Dalarna University's logo and link to the university's website

This study explores how attention mechanisms impact representation distributions within neural networks, focusing on catastrophic forgetting and robustness to input noise. We compare Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), their attention-enhanced counterparts (RNNA, LSTMA, GRUA), and the Transformer model using musical sequences from "Daisy Bell". A key finding is the difference in how these models distribute the information in their representation. Base models like RNN, LSTM, and GRU concentrate information within specific nodes, while attention-enhanced models spread information across more nodes, demonstrating greater robustness to input noise. This is shown by significant differences in performance deterioration between base models and their attention-augmented versions. However, base models such as RNN and GRU exhibit better resistance to catastrophic forgetting compared to their attention-enhanced counterparts. Despite this, attention models show a positive correlation between higher overlap percentages in their representations and improved accuracy for certain tasks, alongside a negative correlation with higher numbers of empty nodes. The Transformer model stands out by maintaining high accuracy across tasks, likely due to its self-attention mechanisms. These results suggest that while attention mechanisms enhance robustness to noise, further research is needed to address catastrophic forgetting in neural networks.

The overall aim of this thesis was to develop and evaluate new knowledge- and data-driven methods for supporting treatment and providing information for better assessment of Parkinson’s disease (PD).

PD is complex and progressive. There is a large amount of inter- and intravariability in motor symptoms of patients with PD (PwPD). The current evaluation of motor symptoms that are done at clinics by using clinical rating scales is limited and provides only part of the health status of PwPD. An accurate and clinically approved assessment of PD is required using frequent evaluation of symptoms.

To investigate the problem areas, the thesis adopted the microdata analysis approach including the stages of data collection, data processing, data analysis, and data interpretation. Sensor systems including smartphone and tri-axial motion sensors were used to collect data from advanced PwPD experimenting with repeated tests during a day. The experiments were rated by clinical experts. The data from sensors and the clinical evaluations were processed and used in subsequent analysis.

The first three papers in this thesis report the results from the investigation, verification, and development of knowledge- and data-driven methods for quantifying the dexterity in PD. The smartphone-based data collected from spiral drawing and alternate tapping tests were used for the analysis. The results from the development of a smartphone-based data-driven method can be used for measuring treatment-related changes in PwPD. Results from investigation and verification of an approximate entropy-based method showed good responsiveness and test-retest reliability indicating that this method is useful in measuring upper limb temporal irregularity.

The next two papers, report the results from the investigation and development of motion sensor-based knowledge- and data-driven methods for quantification of the motor states in PD. The motion data were collected from experiments such as leg agility, walking, and rapid alternating movements of hands. High convergence validity resulted from using motion sensors during leg agility tests. The results of the fusion of sensor data gathered during multiple motor tests were promising and led to valid, reliable and responsive objective measures of PD motor symptoms.

Results in the last paper investigating the feasibility of using the Dynamic Time-Warping method for assessment of PD motor states showed it is feasible to use this method for extracting features to be used in automatic scoring of PD motor states.

The findings from the knowledge- and data-driven methodology in this thesis can be used in the development of systems for follow up of the effects of treatment and individualized treatments in PD.

This thesis consists of four research papers presenting a microdata analysis approach to assess and evaluate the Parkinson’s disease (PD) motor symptoms using smartphone-based systems. PD is a progressive neurological disorder that is characterized by motor symptoms. It is a complex disease that requires continuous monitoring and multidimensional symptom analysis. Both patients’ perception regarding common symptom and their motor function need to be related to the repeated and time-stamped assessment; with this, the full extent of patient’s condition could be revealed. The smartphone enables and facilitates the remote, long-term and repeated assessment of PD symptoms. Two types of collected data from smartphone were used, one during a three year, and another during one-day clinical study. The data were collected from series of tests consisting of tapping and spiral motor tests. During the second time scale data collection, along smartphone-based measurements patients were video recorded while performing standardized motor tasks according to Unified Parkinson’s disease rating scales (UPDRS).

At first, the objective of this thesis was to elaborate the state of the art, sensor systems, and measures that were used to detect, assess and quantify the four cardinal and dyskinetic motor symptoms. This was done through a review study. The review showed that smartphones as the new generation of sensing devices are preferred since they are considered as part of patients’ daily accessories, they are available and they include high-resolution activity data. Smartphones can capture important measures such as forces, acceleration and radial displacements that are useful for assessing PD motor symptoms.

Through the obtained insights from the review study, the second objective of this thesis was to investigate whether a combination of tapping and spiral drawing tests could be useful to quantify dexterity in PD. More specifically, the aim was to develop data-driven methods to quantify and characterize dexterity in PD. The results from this study showed that tapping and spiral drawing tests that were collected by smartphone can detect movements reasonably well related to under- and over-medication.

The thesis continued by developing an Approximate Entropy (ApEn)-based method, which aimed to measure the amount of temporal irregularity during spiral drawing tests. One of the disabilities associated with PD is the impaired ability to accurately time movements. The increase in timing variability among patients when compared to healthy subjects, suggests that the Basal Ganglia (BG) has a role in interval timing. ApEn method was used to measure temporal irregularity score (TIS) which could significantly differentiate the healthy subjects and patients at different stages of the disease. This method was compared to two other methods which were used to measure the overall drawing impairment and shakiness. TIS had better reliability and responsiveness compared to the other methods. However, in contrast to other methods, the mean scores of the ApEn-based method improved significantly during a 3-year clinical study, indicating a possible impact of pathological BG oscillations in temporal control during spiral drawing tasks. In addition, due to the data collection scheme, the study was limited to have no gold standard for validating the TIS. However, the study continued to further investigate the findings using another screen resolution, new dataset, new patient groups, and for shorter term measurements. The new dataset included the clinical assessments of patients while they performed tests according to UPDRS. The results of this study confirmed the findings in the previous study. Further investigation when assessing the correlation of TIS to clinical ratings showed the amount of temporal irregularity present in the spiral drawing cannot be detected during clinical assessment since TIS is an upper limb high frequency-based measure. 

