Full-penetration laser welding (FPLW) is increasingly adopted in manufacturing pipelines, yet its industrial scalability is constrained by in-process defect formation, particularly incomplete penetration. To address this, we propose a sensor-driven framework for non-destructive monitoring and automated defect detection that uses infrared (IR) thermography as the primary in situ sensing modality and applies deep learning to the acquired thermal signals. High-speed IR camera recordings were processed to track spatter and the weld zone, yielding a time series of physically interpretable spatiotemporal features (mean spatter area, mean spatter temperature, number of spatters, and mean welding zone temperature). Defect recognition is formulated as a multi-label classification problem targeting incomplete penetration, sagging, shrinkage groove, and linear misalignment, and multiple temporal models were evaluated on the same sensor-derived feature sequences. Experimental validation on 09G2S pipeline steel demonstrates that the proposed time series pipeline based on a hybrid CNN–transformer achieves a mean Average Precision (mAP) of 0.85 while preserving near-real-time inference on a CPU. The results indicate that IR thermography-based spatter dynamics provide actionable sensing signatures for automated defect prediction and can serve as a foundation for closed-loop quality control in industrial laser pipeline welding.
The paper proposes a method for estimating the state vector of an agent in a multi-agent biological system based on noisy measurements using recurrent filters. It addresses the issues of scalability in existing approaches to monitoring the behavior of laboratory rodents and the absence of a unified mathematical framework. A mathematical description of an agent in the biological system is provided, along with the formulation of the task of estimating its state. The mathematical model is built upon a non-linear discrete-time system in state space. The solution to this problem is demonstrated using the example of skeletal keypoints in a Wistar rat, which are detected using a pre-trained detector. A fully connected neural network is proposed to parameterize the unknown dynamics of the system. The particle filter (a sequential Monte Carlo method) and the unscented Kalman filter were selected for a comparative analysis. The comparison of the methods was conducted on a collected and preprocessed dataset comprising images with a resolution of 1060×548 pixels and annotations of rat skeletal keypoints. The experimental results demonstrate the high efficacy of the proposed method and its advantage over an analytical description of the system's nonlinear dynamics. Among the compared methods, the dual estimation of both the state vector and the neural network parameters using two unscented Kalman filters achieved the minimal mean error of 6.4 pixels. However, for practical applications in real-time scenarios, a single filter employing a pre-trained neural network proves to be more advantageous. Moreover, the unscented Kalman filter in this case demonstrated higher accuracy than the particle filter (mean error of 8.1 pixels vs. 12.0 pixels). The results of this study can be used to solve the task of automated registration of Wistar rat behavior by parameterizing the functions that link state vectors with the output vectors of individual and group behavior.
24 April 2025 in Scientific and Technical Journal of Information Technologies, Mechanics and Optics
The automation of action recognition in laboratory animals is a crucial step in simplifying behavioral tests in the fields of pathophysiology and rehabilitation research. The most common method of action recognition is to analyze the trajectories of key skeletal points. However, the existing methods are strongly tied to the specific animal species, selected skeletal points, and set of activities to be recognized. Furthermore, there is a dearth of mathematical formulations of this problem and research on algorithms for filtering obtained trajectories. The research task involves the collection of a dataset for key points detection of Wistar rats and evaluation of algorithms for filtering trajectories from noisy measurements. In considered skeletal model of the rat, a total of thirteen points were selected for the purpose of estimating the behavior along trajectories. A mathematical description of the dynamics of point movement between frames for use in a Kalman filter is provided. Four filtering algorithms are evaluated in terms of accuracy and curve smoothness. The technique of constructing the covariance matrix of the detector noise by analyzing the key point detection errors is developed. The comparison of filtering algorithms shows that the Unscented Kalman filter with nonlinear model and moving average filter yield the most optimal results in this task. The findings of this study allow the use of a mathematical description of system dynamics to estimate the actual trajectory from noisy measurements. Furthermore, the described methodologies are not exclusive to laboratory animals, but can also be applied to human subjects.
31 March 2026 in Journal of Biomedical Photonics & Engineering
Computer-vision methods have been applied to automated behavior recognition in laboratory rodents. Namely, we explored the possibility of classifying certain behavior classes from still images; compared keypoint-based methods with approaches based on visual embeddings; studied the feasibility of transferring models between rats and mice; and evaluated the relevance of the number and the accuracy of the detected keypoints. We collected a dataset of a freely moving Wistar rat to train six pose-based classifiers with Long Short-Term Memory (LSTM) using six sets of keypoints produced by two detectors and four Convolutional Neural Network (CNN) classifiers using images and optical flow frames. The results demonstrated the highest mean average precision (mAP) of 65.7% for the CNN-based methods and 34.3% for the LSTM classifiers, the feasibility of recognizing visually distinct classes (rearing and body grooming) from still images, and the applicability of a keypoint detector trained on mice. The results of this study can be applied to the design of a computer vision system for automating long-term monitoring of laboratory rodent behavior