Mid Sweden University

Direction-of-arrival (DoA) estimation plays a critical role in applications of acoustic localization. This work proposes a closed-form algorithm for single-source one-dimensional direction of arrival estimation using the uniform circular array. The algorithm estimates the azimuthal direction of arrival using the least-square estimate of the circular harmonics steering vector. The least-square (LS) estimate of the steering vector is estimated as the principal eigenvector of the received data covariance matrix. The proposed algorithm is tested and compare against the time-frequency circular harmonics beamforming (TF-CHB) and eigenbeam MUSIC (EB-MUSIC) algorithms. The comparison shows that the proposed algorithm outperformed the two benchmark methods in accuracy under various test conditions. 

Deep learning (DL) models are increasingly used in industrial applications for precise object measurement, particularly in resource-constrained environments where monocular cameras are favored for their cost-effectiveness. However, the traditional evaluation metrics, such as mean Average Precision (mAP) and validation loss, cannot accurately capture measurement task's accuracy and reliability. This study covers the use of Mean Absolute Error (MAE) and Standard Deviation (Std Dev) as specialized metrics to evaluate the trueness and consistency of DL models in measurement applications. We trained two DL based frameworks with different sets of hyperparameter configurations and assessed their performance using a dataset of images captured under identical conditions. The results show no significant correlation between mAP and the proposed metrics of MAE and Std Dev, further indicating the deficiency of the conventional metrics for measurement quality assessment. Instead, a positive linear relationship between MAE and Std Dev was recorded. Our analysis shows a strong correlation between MAE and Std Dev (0.92 for Detectron2 and 0.93 for YOLO). The random forest algorithm confirmed MAE and Std Dev as the most important feature (0.89 for Detectron2, 0.78 for YOLO), while validation loss and mAP had lower significance in feature importance and correlation. This work highlights the importance of incorporating statistical metrics into evaluating DL models to ensure the selection of configurations that deliver reliable, accurate, and efficient camera-based measurements. 

Heart sound classification is a critical task in automated cardiac diagnostics, yet it is often challenged by the limited availability of labeled data and the dominance of low-frequency components in heart sound signals. This study introduces a novel data augmentation technique, the first-difference method, to address these challenges in convolutional neural network (CNN)- based classification. By enhancing high-frequency components in the time domain, this method enables the model to better capture abnormalities, such as murmurs, present in higher frequency ranges. The effectiveness of this approach was evaluated using three spectral transformations—linear spectrogram, mel-spectrogram, and mel-frequency cepstrum coefficient (MFCC) —across multiple augmentation levels. The results demonstrate substantial improvements in classification metrics, including precision, recall, F1 score, and specificity, with MFCC-based predictors achieving the highest performance gains. The findings highlight the potential of the first-difference augmentation as a simple and effective strategy for improving heart sound classification, paving the way for more robust and generalizable diagnostic tools in real-world clinical applications.

The autoencoder represents an important Artificial Neural Network architecture designed to learn data representations in an unsupervised manner. Its structure, consisting of an encoder and a decoder, allows information to be compressed into a reduced-dimensional latent space and subsequently reconstructed. This process is crucial in many applications, such as dimensionality reduction, data compression, and noise removal. In addition, the autoencoder allows meaningful features to be extracted from the raw data, facilitating tasks such as image analysis and anomaly detection. The importance of this technique lies in its ability to reduce computational complexity and preserve essential information. However, autoencoders, used to compress and reconstruct signals, can exhibit significant variations in reconstruction quality, especially in the presence of noise or anomalies. Therefore, reconstructing the uncertainty band of the input signal to the autoencoder allows for a more accurate assessment of the quality of the reconstructed signal. This paper presents a methodology based on the law of propagation of uncertainty for reconstructing the input uncertainty band to increase the performance of an autoencoder. 

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.