Title: Objective assessment of Parkinson’s disease motor symptoms during leg agility test using motion sensors

Objective: To develop and evaluate machine learning methods for assessment of Parkinson’s disease (PD) motor symptoms using leg agility (LA) data collected with motion sensors during a single dose experiment.

Background: Nineteen advanced PD patients (Gender: 14 males and 5 females, mean age: 71.4, mean years with PD: 9.7, mean years with levodopa: 9.5) were recruited in a single center, open label, single dose clinical trial in Sweden [1].

Methods: The patients performed up to 15 LA tasks while wearing motions sensors on their foot ankle. They performed tests at pre-defined time points starting from baseline, at the time they received a morning dose (150% of their levodopa equivalent morning dose), and at follow-up time points until the medication wore off. The patients were video recorded while performing the motor tasks. and three movement disorder experts rated the observed motor symptoms using 4 items from the Unified PD Rating Scale (UPDRS) motor section including UPDRS #26 (leg agility), UPDRS #27 (Arising from chair), UPDRS #29 (Gait), UPDRS #31 (Body Bradykinesia and Hypokinesia), and dyskinesia scale. In addition, they rated the overall mobility of the patients using Treatment Response Scale (TRS), ranging from -3 (very off) to 3 (very dyskinetic). Sensors data were processed and their quantitative measures were used to develop machine learning methods, which mapped them to the mean ratings of the three raters. The quality of measurements of the machine learning methods was assessed by convergence validity, test-retest reliability and sensitivity to treatment.

Results: Results from the 10-fold cross validation showed good convergent validity of the machine learning methods (Support Vector Machines, SVM) with correlation coefficients of 0.81 for TRS, 0.78 for UPDRS #26, 0.69 for UPDRS #27, 0.78 for UPDRS #29, 0.83 for UPDRS #31, and 0.67 for dyskinesia scale (P<0.001). There were good correlations between scores produced by the methods during the first (baseline) and second tests with coefficients ranging from 0.58 to 0.96, indicating good test-retest reliability. The machine learning methods had lower sensitivity than mean clinical ratings (Figure. 1).

Conclusions: The presented methodology was able to assess motor symptoms in PD well, comparable to movement disorder experts. The leg agility test did not reflect treatment related changes.

Title: Feasibility of a multi-sensor data fusion method for assessment of Parkinson’s disease motor symptoms

Objective: To assess the feasibility of measuring Parkinson’s disease (PD) motor symptoms with a multi-sensor data fusion method. More specifically, the aim is to assess validity, reliability and sensitivity to treatment of the methods.

Background: Data from 19 advanced PD patients (Gender: 14 males and 5 females, mean age: 71.4, mean years with PD: 9.7, mean years with levodopa: 9.5) were collected in a single center, open label, single dose clinical trial in Sweden [1].

Methods: The patients performed leg agility and 2-5 meter straight walking tests while wearing motion sensors on their limbs. They performed the tests at baseline, at the time they received the morning dose, and at pre-specified time points until the medication wore off. While performing the tests the patients were video recorded. The videos were observed by three movement disorder specialists who rated the symptoms using a treatment response scale (TRS), ranging from -3 (very off) to 3 (very dyskinetic). The sensor data consisted of lower limb data during leg agility, upper limb data during walking, and lower limb data during walking. Time series analysis was performed on the raw sensor data extracted from 17 patients to derive a set of quantitative measures, which were then used during machine learning to be mapped to mean ratings of the three raters on the TRS scale. Combinations of data were tested during the machine learning procedure.

Results: Using data from both tests, the Support Vector Machines (SVM) could predict the motor states of the patients on the TRS scale with a good agreement in relation to the mean ratings of the three raters (correlation coefficient = 0.92, root mean square error = 0.42, p<0.001). Additionally, there was good test-retest reliability of the SVM scores during baseline and second tests with intraclass-correlation coefficient of 0.84. Sensitivity to treatment for SVM was good (Figure 1), indicating its ability to detect changes in motor symptoms. The upper limb data during walking was more informative than lower limb data during walking since SVMs had higher correlation coefficient to mean ratings.  

Conclusions: The methodology demonstrates good validity, reliability, and sensitivity to treatment. This indicates that it could be useful for individualized optimization of treatments among PD patients, leading to an improvement in health-related quality of life.

Objectives: The aim of this paper is to investigate whether a smartphone-based system can be used to quantify dexterity in Parkinson’s disease (PD). More specifically, the aim was to develop data-driven methods to quantify and characterize dexterity in PD. Methods: Nineteen advanced PD patients and 22 healthy controls participated in a clinical trial in Uppsala, Sweden. The subjects were asked to perform tapping and spiral drawing tests using a smartphone. Patients performed the tests before, and at pre-specified time points after they received 150% of their usual levodopa morning dose. Patients were video recorded and their motor symptoms were assessed by three movement disorder specialists using three Unified PD Rating Scale (UPDRS) motor items from part III, the dyskinesia scoring and the treatment response scale (TRS). The raw tapping and spiral data were processed and analyzed with time series analysis techniques to extract 37 spatiotemporal features. For each of the five scales, separate machine learning models were built and tested by using principal components of the features as predictors and mean ratings of the three specialists as target variables. Results: There were weak to moderate correlations between smartphone-based scores and mean ratings of UPDRS item #23 (0.52; finger tapping), UPDRS #25 (0.47; rapid alternating movements of hands), UPDRS #31 (0.57; body bradykinesia and hypokinesia), sum of the three UPDRS items (0.46), dyskinesia (0.64), and TRS (0.59). When assessing the test-retest reliability of the scores it was found that, in general, the clinical scores had better test-retest reliability than the smartphone-based scores. Only the smartphone-based predicted scores on the TRS and dyskinesia scales had good repeatability with intra-class correlation coefficients of 0.51 and 0.84, respectively. Clinician-based scores had higher effect sizes than smartphone-based scores indicating a better responsiveness in detecting changes in relation to treatment interventions. However, the first principal component of the 37 features was able to capture changes throughout the levodopa cycle and had trends similar to the clinical TRS and dyskinesia scales. Smartphone-based scores differed significantly between patients and healthy controls. Conclusions: Quantifying PD motor symptoms via instrumented, dexterity tests employed in a smartphone is feasible and data from such tests can also be used for measuring treatment-related changes in patients.

