In this thesis, an anomaly detection framework is assessed and fine-tuned to detect and explain anomalies in a power station, where EtherCAT, an Industrial Control System, is employed for monitoring. The chosen framework is based on a previously published Graph Neural Network (GNN) model, utilizing attention mechanisms to capture complex relationships between diverse measurements within the EtherCAT system. To address the challenges in graph learning and improve model performance and computational efficiency, the study introduces a novel similarity thresholding approach. This approach dynamically selects the number of neighbors for each node based on their similarity instead of adhering to a fixed 'k' value, thus making the learning process more adaptive and efficient. Further in the exploration, the study integrates Extreme Value Theory (EVT) into the framework to set the anomaly detection threshold and assess its effectiveness. The effect of temporal features on model performance is examined, and the role of seconds of the day as a temporal feature is notably highlighted. These various methodological innovations aim to refine the application of the attention based GNN framework to the EtherCAT system. The results obtained in this study illustrate that the similarity thresholding approach significantly improves the model's F1 score compared to the standard TopK approach. The inclusion of seconds of the day as a temporal feature led to modest improvements in model performance, and the application of EVT as a thresholding technique was explored, although it did not yield significant benefits in this context. Despite the limitations, including the utilization of a single-day dataset for training, the thesis provides valuable insights for the detection of anomalies in EtherCAT systems, contributing both to the literature and the practitioners in the field. It lays the groundwork for future research in this domain, highlighting key areas for further exploration such as larger datasets, alternative anomaly detection techniques, and the application of the framework in streaming data environments. / I denna avhandling utvärderas och finslipas ett ramverk för att detektera och förklara anomalier på ett kraftverk, där EtherCAT, ett industriellt styrsystem, används för övervakning. Det valda ramverket är baserat på en tidigare publicerad graf neurala nätverksmodell (GNN) som använder uppmärksamhetsmekanismer för att fånga komplexa samband mellan olika mätningar inom EtherCAT-systemet. För att hantera utmaningar inom grafiskt lärande och förbättra modellens prestanda och beräkningseffektivitet introducerar studien en ny metod för likhetsgränsdragning. Denna metod väljer dynamiskt antalet grannar för varje nod baserat på deras likhet istället för att hålla sig till ett fast 'k'-värde, vilket gör inlärningsprocessen mer anpassningsbar och effektiv. I en vidare undersökning integrerar studien extremvärdesteori (EVT) i ramverket för att sätta tröskeln för detektering av anomalier och utvärdera dess effektivitet. Effekten av tidsberoende egenskaper på modellens prestanda undersöks, och sekunder av dagen som en tidsberoende egenskap framhävs särskilt. Dessa olika metodologiska innovationer syftar till att förädla användningen av det uppmärksamhetsbaserade GNN-ramverket på EtherCAT-systemet. Resultaten som erhållits i denna studie illustrerar att likhetsgränsdragning väsentligt förbättrar modellens F1-poäng jämfört med den standardiserade TopK-metoden. Inkluderingen av sekunder av dagen som en tidsberoende egenskap ledde till blygsamma förbättringar i modellens prestanda, och användningen av EVT som en tröskelmetod undersöktes, även om den inte gav några betydande fördelar i detta sammanhang. Trots begränsningarna, inklusive användningen av ett dataset för endast en dag för träning, ger avhandlingen värdefulla insikter för detektering av anomalier i EtherCAT-system, och bidrar både till litteraturen och praktiker inom området. Den lägger grunden för framtida forskning inom detta område, och belyser nyckelområden för ytterligare utforskning såsom större dataset, alternativa tekniker för detektering av anomalier och tillämpningen av ramverket i strömmande data-miljöer.

