Linköping Electronic Conference Proceedings
Not a member yet
1113 research outputs found
Sort by
Experiments on Automatic Error Detection and Correction for Uruguayan Learners of English
Get PDFThis paper presents an initial experiment on Grammatical Error Correction and Automatic Grading for short texts written by Uruguayan students that are learning English. We present a set of error detection and correction heuristics, and some experiments on using these heuristics for predicting the grade. Although our experiments are limited due to the nature of the dataset, they are a good proof of concept with promising results that might be extended in the future
Automated Assessment of Task Completion in Spontaneous Speech for Finnish and Finland Swedish Language Learners
Get PDFThis study investigates the feasibility of automated content scoring for spontaneous spoken responses from Finnish and Finland Swedish learners. Our experiments reveal that pretrained Transformer-based models outperform the tf-idf baseline in automatic task completion grading. Furthermore, we demonstrate that pre-fine-tuning these models to differentiate between responses to distinct prompts enhances subsequent task completion finetuning. We observe that task completion classifiers exhibit accelerated learning and produce predictions with stronger correlations to human grading when accounting for task differences. Additionally, we find that employing similarity learning, as opposed to conventional classification fine-tuning, further improves the results. It is especially helpful to learn not just the similarities between the responses in one score bin, but the exact differences between the average human scores responses received. Lastly, we demonstrate that models applied to both manual and ASR transcripts yield comparable correlations to human grading
Optimal indoor temperature flexibility for thermal peak shaving in buildings connected to the district heating network
Get PDFBuildings are currently non optimally controlled, using a weather compensation controller that depends only on external temperature. A rich amount of real-time data is available and can be used for better control. This work is focused on developing a general and dynamic model for utilizing the building as an energy storage for a peak-shaving control strategy.A dynamic grey-box model is developed using industry standard operators’ data from a multifamily building, Building A, located in Västerås, Sweden. The training period is set to 408 hours, and the prediction horizon to 48 hours. The model is verified in 4 steps: prediction ability on the historic data, parametric verification on the time constant, simulation of heat supply separated from the historic data and model generality by implementing the model on a second multifamily building, Building B. The modelling errors over a two-month simulated period are 8% for Building A and 9% for Building B. To demonstrate the utilization possibilities, an optimizer is constructed to evaluate a peak shaving control strategy. Different flexibilities for the indoor temperature have been examined yielding heat load peak shaving between 30 to 45%. Flexibility paves the way for improvement in pricing models for the heating sector. This work demonstrates the potential for utilizing building heat storage capacity to reduce peak consumption and costs
Phase Fractions and Velocities in Multiphase Flow – Estimation using Sensor Data Fusion and Machine Learning
Get PDFThere is a strong interest in quantifying the amount of gas and its flow rate to facilitate better control of the processes involved in many industries. There are usually many sensors monitoring these processes, both intrusive and invasive, as well as non-invasive sensors which are usually clamped on to the process pipelines in which the multiphase flow occurs. In the multiphase flow rigs at Equinor and the University of South-Eastern Norway, experiments have been performed with different combinations and velocities of the phases and multiple sensors have been logged. The data from these sensors have been used to estimate volume fractions of the phases as well as their flow rates. This paper presents the estimated results of volume fractions and velocities of selected phases, obtained by fusing data from multiple sensors that monitor density, differential pressure, temperature, and acoustic emission using machine learning (ML) algorithms. These ML algorithms use neural networks with the non-linear input-output type with Levenberg-Marquardt training and provide estimates of volume fractions and phase velocities with RMSE values in the range of 4.6 to 16 m3/h, with the lowest RMSE for gas and the highest for multiphase flow. The total flow rate for the multiphase flow was in the range 30 to 120 m3/h. Results are compared with ML models using data from non-invasive sensors
A Deep Learning Approach for Fault Diagnosis of Hydrogen Fueled Micro Gas Turbines
Get PDFHydrogen fueled gas turbines are susceptible to rigorous health degradation in form of corrosion and erosion in the turbine section of a retrofitted gas turbine due to drastically different thermophysical properties of flue gas stemming from hydrogen combustion. In this context fault diagnosis of hydrogen fueled gas turbines becomes indispensable. To authors knowledge, there is a scarcity of fault diagnosis studies for retrofitted gas turbines considering hydrogen as a potential fuel. The present study, however, develops an artificial neural network (ANN) based fault diagnosis model using MATLAB environment. Prior to fault detection, isolation and identification modules, physics-based performance data of 100 kW micro gas turbine (MGT) was synthesized using GasTurb tool. ANN based classification algorithm showed a 99.4% classification accuracy of fault detection and isolation. Moreover, the feedforward neural network-based regression algorithm showed quite good training, testing and validation accuracies in terms of root mean square error (RMSE). The study revealed that presence of hydrogen induced corrosion fault (both as single corrosion fault or as simultaneous fouling and corrosion) led to false alarms thereby prompting other wrong faults during fault detection and isolation modules. Additionally, performance of fault identification module for hydrogen fuel scenario was found to be marginally lower than that of natural gas case due to assuming small magnitudes of faults arising from hydrogen induced corrosion