This paper is reviewing objective assessments of Parkinson’s disease(PD) motor symptoms, cardinal, and dyskinesia, using sensor systems. It surveys the manifestation of PD symptoms, sensors that were used for their detection, types of signals (measures) as well as their signal processing (data analysis) methods. A summary of this review’s finding is represented in a table including devices (sensors), measures and methods that were used in each reviewed motor symptom assessment study. In the gathered studies among sensors, accelerometers and touch screen devices are the most widely used to detect PD symptoms and among symptoms, bradykinesia and tremor were found to be mostly evaluated. In general, machine learning methods are potentially promising for this. PD is a complex disease that requires continuous monitoring and multidimensional symptom analysis. Combining existing technologies to develop new sensor platforms may assist in assessing the overall symptom profile more accurately to develop useful tools towards supporting better treatment process.

Job Recommendation Systems have traditionally been using machine learning and information retrieval techniques such as Word2Vec, Term Frequency – Inverse Term Frequency and Support Vector Machines. The rise of Large Language Models (LLM) could potentially bridge the shortcomings of the traditional techniques. LLMs have other shortcomings which could be mitigated using the Retrieval Augmented Generation (RAG) architecture, which grounds output of an LLM to a dataset. This thesis explores the use of LLMs and RAG in job recommendation systems. Addressing gaps in existing methods that overlook user’s skills, geographic preferences, and multilingual datasets. LLMs and RAG enhance recommendation systems by processing unstructured data and providing grounded, personalized suggestions. This thesis aims to evaluate the trustworthiness and utility of an LLM in recruitment and competence development. It examines how an LLM can enhance traditional recruitment processes and support job seekers by recommending skills needed for specific job roles and regions, based on a user’s existing skills – resulting in a curated list of skills that could be improved upon. The methodology involves using Swedish IT job market data to provide personalized skill recommendations for job seekers based on their existing skills and desired job titles for specific regions. Data is collected, cleaned, and formatted to meet system requirements, and the models’ usefulness is evaluated through questionnaires, experiments and interviews with IT-students and recruitment professionals and analyzed using thematic and quantitative analysis. The inclusion of LLMs in recruitment and competence development shows promising potential. Findings suggest LLMs can refine the recruitment process, identify skill gaps, and offer insights for both job seekers and recruiters. Both IT-students and professionals in recruitment express optimism about LLMs ability to enhance traditional methods. Respondents given examples of the model’s output showed positive reactions and valuable insights in potential.

I dag går utvecklingen av trådlösa nätverk snabbt framåt. Zigbee är en helt ny teknik som bygger på IEEE 802-15-4 standarden. Zigbee utvecklades av the Zigbee Alliance som består av en rad stora elektronikföretag. Zigbee är en teknik som inriktar sig på låg energiförbrukning och låg kostnad. Tekniken är tänkt att användas för att ställa in och läsa av sensorer av olika slag. Denna rapport är ett resultat av ett examensarbete som går ut på att utreda om Zigbee tekniken kan användas för lite längre avstånd. Arbetet resulterade i två demoapplikationer för ett enkel zigbee system, och färdigskriven kod för en möjlighet att använda Zigbee för länge avstånd.

This project aims to provide understanding the functions of a resistance and capacitance voltage divider and to design one. It is the core principle behind a voltage probe. It should have a bandwidth of 1 MHz, ratio 10:1, τ1 = τ2 and have up to 3 % deviation from the optimal value. Experiments are done to get a better understanding on how the components of the circuit interact with each other. This is done by doing simulations in the software and doing tests in the workshop. Direct current measurements are done to understand how a voltage divider operates and alternate current measurements studies how capacitors in parallel over each resistor in the voltage divider effect the system and therefore the result. It is complicated to design a resistance and capacitance voltage divider since there are many parameters that are in play and impact the result in different ways. Some of the parameters are known and can be calculated and some are parasitic parameters that are more difficult to take in consideration. A more thorough study is needed to find answers on how more of the parameters affect the result. Test number 4 in Appendix 1 exceeds the goal expectation with a bandwidth of 4.24 MHz and a tolerance of 1.2 %. 

This paper presents a novel traffic sign recognition system which can aid in the development of Intelligent Speed Adaptation. This system is based on extracting the speed limit sign from the traffic scene by Circular Hough Transform (CHT) with the aid of colour and non-colour information of the traffic sign. The digits of the speed limit sign are then extracted and classified using SVM classifier which is trained for this purpose. In general, the system detects the prohibitory traffic sign in the first place, specifies whether the detected sign is a speed limit sign, and then determines the allowed speed in case the detected sign is a speed limit sign. The SVM classifier was trained with 270 images which were collected in different light conditions. To check the robustness of this system, it was tested against 210 images which contain 213 speed limit traffic sign and 288 Non-Speed limit signs. It was found that the accuracy of recognition was 98% which indicates clearly the high robustness targeted by this system.

Fuel consumption (FC) is one of the key factors indetermining expenses of operating a heavy-duty vehicle. A customermay therefore request an estimate of the FC of a desired vehicle.The modular design of heavy-duty vehicles allows their constructionby specifying the building blocks, such as gear box, engine andchassis type. If the combination of building blocks is unprecedented,it is unfeasible to measure the FC, since this would first r equire theconstruction of the vehicle. This paper proposes a machine learningapproach to predict FC. This study uses around 40,000 vehiclesspecific a nd o perational e nvironmental c onditions i nformation, suchas road slopes and driver profiles. A ll v ehicles h ave d iesel enginesand a mileage of more than 20,000 km. The data is used to investigatethe accuracy of machine learning algorithms Linear regression (LR),K-nearest neighbor (KNN) and Artificial n eural n etworks ( ANN) inpredicting fuel consumption for heavy-duty vehicles. Performance ofthe algorithms is evaluated by reporting the prediction error on bothsimulated data and operational measurements. The performance of thealgorithms is compared using nested cross-validation and statisticalhypothesis testing. The statistical evaluation procedure finds thatANNs have the lowest prediction error compared to LR and KNNin estimating fuel consumption on both simulated and operationaldata. The models have a mean relative prediction error of 0.3% onsimulated data, and 4.2% on operational data.