Cyber-physical systems (CPS) connect the physical and digital domains and are often realized as spatially distributed. CPS is built on the Internet of Things (IoT) and Internet of Services, which use cloud architecture to link a swarm of devices over a decentralized network. Modern CPSs are undergoing a foundational shift as Industry 4.0 is continually expanding its boundaries of digitization. From automating the industrial manufacturing process to interconnecting sensor devices within buildings, Industry 4.0 is about developing solutions for the digitized industry. An extensive amount of engineering efforts are put to design dynamically scalable and robust automation solutions that have the capacity to integrate heterogeneous CPS. Such heterogeneous systems must be able to communicate and exchange information with each other in real-time even if they are based on different underlying technologies, protocols, or semantic definitions in the form of ontologies. This development is subject to interoperability challenges and knowledge gaps that are addressed by engineers and researchers, in particular, machine learning approaches are considered to automate costly engineering processes. For example, challenges related to predictive maintenance operations and automatic translation of messages transmitted between heterogeneous devices are investigated using supervised and unsupervised machine learning approaches. In this thesis, a machine learning-based collaboration and automation-oriented IIoT framework named Cloud-based Collaborative Learning (CCL) is developed. CCL is based on a service-oriented architecture (SOA) offering a scalable CPS framework that provides machine learning-as-a-Service (MLaaS). Furthermore, interoperability in the context of the IIoT is investigated. I consider the ontology of an IoT device to be its language, and the structure of that ontology to be its grammar. In particular, the use of aggregated language and structural encoders is investigated to improve the alignment of entities in heterogeneous ontologies. Existing techniques of entity alignment are based on different approaches to integrating structural information, which overlook the fact that even if a node pair has similar entity labels, they may not belong to the same ontological context, and vice versa. To address these challenges, a model based on a modification of the BERT_INT model on graph triples is developed. The developed model is an iterative model for alignment of heterogeneous IIoT ontologies enabling alignments within nodes as well as relations. When compared to the state-of-the-art BERT_INT, on DBPK15 language dataset the developed model exceeds the baseline model by (HR@1/10, MRR) of 2.1%. This motivated the development of a proof-of-concept for conducting an empirical investigation of the developed model for alignment between heterogeneous IIoT ontologies. For this purpose, a dataset was generated from smart building systems and SOSA and SSN ontologies graphs. Experiments and analysis including an ablation study on the proposed language and structural encoders demonstrate the effectiveness of the model. The suggested approach, on the other hand, highlights prospective future studies that may extend beyond the scope of a single thesis. For instance, to strengthen the ablation study, a generalized IIoT ontology that is designed for any type of IoT devices (beyond sensors), such as SAREF can be tested for ontology alignment. Next potential future work is to conduct a crowdsourcing process for generating a validation dataset for IIoT ontology alignment and annotations. Lastly, this work can be considered as a step towards enabling translation between heterogeneous IoT sensor devices, therefore, the proposed model can be extended to a translation module in which based on the ontology graphs of any device, the model can interpret the messages transmitted from that device. This idea is at an abstract level as of now and needs extensive efforts and empirical study for full maturity.

The condition of multimorbidity — the presence of two or more medical conditions in an individual — is a growing phenomenon worldwide. In the United States, multimorbid patients represent more than a third of the population and the trend is steadily increasing in an already aging population. There is thus a pressing need to understand the patterns in which multimorbidity occurs, and to better understand the nature of the care that is required to be provided to such patients. In this thesis, we use data from the Medical Expenditure Panel Survey (MEPS) from the years 2011 to 2015 to identify combinations of multiple chronic conditions (MCCs). We first quantify the significant heterogeneity observed in these combinations and how often they are observed across the five years. Next, using two criteria associated with each combination -- (a) the annual prevalence and (b) the annual median expenditure -- along with the concept of non-dominated Pareto fronts, we determine the degree of impact each combination has on the healthcare system. Our analysis reveals that combinations of four or more conditions are often mixtures of diseases that belong to different clinically meaningful groupings such as the metabolic disorders (diabetes, hypertension, hyperlipidemia); musculoskeletal conditions (osteoarthritis, spondylosis, back problems etc.); respiratory disorders (asthma, COPD etc.); heart conditions (atherosclerosis, myocardial infarction); and mental health conditions (anxiety disorders, depression etc.). Next, we use unsupervised learning techniques such as association rule mining and hierarchical clustering to visually explore the strength of the relationships/associations between different conditions and condition groupings. This interactive framework allows epidemiologists and clinicians (in particular primary care physicians) to have a systematic approach to understand the relationships between conditions and build a strategy with regards to screening, diagnosis and treatment over a longer term, especially for individuals at risk for more complications. The findings from this study aim to create a foundation for future work where a more holistic view of multimorbidity is possible.