Information extraction from operator interface images using computer vision and machine learning
Get PDFIn the process of system upgrades or migrations, the utilization of existing layouts and object structures for designing new Human Machine Interfaces (HMI) can significantly save time and effort. Operator interface images, commonly referred to as HMI´s, contain valuable information crucial to industrial operations, but access to source code or design files can be limited. Modern frameworks for object detection and text recognition offer a solution by extracting information directly from images. However, these methods require time-consuming data acquisition and manual effort to initiate. This paper proposes a novel approach utilizing traditional Computer Vision (CV) and Machine Learning (ML) techniques to extract objects from images. The extracted objects are used as training data to transfer learn a ResNet model for multi-label image classification. The combination of this model with techniques such as sliding window, pyramid scaling, and non-maximum suppression forms the basis for a semi-automated annotation tool. This tool generates training data for more optimized object detection methods, specifically the YOLO (You Only Look Once) one-stage object detector. The semi-automated annotation tool allows engineers to manually refine the training data and export state-of-the-art training images for YOLO. The YOLO model achieves an impressive mean Average Precision at IoU 50% (mAP50) score of 95.5% when transfer learned on the annotated data. Additionally, an Optical Character Recognition (OCR) engine is utilized to extract text information from preprocessed images, followed by postprocessing to filter tag data. An algorithm is then employed to link objects and tags together. The final solution is implemented in software designed to optimize user interaction, resulting in an analysis document in Excel format, which can be easily exported for end-user access. With the novel use of this software to automate image analysis, the time required to analyze HMI images prior to migration or rebuild can be reduced by an estimate of 90%
Automatic Translator from System Dynamics to Modelica with Application to Socio-Bio-Physical Systems
No full textSystem Dynamics is a modelling paradigm that has been applied to a wide range of systems, from economic to physical and from managerial to ecological. The main strength of the paradigm is its ease of use. A System Dynamics modeller does not need to focus on equations; instead, models are expressed in terms of stocks and flows. Modelica, on the other hand, is an equation-based modelling language capable of multi-domain modelling using equations. It gives the user more freedom but requires more mathematical focus and skills. Therefore, a unification of equation-based modelling and the System Dynamics paradigm is seen as highly beneficial. Advantages include the ability for System Dynamics modellers to use the tools available in the Modelica ecosystem. Furthermore, it allows the integration of System Dynamics models into Modelica models. To achieve this goal, we developed an XMILE-to-Modelica translator that maps System Dynamics models represented in the XMILE standard exchange format to Modelica models. We also applied a Modelica-to-Julia translator to demonstrate the generality of the approach.We translated several models to test the correctness of the translator. In particular, the Earth System Climate Interpretable Model (ESCIMO) was translated from its original version in the Vensim toolkit into the OpenModelica toolkit, and a correct validation was obtained by comparing simulation results between simulators. Our work improves tool interoperability and further demonstrates the feasibility of using Modelica as a unified, standard language to integrate models created using System Dynamics, including large and complex socio-biophysical systems
Mapping Simulation optimization requirements for construction sites: A study in heavy-duty vehicles industry
Get PDFThe Construction and mining Industry comprises complex operations and interactions between various actors at different levels. Simulation has emerged as a valuable tool in this domain to better understand the site's behavior and optimize its operation. However, developing a simulation platform that can handle all the operations on the site is challenging due to the computational cost of the digital representation of reality along with the required accuracy level.This paper aims at extracting and mapping the optimization requirements of construction sites at three main levels: site level, operational level and dynamics level. More precisely, this work seeks to define and map the most important requirements between these levels that ensure simulation credibility and reliability.Based on interviews with experts in the domain, both from academia and industry, several key insights and recommendations emerged: at the site level, the layout and the key performance indicators, such as productivity, time, cost, number of machines and workers, need to be modeled and simulated. At the operational level, the simulation platform must include the main activities, such as loading, excavating, transporting and dumping. Moreover, the dynamics level should involve machine models and their interactions with the site's environment, such as earthmoving, drilling, excavating and blasting
Civila insatspersoner från mikro- till makro-nivå – att sprida och stödja lokala initiativ
No full textFrivilliga i samarbete med kommunal räddningstjänst – civila insatspersoner, startade med initiativet Förstärkt Medmänniska i Medelpad för över ett decennium sedan. Det anammades först av landsbygd och små kommuner för att runt 2018 introduceras i storstad av Södertörns Brandsvarförbund. Vi har studerat konceptet från olika håll och som en form av hybrid samproduktion (de frivilliga är inte organiserade d v s tillhör en organisation men de är heller inte spontanfrivilliga) och tittat på exempelvis möjligheter, nytta, utmaningar, hinder och behov. Vi har också jämfört landsbygds- och storstadsmiljö. Vi vet civila insatspersoner där de finns – gör nytta. Liv har räddats, liksom material och restvärden. Trots detta har konceptet haft långsam spridning och många räddningstjänster uttrycker fortfarande ett motstånd mot samverkansformen. Hindren är dels organisatoriska, dels kopplade till IT-stödet för utlarmning, och är ofta interagerar de.