The aim of this thesis is to investigate computerized voice assessment methods to classify between the normal and Dysarthric speech signals. In this proposed system, computerized assessment methods equipped with signal processing and artificial intelligence techniques have been introduced. The sentences used for the measurement of inter-stress intervals (ISI) were read by each subject. These sentences were computed for comparisons between normal and impaired voice. Band pass filter has been used for the preprocessing of speech samples. Speech segmentation is performed using signal energy and spectral centroid to separate voiced and unvoiced areas in speech signal. Acoustic features are extracted from the LPC model and speech segments from each audio signal to find the anomalies. The speech features which have been assessed for classification are Energy Entropy, Zero crossing rate (ZCR), Spectral-Centroid, Mean Fundamental-Frequency (Meanf0), Jitter (RAP), Jitter (PPQ), and Shimmer (APQ). Naïve Bayes (NB) has been used for speech classification. For speech test-1 and test-2, 72% and 80% accuracies of classification between healthy and impaired speech samples have been achieved respectively using the NB. For speech test-3, 64% correct classification is achieved using the NB. The results direct the possibility of speech impairment classification in PD patients based on the clinical rating scale.

Much focus in the joint retailing and transportation domain has been on the transition to e-tailing and the reformation of supply-chain logistics. However, traditional retailing, where consumers visit stores for shopping, dominates and will continue to do so for the foreseeable future. Retailers continuously expand, contract, and reconfigure their store network for strategic reasons. This paper reports on a project aiming to facilitate the incorporation of environmental consequences into the retailer’s reconfiguration decision process. It describes the design and deployment process of eCOmpass, an online decision support tool that enables retailers to estimate the change in transportation-related CO2 emissions caused by a reconfiguration of their store network. This description encompasses the judgmental choices of data acquisition, optimization technology, and user interface. 

Walking in daily life places high demands on the interplay between cognitive and motor functions. A well-functioning dual-tasking ability is thus essential for walking safely. The aims were to study longitudinal changes in gait function during single- and dual-tasking over a period of two years among people with initially mild AD (n = 21). Data were collected on three occasions, twelve months apart. An optical motion capture system was used for three-dimensional gait analysis. Gait parameters were examined at comfortable gait speed during single-tasking, dual-tasking naming names, and naming animals. The dual-task cost for gait speed was pronounced at baseline (names 26%, animals 35%), and remained so during the study period. A significant (p < 0.05) longitudinal decline in gait speed and step length during single- and dual-tasking was observed, whereas double support time, step width and step height showed inconsistent results. Systematic visual examination of the motion capture files revealed that dual-tasking frequently resulted in gait disturbances. Three main characteristics of such disturbances were identified: Temporal disturbance, Spatial disturbance and Instability in single stance. These aberrant gait performances may affect gait stability and increase the risk of falling. Furthermore, the observed gait disturbances can contribute to understanding and explaining previous reported gait variability among individuals with AD. However, the role that dual-task testing and aberrant dual-task gait performance play in the identification of individuals with early signs of cognitive impairment and in predicting fall risk in AD remains to be studied.

This paper presents a new boosting algorithm called NormalBoost which is capable of classifying a multi-dimensional binary class dataset. It adaptively combines several weak classifiers to form a strong classifier. Unlike many boosting algorithms which have high computation and memory complexities, NormalBoost is capable of classification with low complexity. Since NormalBoost assumes the dataset to be continuous, it is also noise resistant because it only deals with the means and standard deviations of each dimension. Experiments conducted to evaluate its performance shows that NormalBoost performs almost the same as AdaBoost in the classification rate. However, NormalBoost performs 189 times faster than AdaBoost and employs a very little amount of memory when a dataset of 2 million samples with 50 dimensions is invoked.

Forgetting is an oft-forgotten art. Many artificial intelligence (AI) systems deliver good performance when first implemented; however, as the contextual environment changes, they become out of date and their performance degrades. Learning new knowledge is part of the solution, but forgetting outdated facts and information is a vital part of the process of renewal. However, forgetting proves to be a surprisingly difficult concept to either understand or implement. Much of AI is based on analogies with natural systems, and although all of us have plenty of experiences with having forgotten something, as yet we have only an incomplete picture of how this process occurs in the brain. A recent judgment by the European Court concerns the "right to be forgotten" by web index services such as Google. This has made debate and research into the concept of forgetting very urgent. Given the rapid growth in requests for pages to be forgotten, it is clear that the process will have to be automated and that intelligent systems of forgetting are required in order to meet this challenge.

This report investigates the possibility to use a different algorithm for predictingthe conductivity [2] in HVDC pure water cooling systems.The study shows that it is possible to use a different algorithm that will raise theaccuracy and reliability significantly.

Objectives: To explore events and describe phases for temporal coordination of the sit-to-walk (STW) task, within a semistandardized set up, in subjects with stroke and matched controls. In addition, to assess variability of STW phase duration and to compare the relative duration of STW phases between the 2 groups.

Design: Cross-sectional.

Setting: Research laboratory.

Participants: A convenience sample of persons with hemiparesis (n=10; age 50–67y), more than 6 months after stroke and 10 controls matched for sex, age, height, and body mass index.

Interventions: Not applicable.

Main Outcome Measures: Relative duration of STW phases, SE of measurement in percentage of the mean, and intraclass correlation coefficients (ICCs).

Results: Four STW phases were defined: rise preparation, transition, primary gait initiation, and secondary gait initiation. The subjects with stroke needed 54% more time to complete the STW task than the controls did. ICCs ranged from .38 to .66 and .22 to .57 in the stroke and control groups, respectively. SEs of measurement in percentage of the mean values were high, particularly in the transition phase: 54.1% (stroke) and 50.4% (controls). The generalized linear model demonstrated that the relative duration of the transition phase was significantly longer in the stroke group.