This project is about trying to apply machine learning theories on a selection of data points in order to see if an improvement of current methodology within stress detection and measure selecting could be applicable for the company Linkura AB. Linkura AB is a medical technology company based in Linköping and handles among other things stress measuring for different companies employees, as well as health coaching for selecting measures. In this report we experiment with different methods and algorithms under the collective name of Unsupervised Learning, to identify visible patterns and behaviour of data points and further on we analyze it with the quantity of data received. The methods that have been practiced on during the project are “K-means algorithm” and a dynamic hierarchical clustering algorithm. The correlation between the different data points parameters is analyzed to optimize the resource consumption, also experiments with different number of parameters are tested and discussed with an expert in stress coaching. The results stated that both algorithms can create clusters for the risk groups, however, the dynamic clustering method clearly demonstrate the optimal number of clusters that should be used. Having consulted with mentors and health coaches regarding the analysis of the produced clusters, a conclusion that the dynamic hierarchical cluster algorithm gives more accurate clusters to represent risk groups were done. The conclusion of this project is that the machine learning algorithms that have been used, can categorize data points with stress behavioral correlations, which is usable in measure testimonials. Further research should be done with a greater set of data for a more optimal result, where this project can form the basis for the implementations. / Detta projekt handlar om att försöka applicera maskininlärningsmodeller på ett urval av datapunkter för att ta reda på huruvida en förbättring av nuvarande praxis inom stressdetektering och  åtgärdshantering kan vara applicerbart för företaget Linkura AB. Linkura AB är ett medicintekniskt företag baserat i Linköping och hanterar bland annat stressmätning hos andra företags anställda, samt hälso-coachning för att ta fram åtgärdspunkter för förbättring. I denna rapport experimenterar vi med olika metoder under samlingsnamnet oövervakad maskininlärning för att identifiera synbara mönster och beteenden inom datapunkter, och vidare analyseras detta i förhållande till den mängden data vi fått tillgodosett. De modeller som har använts under projektets gång har varit “K-Means algoritm” samt en dynamisk hierarkisk klustermodell. Korrelationen mellan olika datapunktsparametrar analyseras för att optimera resurshantering, samt experimentering med olika antal parametrar inkluderade i datan testas och diskuteras med expertis inom hälso-coachning. Resultaten påvisade att båda algoritmerna kan generera kluster för riskgrupper, men där den dynamiska modellen tydligt påvisar antalet kluster som ska användas för optimalt resultat. Efter konsultering med mentorer samt expertis inom hälso-coachning så drogs en slutsats om att den dynamiska modellen levererar tydligare riskkluster för att representera riskgrupper för stress. Slutsatsen för projektet blev att maskininlärningsmodeller kan kategorisera datapunkter med stressrelaterade korrelationer, vilket är användbart för åtgärdsbestämmelser. Framtida arbeten bör göras med ett större mängd data för mer optimerade resultat, där detta projekt kan ses som en grund för dessa implementeringar.

Being able to read and interpret affective states plays a significant role in human society. However, this is difficult in some situations, especially when information is limited to either vocal or visual cues. Many researchers have investigated the so-called basic emotions in a supervised way. This thesis holds the results of a multimodal supervised and unsupervised study of a more realistic number of emotions. To that end, audio and video features are extracted from the GEMEP dataset employing openSMILE and OpenFace, respectively. The supervised approach includes the comparison of multiple solutions and proves that multimodal pipelines can outperform unimodal ones, even with a higher number of affective states. The unsupervised approach embraces a traditional and an exploratory method to find meaningful patterns in the multimodal dataset. It also contains an innovative procedure to better understand the output of clustering techniques. / Att kunna läsa och tolka affektiva tillstånd spelar en viktig roll i det mänskliga samhället. Detta är emellertid svårt i vissa situationer, särskilt när information är begränsad till antingen vokala eller visuella signaler. Många forskare har undersökt de så kallade grundläggande känslorna på ett övervakat sätt. Det här examensarbetet innehåller resultaten från en multimodal övervakad och oövervakad studie av ett mer realistiskt antal känslor. För detta ändamål extraheras ljud- och videoegenskaper från GEMEP-data med openSMILE respektive OpenFace. Det övervakade tillvägagångssättet inkluderar jämförelse av flera lösningar och visar att multimodala pipelines kan överträffa unimodala sådana, även med ett större antal affektiva tillstånd. Den oövervakade metoden omfattar en konservativ och en utforskande metod för att hitta meningsfulla mönster i det multimodala datat. Den innehåller också ett innovativt förfarande för att bättre förstå resultatet av klustringstekniker.

