation tool for WAAM videos was developed based upon the computer vision foundation model SAM and the video object tracking model XMem. The tool can enable annotation of the video frames hundreds of times faster than traditional manual annotation methods, thus making it possible to achieve rapid quantitative analysis of WAAM and welding videos with minimal user intervention. To demonstrate the effectiveness of the tool, three cases are demonstrated: online wire position closed-loop control, droplet transfer behaviour analysis, and assembling a dataset for dedicated deep learning segmentation models. This work provides a broader perspective on how to exploit large models in WAAM and weld deposits.