Direction of arrival estimation using the spherical microphone array usually requires a search in the whole 3-dimensional space, hence computationally demanding. This work presents a machine learning approach to sectorizing the 3-dimensional space, as an intermediate step for direction-of-arrival estimation using spherical microphone array. A new feature based on the outer product of spherical harmonic vectors was proposed for the classification. This spherical harmonic matrix nominally offers lower dimensionality compared to the commonly used covariance matrix of received data. The dimension of the input matrix was further reduced using the neighborhood component analysis. The extracted features were then used to train a support vector machine (SVM), 2-layer multilayer perceptron (MLP) and a convolutional neural network (CNN) for classification purposes. The results show that the models were able to classify the spherical sector with up to 90 % accuracy for all models and number of sectors under consideration. Also, the MLP and CNN trained with simulated samples were able to accurately classify samples from real data that were not included in training samples.

The field of Metrology has seen great use of Machine Learning and Deep Learning models, improving existing Metrology and enabling measurements and estimations that were previously not possible. In the challenging task of gathering training data in various areas of Metrology, a question arises; is it necessary to gather a completely new dataset every time a quality upgrade is done to a measurement system, method or model, or can the formerly trained model be used for new data with higher Signal-to-Noise Ratio (SNR)? This paper investigates how trained neural networks react to new data coming into the testing, with a higher SNR than the training data. In the experiments, Convolutional Neural Networks (CNN), in 1D and 2D, are used on heart sound data, as a test case. The initial results show that the classification accuracy for the new data, with a higher SNR, coming into the 1D CNN is almost as high as if the network had been trained on the higher SNR data. For a 2D CNN working with spectrograms instead of time series data, the change in accuracy is not nearly as high, as the 2D CNN model seems more robust to noise differences. 

The increase in Cervical Spondylosis cases and the expansion of the affected demographic to younger patients have escalated the demand for X-ray screening. Challenges include variability in imaging technology, differences in equipment specifications, and the diverse experience levels of clinicians, which collectively hinder diagnostic accuracy. In response, a deep learning approach utilizing a ResNet-34 convolutional neural network has been developed. This model, trained on a comprehensive dataset of 1235 cervical spine X-ray images representing a wide range of projection angles, aims to mitigate these issues by providing a robust tool for diagnosis. Validation of the model was performed on an independent set of 136 X-ray images, also varied in projection angles, to ensure its efficacy across diverse clinical scenarios. The model achieved a classification accuracy of 89.7%, significantly outperforming the traditional manual diagnostic approach, which has an accuracy of 68.3%. This advancement demonstrates the viability of deep learning models to not only complement but enhance the diagnostic capabilities of clinicians in identifying Cervical Spondylosis, offering a promising avenue for improving diagnostic accuracy and efficiency in clinical settings.

One of the main challenges in Artificial Neural Networks (ANNs) is the development of reliable, valid, and reproducible systems. Prediction networks have had a disruptive impact, bringing numerous advantages in various fields, but for their common usage, it's necessary to quantify their quality. In particular, evaluating the uncertainty of the measurements obtained with these approaches allows their correct utilization. This work aims to analyze the covariances of the inputs of different neurons, particularly in those of the hidden layers of ANNs. Evaluating the covariance of the inputs of a single neuron finds primary use in the law of propagation of uncertainty, particularly for evaluating the correlation term in mathematical development, as defined by ISO GUM. Based on numerical evaluation, the proposed procedure aims to evaluate the PDFs of inputs to individual nodes and, therefore, the correlations among all inputs propagating within the network architecture. 

Snow accumulations, especially if of great intensity, as is the case in northern countries, for example, can be very damaging, especially if they occur in urban environments. The damage provoked by snow is not only related to the weight of the accumulations, causing damage to structures but also to the pollution retained by the structure of the snowflakes. However, snow weight monitoring is a complex task, both because of the placement of the sensors and the specific operating ranges they must have in terms of operating temperature. These complications can be overcome by the design and use of a soft sensor, that is, a sensor capable of making indirect measurements from other parameters related to the measurement under consideration. This paper presents the design and metrological validation of a soft sensor for indirect weight measurement of snow accumulations. The designed soft sensor has been based on Artificial Neural Network and achieved, as a result, a Root-Mean-Square Error (RMSE) of 114g and a maximum extended uncertainty, evaluated by Monte Carlo simulation, of 300g in a measurement range from 150g to 5200g.