No description available.

No description available.

Quality software usually means high reliability, which in turn has two main components; the software should provide correctness, which means it should perform the specified task, and robustness in the sense that it should be able to manage unexpected situations. In other words, reliable systems are systems without bugs. Because of this, testing and debugging are recurrent and resource expensive tasks in software development, notably in large software systems. This thesis investigate the potential of using unsupervised machine learning on Ericsson bug reports to avoid unnecessary debugging by identifying duplicate bug reports. The bug report data that is considered are crash dumps from software crashes. The data is clustered using the clustering algorithms k-modes, k-prototypes and expectation maximization where-after the generated clusters are used to assign new incoming bug reports to the previously generated clusters, thus indicating whether an old bug report is similar to the newly submitted one. Due to the dataset only being partially labeled both internal and external validity indices are applied to evaluate the clustering. The results indicate that many, small clusters can be identified using the applied method. However, for the results to have high validity the methods could be applied on a larger data set. / Mjukvara av hög kvalitet innebär ofta hög tillförlitlighet, vilket i sin tur har två huvudkomponenter; mjukvaran bör vara korrekt, den ska alltså uppfylla dom specifierade kraven, och dessutom robust vilket innebär att den ska kunna hantera oväntade situationer. Med andra ord, tillförlitliga system är system utan buggar. På grund av detta är testning och felsökning återkommande och resurskrävande uppgifter inom mjukvaruutveckling, i synnerhet för stora mjukvarusystem. Detta arbete utforskar vilken potential oövervakad maskininlärning på Ericssons felrapporter har för att undvika onödig felsökning genom att identifiera felrapporter som är dubletter. Felrapporterna som används i detta arbete innehåller data som sparats i minnet vid en mjukvarukrasch. Data klustras sedan med klustringsalgoritmerna k-modes, k-prototypes och expectation maximization varpå dom genererade klustren används för att tilldela nya inkommande felrapporter till de tidigare generade klustren, för att på så sätt kunna identifiera om en gammal felrapport är lik en ny felrapport. Då de felrapporter som behandlas endast till viss del redan är märkta som dubletter används både externa och interna valideringsmått för att utvärdera klustringen. Resultaten tyder på att många, små kluster kunde identifieras med de använda metoderna. Dock skulle metoderna behöva appliceras på ett dataset med större antal felrapporter för att resultaten ska få hög validitet.

This work proposes a new approach to fast detection of abnormal behaviour of cooling, IT, and power distribution systems in micro data centers based on machine learning techniques. Conventional protection of micro data centers focuses on monitoring individual parameters such as temperature at different locations and when these parameters reach certain high values, then an alarm will be triggered. This research employs machine learning techniques to extract normal and abnormal behaviour of the cooling and IT systems. Developed data acquisition system together with unsupervised learning methods quickly learns the physical dynamics of normal operation and can detect deviations from such behaviours. This provides an efficient way for not only producing health index for the micro data center, but also a rich label logging system that will be used for the supervised learning methods. The effectiveness of the proposed detection technique is evaluated on an micro data center placed at Computing Infrastructure Research Center (CIRC) in McMaster Innovation Park (MIP), McMaster University. / Thesis / Master of Science (MSc)