Conclusions: The present results extend existing knowledge by presenting 4 new phases of temporal coordination of STW, within a semistandardized set-up, in persons with stroke and in controls. The high degree of variability regarding relative STW phase duration was probably a result of both the semistandardized set up and biological variability. The significant difference in the transition phase across the 2 groups requires further study.

This thesis investigates the integration of Artificial Intelligence (AI) into the generation, summarization, and presentation of test reports. The thesis is a cooperation with us, the authors, and Hitachi Energy in Ludvika, Sweden. It focuses on identifying factors that enable a transition to AI tools in the industrial sector. The research evaluates AI tools' feasibility for improving the reporting process within a manufacturing environment. Key challenges identified include significant data integration obstacles and the need to align AI implementation with organizational goals. The study uses a case study approach with empirical data from Hitachi Energy where we evaluate the feasibility of AI tools using a design and creation method. We assess AI's capability to automate the test reporting processes, given the quality and accessibility of data. The study concludes with a set of practical guidelines for AI implementation in industrial settings. These guidelines highlight the importance of high-quality data, integrating AI with existing processes. We discuss employee training to ensure a smooth transition to digitized operations. This research provides insights for industrial companies planning digital transformation initiatives. It focuses on conditions for deploying AI technologies to suit fast-paced technological changes and dynamic industrial demands.

This paper presents a new approach to segment low quality finger-print images which are collected by low quality fingerprint scanners. Images collected using such readers are easy to collect but difficult to segment. The proposed approach focuses on automatically segment and enhance these fingerprint images to reduce the detection of false minutiae and hence improve the recognition rate. There are four major contributions of this paper. Firstly, segmentation of fingerprint images is achieved via morphological filters to find the largest object in the image which is the foreground of the fingerprint. Secondly, specially designed adaptive thresholding algorithm to deal with fingerprint images. The algorithm tries to fit a curve between the gray levels of the pixels of each row or column in the fingerprint image. The curve represents the binarization threshold of each pixel in the corresponding row or column. Thirdly, noise reduction and ridge enhancement is achieved by invoking a rotational invariant anisotropic diffusion filter. Finally, an adaptive thinning algorithm which is immune against spurs is invoked to generate the recognition ready fingerprint image. Segmentation of 100 images from databases FVC2002 and FVC2004 was performed and the experiments showed that 96 % of images under test are correctly segmented.

This chapter presents an overview of traffic sign detection and recognition. It describes the characteristics of traffic signs and the requirements and difficulties when dealing with traffic sign detection and recognition in outdoor images. The chapter also covers the different techniques invoked to segment traffic signs from the different traffic scenes and the techniques employed for the recognition and classification of traffic signs. It points many problems regarding the stability of the received colour information, variations of these colours with respect to the daylight conditions, and absence of a colour model that can led to a good solution. It also proposes an adaptive colour segmentation model based on Neural Networks. The chapter demonstrates the way to classify segmented traffic signs by employing one of widely used classifiers, AdaBoost , based on a set of features, in this case HOG descriptors, which was developed for pedestrian recognition but found the way for many applications in different fields. The chapter ends by showing examples where traffic sign recognition is applicable in vehicle industry

Color represents an important attribute in the field of traffic sign recognition. However, when the color of the traffic sign fades or the traffic scene is collected in gray as in the case of Infrared imaging, then color based recognition systems fail. Other problems related to color are simply that different countries use different colors. Even within the European Union, colors of traffic signs are not the same.

This paper aims to present a new approach to detect traffic signs without color attributes. It is based a two-stage sliding window which detects traffic signs in the multi-scale image. Histogram of Oriented Gradients (HOG) descriptors are computed as a quality function which are evaluated by two SVM classifier; the coarse and the fine detectors. 

Different objects detected by the coarse detectors are clustered and a fine search is conducted in the areas where traffic signs are more probable to exist. 

Experiments conducted to detect traffic signs under different light conditions such as sunny, cloudy, fog and snow fall have showed a performance of 98% and very low false positive rate.  The proposed approach was tested on the Yield traffic signs because it has a simple triangular shape which can be found in many places other than the traffic signs and represent a challenge to the proposed approach.

Color represents an important attribute in the field of traffic sign recognition. However, when the color of the traffic sign fades or the traffic scene is collected in gray as in the case of Infrared imaging, then color based recognition systems fail. Other problems related to color are simply that different countries use different colors. Even within the European Union, colors of traffic signs are not the same. This paper aims to present a new approach to detect traffic signs without color attributes. It is based a two-stage sliding window which detects traffic signs in the multi-scale image. Histogram of Oriented Gradients HOG descriptors are computed as a quality function which are evaluated by two SVM classifier; the coarse and the fine detectors. Different objects detected by the coarse detectors are clustered and a fine search is conducted in the areas where traffic signs are more probable to exist. Experiments conducted to detect traffic signs under different light conditions such as sunny, cloudy, fog and snow fall have showed a performance of 98% and very low false positive rate. The proposed approach was tested on the Yield traffic signs because it has a simple triangular shape which can be found in many places other than the traffic signs which represent a challenge to the proposed approach.

This paper elaborates the routing of cable cycle through available routes in a building in order to link a set of devices, in a most reasonable way. Despite of the similarities to other NP-hard routing problems, the only goal is not only to minimize the cost (length of the cycle) but also to increase the reliability of the path (in case of a cable cut) which is assessed by a risk factor. Since there is often a trade-off between the risk and length factors, a criterion for ranking candidates and deciding the most reasonable solution is defined. A set of techniques is proposed to perform an efficient and exact search among candidates. A novel graph is introduced to reduce the search-space, and navigate the search toward feasible and desirable solutions. Moreover, admissible heuristic length estimation helps to early detection of partial cycles which lead to unreasonable solutions. The results show that the method provides solutions which are both technically and financially reasonable. Furthermore, it is proved that the proposed techniques are very efficient in reducing the computational time of the search to a reasonable amount.