Idag är civila insatspersoner mer aktuella än någonsin och ett antal räddningstjänster inför konceptet under 2023. Brandskyddsföreningen har tagit på sig en koordinerande roll för att stödja räddningstjänster i processen. Ökad krisberedskap och ett civilt försvar har också aktualiserat frågan om civila insatspersoner har en roll att spela även här. Den här studien är baserad på över 10 års forskning och erfarenhet runt konceptet civila insatserpersoner. Vi presenterar tre olika modeller som tillsammans kan stödja lokala initiativ och sprida konceptet nationellt. Två av modellerna är lokala (mikronivå) och är anpassade mot landsbygds- respektive storstadsmiljö. Den tredje är övergripande (makronivå) och ska kunna erbjuda stöd till de lokala modellerna och koordinering av civila insatspersoner. IT-stödet spelar en central roll i samtliga modeller och studiens resultat diskuteras i termer av implikationer för digitaliserad samproduktion
Trafikolyckor under polisens utryckningskörning
No full textPolisens utryckningskörning innebär alltid en ökad risk för olyckor. Syftet med denna studie var därför att öka kunskapen om trafikolyckor som skett under polisers utryckningskörningar. Resultaten i presentation kommer från en intervjustudie med sex poliser som varit inblandade i minst en olycka under brådskande ärende eller trängande fall. Intervjustudien ingår som en av tre delstudier i ett större projekt som slutfördes under 2022. Resultaten visar bland annat att förarutbildningen skiljer sig åt mellan olika orter och att det finns en brist på fort-/vidareutbildning. Formella mentorer saknas men äldre och mer erfarna kollegor stöttar ändå ofta de yngre. Förutom att nyutexaminerade poliserna ofta är unga, och ännu inte hunnit skaffa sig erfarenhet inom polisyrket, är många av dem även oerfarna bilförare. Samtliga deltagare körde oftare i brådskande ärende än i trängande fall där de använde blåljus och siren, åtminstone delar av körningen. Att köra i trängande fall upplevdes som en positiv kick, men även som påfrestande och riskfyllt. Det stresspåslag som upplevdes behövde dessutom inte bara vara kopplat till själva körningen utan kunde även vara kopplat till ärendet. Deltagarna hade tillsammans varit förare vid sju olyckor. Två av olyckorna inträffade under körning i brådskande ärende (förföljande av oskyddade trafikanter) medan fem inträffade under körning i trängande fall (två singelolyckor i cirkulationsplatser, två korsningsolyckor och ett förföljande). I ett av de trängande fallen användes varken blåljus eller siren, i ett av fallen enbart blåljus och i de resterande tre fallen både blåljus och siren. Bidragande orsaker till olyckorna ansågs dels vara yttre omständigheter utanför Polismyndighetens kontroll, dels bristande tajming när informations ges samt en balansgång mellan yrkesutövande och trafiksäkerhet. Bidragande orsaker kopplade till föraren ansågs framför allt handla om bristande erfarenhet. Slutligen framhöll flera deltagare att det är viktigare att över huvud taget komma fram, än att komma fram så fort som möjligt. Utifrån dessa resultat kommer ett antal rekommendationer att presenteras