Recent progress on deep neural networks has shown remarkable action recognition performance from videos. The remarkable performance is often achieved by transfer learning: training a model on a large-scale labeled dataset (source) and then fine-tuning the model on the small-scale labeled datasets (targets). However, existing action recognition models do not always generalize well on new tasks or datasets because of the following two reasons. i) Current action recognition datasets have a spurious correlation between action types and background scene types. The models trained on these datasets are biased towards the scene instead of focusing on the actual action. This scene bias leads to poor generalization performance. ii) Directly testing the model trained on the source data on the target data leads to poor performance as the source, and target distributions are different. Fine-tuning the model on the target data can mitigate this issue. However, manual labeling small- scale target videos is labor-intensive. In this dissertation, I propose solutions to these two problems. For the first problem, I propose to learn scene-invariant action representations to mitigate the scene bias in action recognition models. Specifically, I augment the standard cross-entropy loss for action classification with 1) an adversarial loss for the scene types and 2) a human mask confusion loss for videos where the human actors are invisible. These two losses encourage learning representations unsuitable for predicting 1) the correct scene types and 2) the correct action types when there is no evidence. I validate the efficacy of the proposed method by transfer learning experiments. I trans- fer the pre-trained model to three different tasks, including action classification, temporal action localization, and spatio-temporal action detection. The results show consistent improvement over the baselines for every task and dataset. I formulate human action recognition as an unsupervised domain adaptation (UDA) problem to handle the second problem. In the UDA setting, we have many labeled videos as source data and unlabeled videos as target data. We can use already exist- ing labeled video datasets as source data in this setting. The task is to align the source and target feature distributions so that the learned model can generalize well on the target data. I propose 1) aligning the more important temporal part of each video and 2) encouraging the model to focus on action, not the background scene, to learn domain-invariant action representations. The proposed method is simple and intuitive while achieving state-of-the-art performance without training on a lot of labeled target videos. I relax the unsupervised target data setting to a sparsely labeled target data setting. Then I explore the semi-supervised video action recognition, where we have a lot of labeled videos as source data and sparsely labeled videos as target data. The semi-supervised setting is practical as sometimes we can afford a little bit of cost for labeling target data. I propose multiple video data augmentation methods to inject photometric, geometric, temporal, and scene invariances to the action recognition model in this setting. The resulting method shows favorable performance on the public benchmarks. / Doctor of Philosophy / Recent progress on deep learning has shown remarkable action recognition performance. The remarkable performance is often achieved by transferring the knowledge learned from existing large-scale data to the small-scale data specific to applications. However, existing action recog- nition models do not always work well on new tasks and datasets because of the following two problems. i) Current action recognition datasets have a spurious correlation between action types and background scene types. The models trained on these datasets are biased towards the scene instead of focusing on the actual action. This scene bias leads to poor performance on the new datasets and tasks. ii) Directly testing the model trained on the source data on the target data leads to poor performance as the source, and target distributions are different. Fine-tuning the model on the target data can mitigate this issue. However, manual labeling small-scale target videos is labor-intensive. In this dissertation, I propose solutions to these two problems. To tackle the first problem, I propose to learn scene-invariant action representations to mitigate background scene- biased human action recognition models for the first problem. Specifically, the proposed method learns representations that cannot predict the scene types and the correct actions when there is no evidence. I validate the proposed method's effectiveness by transferring the pre-trained model to multiple action understanding tasks. The results show consistent improvement over the baselines for every task and dataset. To handle the second problem, I formulate human action recognition as an unsupervised learning problem on the target data. In this setting, we have many labeled videos as source data and unlabeled videos as target data. We can use already existing labeled video datasets as source data in this setting. The task is to align the source and target feature distributions so that the learned model can generalize well on the target data. I propose 1) aligning the more important temporal part of each video and 2) encouraging the model to focus on action, not the background scene. The proposed method is simple and intuitive while achieving state-of-the-art performance without training on a lot of labeled target videos. I relax the unsupervised target data setting to a sparsely labeled target data setting. Here, we have many labeled videos as source data and sparsely labeled videos as target data. The setting is practical as sometimes we can afford a little bit of cost for labeling target data. I propose multiple video data augmentation methods to inject color, spatial, temporal, and scene invariances to the action recognition model in this setting. The resulting method shows favorable performance on the public benchmarks.