Prohibitory traffic signs represent an important group of traffic signs which are used to prohibit certain types of manoeuvres or some types of traffic. Speed limits signs belong to this group and speed is the main cause of many deadly accidents. Detecting this group in good time may be helpful to avoid many fatal accidents. This paper presents a new approach to detecting prohibitory traffic signs which is based on colour segmentation using LVQ and windowed Hough Transform. Experiments conducted to check the robustness of this approach indicated that 98.5% of the traffic signs invoked for this test were successfully detected. This test was carried out using images collected under a wide range of environmental conditions.

This paper describes an intelligent algorithm for traffic sign recognition which converges quickly, is accurate in its segmentation and adaptive in its behaviour. The proposed approach can segment images of traffic signs in different lighting and environmental conditions and in different countries. It is based on using Kohonen's Self-Organizing Maps (SOM) as a clustering tool and it is developed for Intelligent Vehicle applications. The current approach does not need any prior training. Instead, a slight portion, which is about 1% of the image under investigation, is used for training. This is a key issue to ensure fast convergence and high adaptability. The current approach was tested by using 442 images which were collected under different environmental conditions and from different countries. The proposed approach shows promising results; good improvement of 73% is observed in faded traffic sign images compared with 53.3% using the traditional algorithm. The adaptability of the system is evident from the segmentation of the traffic sign images from various countries where the result is 96% for the nine countries included in the test.

This paper aims to present a new approach to detect traffic signs which is based on color segmentation using AdaBoost binary classifier and circular Hough Transform.The Adaboost classifier was trained to segment traffic signs images according to the desired color. A voting mechanism was invoked to establish a property curve for each of the candidates. SVM classifier was trained to classify the property curves of each object into their corresponding classes.

Experiments conducted on Adaboost color segmentation under different light conditions such as sunny, cloudy, fog and snow fall have showed a performance of 95%. The proposed system was tested on two different groups of traffic signs; the warning and the prohibitory signs. In the case of warning signs, a recognition rate of 98.4% was achieved while it was 97% for prohibitory traffic signs. This test was carried out under a wide range of environmental conditions.

NormalBoost is a new boosting algorithm which is capable of classifying a multi-dimensional binary class dataset. It adaptively combines several weak classifiers to form a strong classifier. Unlike many boosting algorithms which have high computation and memory complexities, NormalBoost is capable of classification with low complexity. The purpose of this paper is to present NormalBoost as a framework which establishes a platform to solve classification problems. The approach was tested with a dataset which was extracted automatically from real-world traffic sign images. The dataset contains both images of traffic sign borders and speed limit pictograms. This framework involves the computation of Haar-like features of these images and then employs the NormalBoost classifier to classify these traffic signs. The total number of images which were classified was 6500 binary images. A -fold validation was invoked to check the validity of the classification which resulted in a classification rate of 98.4% and 98.9% being achieved for these two databases. This framework is distinguished by its invariance to in-plane rotation of the images under consideration. Experiments show that the classification rate remains almost constant when traffic sign images with different angles of rotations were tested.

Speed is one of the major factors by which the traffic safety is affected. If the speed limit traffic signs on the road are recognised and displayed to a driver, this will be a motivation to keep the vehicle's speed within the permitted range. The purpose of this paper is to investigate Eigen-based traffic sign recognition which can aid in the development of Intelligent Speed Adaptation. This system is based on invoking the PCA technique to detect the unknown speed limit traffic sign and computes its best effective Eigen vectors. The traffic sign is then recognized and classified by using the shortest Euclidean distance to the different speed limit traffic sign classes. The system was trained using 24 037 images which were collected in different light conditions. To check the robustness of this system, it was tested against 1429 images and it was found that the accuracy of recognition was 97.5% which indicates clearly the high robustness targeted by this system.

This paper presents a multi-class AdaBoost based on incorporating an ensemble of binary AdaBoosts which is organized as Binary Decision Tree (BDT). It is proved that binary AdaBoost is extremely successful in producing accurate classification but it does not perform very well for multi-class problems. To avoid this performance degradation, the multi-class problem is divided into a number of binary problems and binary AdaBoost classifiers are invoked to solve these classification problems. This approach is tested with a dataset consisting of 6500 binary images of traffic signs. Haar-like features of these images are computed and the multi-class AdaBoost classifier is invoked to classify them. A classification rate of 96.7% and 95.7% is achieved for the traffic sign boarders and pictograms, respectively. The proposed approach is also evaluated using a number of standard datasets such as Iris, Wine, Yeast, etc. The performance of the proposed BDT classifier is quite high as compared with the state of the art and it converges very fast to a solution which indicates it as a reliable classifier.

This paper presents a new approach to segment low quality fingerprint images which are collected by low quality fingerprint readers. Images collected using such readers are easy to collect but difficult to segment. The proposed approach is based on combining global and local processing to achieve segmentation of fingerprint images. On the global level, the fingerprint is located and extracted from the rest of the image by using a global thresholding followed by dilation and edge detection of the largest object in the image. On the local level, fingerprint's foreground and its border image are treated using different fuzzy rules which the two images are segmented. These rules are based on the mean and variance of the block under consideration. The approach is implemented in three stages; preprocessing, segmentation, and post-processing. Segmentation of 100 images was performed and compared with manual examinations by human experts. The experiments showed that 96% of images under test are correctly segmented. The results from the quality of segmentation test revealed that the average error in block segmentation was 2.84% and the false positive and false negatives were approximately 1.4%. This indicates the high robustness of the proposed approach.

