Mid Sweden University

The importance of preventing failures in rotating machines has led to extensive research on diagnosis and prognosis methods based on vibration data. To achieve a high generalizability for these solutions, they need to handle high discrepancies between the data used for developing the method and the data from the target machine, a lack of historical events from the target machine and multi-component fault scenarios. Currently, state-of-the-art research has successfully found solutions targeting the first two challenges. However, these almost exclusively focus on single-component fault scenarios, meaning it is unclear how a method should be constructed considering a multi-component system. Therefore, this study constructs a framework for multi-component fault scenarios called Rotating Machinery Prognostic Framework (RoMaP) that leverages the advancements in anomaly detection, fault diagnosis and RUL prediction based on vibration data. To evaluate the potential of RoMaP, an instance based on state-of-the-art research was implemented and compared to other methods based on alternative frameworks. The results show that the instance of RoMaP had the best performance across many scenarios expected in an industrial environment, suggesting that it is a suitable approach for monitoring the health of rotating machines. 

Accurate distance estimation is crucial for applications such as robotics and autonomous systems, where reliable measurements are needed for navigation and interaction with the environment. ArUco markers are a robust solution for distance estimation, offering precise measurements based on their geometric properties. In this paper, we, first, systematically investigate the impact of various contextual factors in an indoor environment, such as angle of observation and illumination, on the systematic error and uncertainty of distance measurements. Our experiments show that illumination (exposure) significantly influence the performance of distance estimation systems, introducing notable systematic errors in real-world settings. Second, we propose a machine learning-based approach, using a neural network, to address the challenge posed by these factors and improve the systematic error of distance estimation. Our results in dynamic outdoor environment, demonstrate the effectiveness of this approach, significantly enhancing the systematic error of distance under dynamic environmental conditions. 

Automated advertisement detection in newspapers is a challenging task due to the diversity in print layouts, formats, and design styles. This task has critical applications in media monitoring, content analysis, and advertising analytics. To address these challenges, we introduce AdVision, a deep-learning-based solution that treats advertisements as unique visual objects. We provide a comparative study of various detection architectures, including one-stage, two-stage, and transformer-based detectors, to identify the most effective approach for detecting advertisements. Our results are validated through extensive experiments conducted under different conditions and metrics. Newspapers from four different countries - Denmark, Norway, Sweden, and the UK - were selected to demonstrate the variety of languages and print formats. Additionally, we conduct a cross-analysis to show how training on one language can generalize to another. To enhance the explainability of our results, we employ GradCAM++ (Chattopadhay et al., 2018) heatmaps. Our experiments demonstrate that the YOLOv8 model achieves superior performance, balancing high precision and recall with minimal inference latency, making it particularly suitable for high-throughput advertisement detection.

The demand for high-quality LiDAR datasets is increasing as LiDAR technology is being used in various applications, including autonomous vehicles, robotics, and 3D mapping. However, generating accurate ground truth data for LiDAR datasets remains a challenge due to issues like multipath interference (MPI) and other disturbances. The method for generating high-quality ground truth data for LiDAR applications based on cross-section is introduced is this paper. The key concept is based on precisely registering a digital twin, CAD-based which is later converted to a mesh, with LiDAR depth images. By utilizing selective point usage in its cross-sections, the method demonstrates greater robustness to MPI compared to standard approaches. The technique is evaluated using a dataset of LiDAR-acquired point clouds of a living room scene. The results show that the proposed technique achieves significantly better accuracy in affected regions than standard Iterative Closest Point (ICP) based methods. Additionally, the paper proposes a new set of metrics for evaluating the quality of ground truth data, which is more robust to MPI than standard metrics such as RMSE and Chamfer Distance. The proposed technique is a valuable tool for generating large-scale, high-quality datasets for LiDAR applications. Lastly we compared our dataset with latest available datasets. 

Recent advancements have explored parallel partitioning of Transformers and Convolutional Neural Network (CNN) based models across networks of edge devices to accelerate deep neural network (DNN) inference. However, partitioning strategies for hybrid Vision Transformers-models integrating convolutional and attention layers-remain underdeveloped, particularly in scenarios with low communication data rates. This work introduces a novel partitioning scheme tailored for hybrid Vision Transformers, addressing communication latency through efficient compressed communication and model size reduction. The proposed approach incorporates a trainable quantization and JPEG compression pipeline to minimize overhead. We evaluate our scheme on two state-of-the-art architectures, edgeViT and CoatNet. For a communication data rate of 10 MB/s and partitioning across 12 devices, we achieve up to a 1.74x speed-up and a 5.34x model size reduction for edgeViT-XXS. Similarly, on a customized CoatNet-0, our method achieves a 1.40x speed-up and a 2.66x reduction in model size, demonstrating the efficacy of the approach in real-world scenarios.