The recent days have seen huge developments in deep learning with specific reference to artificial neural networks (ANN).However, ANNs cannot address when data is impregnated with ambiguity, uncertainty of non statistical kind, vagueness,and noise. These factors are detrimental to efficient learning of deep networks. It is exactly here that the role of deep fuzzymodels comes to play. These models can effectively capture the mentioned vagaries of data and are the best to accommodatehumanistic notions, approximations, and tolerance to imprecision. The fruitions of the capabilities of deep fuzzy notionshas led to development of models. In this direction, this paper makes an overall view of ongoing research work related todeep fuzzy models in the individual capacity and hybridized models. This article explores application of the concept in therealm of data processing, fault diagnosis, image processing, Robotics, vulnerability detection systems, and many more. It ishoped that this article of review will facilitate the novice researchers who have set forth in this direction to apply deep fuzzyconcepts to achieve high accuracy in conventional as well as widely used learning tasks such as object recognition,computer vision, and in certain AI applications within a short time.

This paper presents a new algorithm to segment fingerprint images. The algorithm uses four features, the global mean, the local mean, variance and coherence of the image to achieve the fingerprint segmentation. Based on these features, a rule based system is built to segment the image. The proposed algorithm is implemented in three stages; pre-processing, segmentation, and post-processing. Gaussian filter and histogram equalization are applied in the pre-processing stage. Segmentation is applied using the local features. Finally, fill the gaps algorithm and a modified version of Otsu thresholding are invoked in the post-processing stage. In order to evaluate the performance of this method, experiments are performed on FVC2000 DB1. Segmentation of 100 images is performed and compared with manual examinations of human experts. It shows that the proposed algorithm achieves a correct segmentation of 82% of images under test.

This paper presents a new approach to segment low quality fingerprint imageswhich are collected by low quality fingerprint readers. Images collected using such readersare easy to collect but difficult to segment. The proposed approach is based on combiningglobal and local processing to achieve segmentation of fingerprint images. On the globallevel, the fingerprint is located and extracted from the rest of the image by using a globalthresholding followed by dilation and edge detection of the largest object in the image.On the local level, fingerprint’s foreground and its border image are treated using differentfuzzy rules. These rules are based on the mean and variance of the block under consideration.The approach is implemented in three stages: pre-processing, segmentation, andpost-processing.Segmentation of 100 images was performed and compared with manual examinationsby human experts. The experiments showed that 96% of images under test are correctlysegmented. The results from the quality of segmentation test revealed that the averageerror in block segmentation was 2.84% and the false positive and false negatives wereapproximately 1.4%. This indicates the high robustness of the proposed approach.

This paper addresses the problem of traffic sign recognition in real-time conditions. The algorithm presented in this paper is based on detecting traffic signs in life images and videos using pattern matching of the unknown sign’s shape with standard shapes of the traffic signs. The pattern matching algorithm works with shape vertices rather than the whole image. This reduces the computation time which is a crucial factor to fit real-time demands. The algorithm is translation and scaling invariant. It shows high robustness as it is tested with 500 images and several videos and a recognition rate of 97% is achieved.

Night driving is one of the major factors which affects traffic safety. Althoughdetecting oncoming vehicles at night time is a challenging task, it may improve trafficsafety. If the oncoming vehicle is recognised in good time, this will motivate drivers tokeep their eyes on the road. The purpose of this paper is to present an approach to detectvehicles at night based on the employment of a single onboard camera. This system isbased on detecting vehicle headlights by recognising their shapes via an SVM classifierwhich was trained for this purpose. A pairing algorithm was designed to pair vehicleheadlights to ensure that the two headlights belong to the same vehicle. A multi-objecttracking algorithm was invoked to track the vehicle throughout the time the vehicle isin the scene. The system was trained with 503 single objects and tested using 144 587single objects which were extracted from 1410 frames collected from 15 videos and 27moving vehicles. It was found that the accuracy of recognition was 97.9% and the vehiclerecognition rate was 96.3% which indicates clearly the high robustness attained by thissystem.

The purpose of this paper is to analyze the performance of the Histograms of Oriented Gradients (HOG) as descriptors for traffic signs recognition. The test dataset consists of speed limit traffic signs because of their high inter-class similarities.

  HOG features of speed limit signs, which were extracted from different traffic scenes, were computed and a Gentle AdaBoost classifier was invoked to evaluate the different features. The performance of HOG was tested with a dataset consisting of 1727 Swedish speed signs images. Different numbers of HOG features per descriptor, ranging from 36 features up 396 features, were computed for each traffic sign in the benchmark testing. The results show that HOG features perform high classification rate as the Gentle AdaBoost classification rate was 99.42%, and they are suitable to real time traffic sign recognition. However, it is found that changing the number of orientation bins has insignificant effect on the classification rate. In addition to this, HOG descriptors are not robust with respect to sign orientation.

Paper manufacturing is energy demanding and improvedmodelling of the pulp bleach process is the main non-invasive means ofreducing energy costs. In this paper, time it takes to bleach paper pulpto desired brightness is examined. The model currently used is analysedand benchmarked against two machine learning models (Random Forestand TreeBoost). Results suggests that the current model can be super-seded by the machine learning models and it does not use the optimalcompact subset of features. Despite the differences between the machinelearning models, a feature ranking correlation has been observed for thenew models. One novel, yet unused, feature that both machine learningmodels found to be important is the concentration of bleach agent.

Many studies have been done on the identification of the causal effect of a treatment (or intervention) but still the field of causal effect identification is highly debated in many disciplines e.g. data science, econometrics, biostatistics etc. This study examined the theoretical motivation of predictive modelling approach to estimate and check the sensitivity analysis for causal effects by using structural equations modelling framework with application of real data from Landstinget Dalarna. In this study, the link between causal inference using structural equations modelling approach of Heckman's two-step estimation framework and predictive modeling approach has been established. Furthermore, two different simulation studies under linearity and nonlinearity assumptions were conducted to see the finite sample properties of the predictive modelling approaches such as Support Vector Machine, Random Forest, Gaussian boosted regression trees and Bayesian additive regression trees. Finally, the predictive modelling approaches were used to estimate the treatment effects on patient outcomes in terms of length of stay, unplanned readmission and mortality within 30 days after discharge. The results were also compared with the traditional approaches: propensity score matching, propensity weighting and two-step estimation method. This study used two types of estimates, average treatment effects and treatment on treated. The estimates and their standard errors were calculated for the real data from Landstinget Dalarna. The study found that, non-outlying patients are staying in the hospital for longer periods of time (in days) compared to outlying patients though the reasons that make long of stay remained unknown. For readmission and mortality, the results varied a lot between the alternative models, and we can therefore not conclude that non-outlying patients have higher mortality and readmission rates compared to outlying.

The usage of energy directly leads to a great amount of consumption of the non-renewable fossil resources. Exploiting fossil resources energy can influence both climate and health via ineluctable emissions. Raising awareness, choosing alternative energy and developing energy efficient equipment contributes to reducing the demand for fossil resources energy, but the implementation of them usually takes a long time. Since building energy amounts to around one-third of global energy consumption, and systems in buildings, e.g. HVAC, can be intervened by individual building management, advanced and reliable control techniques for buildings are expected to have a substantial contribution to reducing global energy consumptions. Among those control techniques, the model-free, data-driven reinforcement learning method seems distinctive and applicable. The success of the reinforcement learning method in many artificial intelligence applications has brought us an explicit indication of implementing the method on building energy control. Fruitful algorithms complement each other and guarantee the quality of the optimisation. As a central brain of smart building automation systems, the control technique directly affects the performance of buildings. However, the examination of previous works based on reinforcement learning methodologies are not available and, moreover, how the algorithms can be developed is still vague. Therefore, this paper briefly analyses the empirical applications from the methodology point of view and proposes the future research direction.

In a global economy, manufacturers mainly compete with cost efficiency of production, as the price of raw materials are similar worldwide. Heavy industry has two big issues to deal with. On the one hand there is lots of data which needs to be analyzed in an effective manner, and on the other hand making big improvements via investments in cooperate structure or new machinery is neither economically nor physically viable. Machine learning offers a promising way for manufacturers to address both these problems as they are in an excellent position to employ learning techniques with their massive resource of historical production data. However, choosing modelling a strategy in this setting is far from trivial and this is the objective of this article. The article investigates characteristics of the most popular classifiers used in industry today. Support Vector Machines, Multilayer Perceptron, Decision Trees, Random Forests, and the meta-algorithms Bagging and Boosting are mainly investigated in this work. Lessons from real-world implementations of these learners are also provided together with future directions when different learners are expected to perform well. The importance of feature selection and relevant selection methods in an industrial setting are further investigated. Performance metrics have also been discussed for the sake of completion.

The demand of electricity is constantly increasing and requires an expansion of the Swedish power system. The safety of electric power transmission is taken for granted by most of the customers, most of them do not have an idea that earth fault occurs regularly and that such event could present a hazard. Currently, measurements with the heavy-current injection method are used to ensure safety. This method requires that the transmission lines be taken out of service for several days with large costs due to reduced power transfer capability. The safety requirements are regulated by law and the purpose is to make sure that no touch voltages of dangerous level occur in the low voltage grid. This study investigates if computer simulations with the software XGSLab can replace present method of measurement. If simulation of touch voltages could be used in compliance with regulations, there is a great opportunity to save both time and money. The method involves modeling electricity grids with both high and low voltage. The high voltage grid is injected with a fault current which leads to an increase in earth potential. This increase means that touch voltages occur in an adjacent low-voltage grid. The values from the simulations in XGSLab is then compared with the results from previous measurements using the heavy-current injection method. To better understand how the assumptions made affect the results, a study was conducted that examined the various parameters available in the program. With some simple adaptations the method managed to identify 83 % of the measurement points that the heavy-current injection method found where the level of touch voltages exceeded the acceptable value. The study also points out some difficulties which requires more information to be analyzed to make the method even more reliable. 

Accurately estimating the charging behaviours of electric vehicles (EVs) is crucial for various applications, such as charging station planning and grid impact estimation. However, the analysis of EV charging behaviours using real-world data remains limited due to (confidential) data availability constraints. Furthermore, while existing modelling studies have demonstrated EVs as effective tools for electricity balancing and delivery between locations, their potential remains unexplored empirically. This study aims to bridge the research gap by studying EV charging behaviours and their capacity for electricity balancing and delivery. Using data from 179,665 realworld charging sessions in Sweden, we employed statistical and clustering analysis to scrutinize charging behaviours comprehensively. Synthetic weekly charging load profiles are generated for both residential areas and workplaces, considering varying charging power levels, which can be used as inputs for large-scale EV charging load modelling. Furthermore, performance indicators are proposed to quantify the potential of EVs for electricity balancing and delivery. Results show that EVs exhibit significant potential for electricity balancing (up to 51.5 kWh daily). Many EV owners underutilize their EV battery capacity, providing an opportunity for active electricity delivery across locations. This study can help understand EV charging behaviours and recognize their significant potentials for electricity regulation and integrating more renewables in the future power system.

The use of electric vehicles (EVs) has been on the rise. Most of the existing EV smart charging controls can be categorized into three approaches according to their optimization principles: individual, bottom-up and top-down. Until now, systematic comparison and analysis of the different approaches are still lacking. It is still unknown whether a control approach performs better than others and, if yes, why is it so. This study aims to fill in such knowledge gaps by conducting a systematic comparison of these three different control approaches and analyzing their performances in depth. A representative control algorithm will be selected from each control approach, then the selected algorithms will be applied for optimizing EV charging loads in a building community in Sweden. Their power regulation performances will be comparatively investigated. This study will help pave the way for the developments of more sophisticated control algorithms for EV smart charging. © 2022 The Authors, published by EDP Sciences.

The demands of image processing related systems are robustness, high recognition rates, capability to handle incomplete digital information, and magnanimous flexibility in capturing shape of an object in an image. It is exactly here that, the role of convex hulls comes to play. The objective of this paper is twofold. First, we summarize the state of the art in computational convex hull development for researchers interested in using convex hull image processing to build their intuition, or generate nontrivial models. Secondly, we present several applications involving convex hulls in image processing related tasks. By this, we have striven to show researchers the rich and varied set of applications they can contribute to. This paper also makes a humble effort to enthuse prospective researchers in this area. We hope that the resulting awareness will result in new advances for specific image recognition applications.