The localization and identification of advertisements play a pivotal role in content analysis and information retrieval. Acknowledging this significance, this paper focuses on the critical task of model evaluation and selection for robust and efficient advertisement detection. Employing a comprehensive methodology, we assess various deep learning models based on their accuracy, efficiency, and reliability in detecting and localizing advertisements within diverse print media formats. Our study reveals that certain models significantly outperform others in terms of mean Average Precision (mAP) and F1 scores, while also maintaining low inference latencies. These findings have profound implications for the development of more effective and efficient advertisement detection systems in print media. The conclusions drawn from our research provide valuable insights for future advancements in digital advertising and media analytics. 

We enhance Convolutional Neural Network (CNN) deployment through optimized model partitioning for efficient implementation on resource-constrained IoT devices. Our strategy involves (a) a novel trainable activation sharing function and (b) combining lossy and lossless compression at the partitioning point. Using an autoencoder, trained via Knowledge Distillation, with activation sharing and LZMA compression, we maximize intermediate feature map compression and accelerate IoT-Server communication. Experiments on ResNet50 and YOLOv10n, implemented on a Raspberry Pi 4 with LTE-C4 and WiFi(ah), demonstrate up to 7× speed-up compared to full in-node processing, with less than 1% accuracy drop and 99% IoT-model size reduction via 3-bit quantization (ResNet50 on Tiny-ImageNet-200). We analyze latency across key processes, highlighting significant improvements in IoT-node latency and performance preservation. Our approach enhances existing methods in model partitioning and quantization, offering solutions for reducing latency on the IoT-node while maintaining performance. 

This paper introduces an enhanced approach for deploying deep learning models on resource-constrained IoT devices by combining model partitioning, autoencoder-based compression, quantization with Time Dependent Clustering Loss (TCL) regularization, and lossless compression, to reduce communication overhead, minimizing latency while maintaining accuracy. The autoencoder compresses feature maps at the partitioning point before quantization, effectively reducing data size and preserving accuracy. TCL regularization clusters activations at the partitioning point to align with quantization levels, minimizing quantization error and ensuring accuracy even with extreme low-bitwidth quantization. Our method is evaluated on classification models (ResNet-50, EfficientNetV2-S) and an object detection model (YOLOv10n) using the TinyImageNet-200 and Pascal VOC datasets. Deployed on Raspberry Pi 4 B and GPU, each model is tested across various partitioning points, quantization bit-widths (1-bit, 2-bit, and 3-bit), communication datarate (1MB/s to 10MB/s), and LZMA lossless compression. For a partitioned ResNet-50 after the convolutional stem block, the speed-up against a server solution is 2.33× and 1.85x compared to the all-in-node solution, with only a minimal accuracy drop of less than one percentage points. The proposed framework offers a scalable solution for deploying high-performance AI models on IoT devices, extending the feasibility of real-time inference in resource-constrained environments. 

The accurate classification of plant growth stages from image sequences is a challenging problem due to the gradual and continuous nature of plant development. Significant variability exists within the same category of growth, and temporal boundaries between successive categories often exhibit visual similarities, complicating the classification task. Whereas the existing approaches are limited and rely on the analysis of single frames, they do not take temporal information embedded in sequential data, reducing robustness and accuracy. They also show vulnerability to intra-category variability among the images and also to inter-class overlap at the temporal boundaries. To address these challenges, we propose a novel approach, PSO-SqueezeTempVote, for plant temporal state categorisation that combines PSO for hyperparameter optimisation of the SqueezeNet architecture with Temporal Image Sequence Fusion (TISF) for temporal ensemble learning. The PSO part effectively optimises some critical hyperparameters, hence improving the performance of the model on individual frames. Meanwhile, the component of TISF exploits temporal continuity using a sequence of images' aggregation through majority voting. This enables classifying the segment of an image much more precisely than the previously done approaches. Results on our datasets confirm that the PSOSqueezeTempVot has improved the state-of-the-art techniques by almost 13%, out of which approximately 6% was contributed by TISF itself. 

The rapid progress in sensor technology and computational capabilities has significantly improved real-time data collection, enabling precise monitoring of various phenomena and industrial processes. However, the volume and complexity of heterogeneous data present substantial processing challenges. Traditional data-processing techniques, such as data aggregation, filtering, and statistical analysis, are increasingly supplemented by data fusion methods. These methods can be broadly categorised into traditional analytics-based approaches, like the Kalman Filter and Particle Filter, and learning-based approaches, utilising machine learning and deep learning techniques such as Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN). These techniques combine data from multiple sources to provide a comprehensive and accurate representation of information, which is critical in number of fields. Despite this, a comprehensive review of learning-based, particularly deep learning-based, data fusion strategies is lacking. This paper presents a thorough review of deep learning-based data fusion methodologies across various fields, examining their evolution over the past five years. It highlights applications in remote sensing, healthcare, industrial fault diagnosis, intelligent transportation, and other domains. The paper categories fusion strategies into early-level, intermediate-level, late-level, and hybrid fusion, emphasising their synergies, challenges, and suitability. It outlines significant advancements, the comparative advantages of deep learning-based methods over traditional approaches, and emerging trends and future directions. To ensure a comprehensive analysis, the review is structured using the ProKnow-C methodology, a rigorous selection process that focuses on relevant literature from recent years.