Infotech Oulu Annual Report 2015 - Biomimetics and Intelligent Systems Group (BISG)

Professor Juha Röning and Dr. Heli Koskimäki, Faculty of Information Technology and Electrical Engineering, and Professor Seppo Vainio, Faculty of Biochemistry and Molecular Medicine, University of Oulu
juha.roning(at)oulu.fi, heli.koskimaki(at)oulu.fi, seppo.vainio(at)oulu.fi
http://www.oulu.fi/bisg

Background and Mission

Biomimetics and Intelligent Systems Group (BISG) is a fusion of expertise from the fields of computer science and biology. In BISG, our basis are intelligent systems and our research areas include data mining, machine learning, robotics, and information security. More precise research topics vary from data mining algorithm development and optimization of industrial manufacturing processes all the way to environmental monitoring with mobile robots.

Bringing expertise from ICT and Biotech together we will reach the skills to make use of the mechanisms common in information processing and the biological data processing system and extrapolate this to intelligent solution making in ICT. One important goal of this program is to be able to physically link living cells via identified signaling systems to establish learning complex that involves Bio and ICT in a unified bifunctional interactive machine.

The group consists of four sub-groups: Data Analysis and Inference Group, Developmental Biology, Robotics and Secure Programming

We have conducted basic research in intelligent systems and developmental biology for over ten years as individual groups. Now we have joint our efforts. Our team consists of two professors, 10 post-doctoral researchers and 15 doctoral students. The annual external funding of the group is more than two million Euros, in addition to our basic university funding. There have been two completed doctoral degrees from the group. From the research of the group, eleven spin-out companies have been established so far: Codenomicon, Clarified Networks, Hearth Signal, Nose Laboratory, Nelilab, Atomia, Indalgo Probot, Aquamarine Robots, Radai and IndoorAtlas.

We co-operate with many international and domestic partners. In applied research, we are active in European projects. In addition, several joint projects are funded by the Finnish Funding Agency for Technology and Innovation (Tekes) and industry. We were a research partner in the, Internet of Things (IoT), SIMP and CyberTrust SHOKs. Prof. Juha Röning was selected as ACO (Academic coordinator) of the Cyber Trust program.

We are active in the scientific community. For example, Prof. Juha Röning is acting as visiting professor of Tianjin University of Technology and as the Robot Science Adviser of Tianjin Science and Technology Center for Juveniles. He served as a member of the Board of Directors in euRobotics and as a member of the SAFECode International Board of Advisors, and as a chief judge in the euRathlon 2015 (air+land+sea) competition, which took place Piombino, Italy, from the 17th to 25th September. He chaired the euRathlon / SHERPA Summer School 2015 in Oulu, Finland, 1st to 5th of June. It was a five-day course to provide participants with a full overview and hands-on experience with multi-domain real robotic systems. He also chaired with prof. Othmane the First International Workshop on Agile Development of Secure Software (ASSD’15) in Toulouse 24th of August 2015. In Tekniikan päivät (Tampere 23-24.10.2015) he lectured about Cyber Security. Prof. Seppo Vainio has been the chair in the Personalized Medicine day (2015), course on 580402S Biomedical imaging methods (1-­4 ECTS; 2015) and BIG data seminar (2015. Prof. Seppo Vainio is part of a European nanotechnology ”HyNanoDend” network.

During the reporting year, the group organized the 8th International Crisis Management Workshop and Winter School (CrIM’15), which brought together both Finnish and international information security experts. The group also organized Big Data Seminar 14th of September bringing together people interesting genome, and morphogenesis.

Scientific Progress

Intelligent Systems Incorporating Security

Within the Biometics and Intelligent Systems Group, the Oulu University Secure Programming Group (OUSPG) has continued research on security and safety in intelligent systems. Security and safety challenges in intelligent systems are threefold: increasing complexity leads to unforeseeable failure modes, quality is not the priority and awareness is lacking. We have approached the challenges from these three directions in our research.

Complexity - Model Inference and Pattern Recognition: we work under the premises of unmanageable growth in software and system complexity and emergent behaviour (unanticipated, not designed) having a major role in any modern non-trivial system. We have worked on natural science approaches to understanding artificial information processing systems. We have developed and applied model inference and pattern recognition to both content and causality of signalling between different parts of systems.

Quality - Building Security In: software quality problems, wide impact vulnerabilities, phishing, botnets, and criminal enterprise have proven that software and system security is not just an add-on, despite the past focus of the security industry. Instead, security, trust, dependability and privacy have to be considered over the whole life-cycle of the system and software development, from requirements all the way to operations and maintenance. This is furthermore emphasized by the fact that large intelligent systems are emergent and do not follow a traditional development life-cycle. Building security in not only makes us safer and secure, but also improves overall system quality and development efficiency. Security and safety are transformed from inhibitors to enablers. We have developed and applied black-box testing methods to set quantitative robustness criteria. International recognition of the Secure Development Life Cycle has provided us with a way to map our research on different security issues.

Awareness - Vulnerability Life Cycle: Intelligent systems are born with security flaws and vulnerabilities, new ones are introduced, old ones are eliminated. Any deployment of system components comes in generations that have different sets of vulnerabilities. Technical, social, political and economic factors all affect this process. We have developed and applied processes for handling the vulnerability life-cycle. This work has been adopted in critical infrastructure protection. Awareness of vulnerabilities and the processes to handle them all increase the survivability of emergent intelligent systems for developers, users and society.

These research goals are reached through a number of research activities.

Secure Software Development Lifecycle as a part of the Cyber Trust project - we approach all three goals by researching practical ways of building security into Secure Platforms, Cloud Computing services and Critical Infrastructure, from the design phase to actual operational use (Figure 1).

Figure 1. Dependencies of a single cloud based web service visualized by technology and location.

Situational Awareness in Information and Cyber Security aims to understand critical environments and accurately predict and respond to potential problems that might occur. Networked systems and networks have vulnerabilities that present significant risks to both individual organizations and critical infrastructure. By anticipating what might happen to these systems, leaders can develop effective countermeasures to protect their assets (Figure 2).

Figure 2. Port scanning visualized in an industrial automation network.

Coverage based robustness testing: Modern web browsers are feature rich software applications available for different platforms ranging from home computers to mobile phones and modern TVs. Because of this variety, the security testing of web browsers is a diverse field of research. Typical publicly available tools for browser security testing are fuzz test case generators designed to target a single feature of a browser on a single platform. This work introduces a cross-platform testing harness for browser fuzz testing, called NodeFuzz. In the design of NodeFuzz, test case generators and instrumentation are separated from the core into separate modules. This allows the user to implement feature specific test case generators and platform specific instrumentations, and to execute those in different combinations. During development, NodeFuzz was tested with ten different test case generators and six different instrumentation modules. Over 50 vulnerabilities were uncovered from the tested web browsers during the development and testing of NodeFuzz

Identification of a protocol gene: this research, PROTOS-GENOME, approaches the problems of complexity and quality by developing tools and techniques for reverse-engineering, and identification of protocols based on using protocol genes - the basic building blocks of protocols. The approach is to use techniques developed for bioinformatics and artificial intelligence. Samples of protocols and file formats are used to infer structure from the data. This structural information can then be used to effectively create large numbers of test cases for this protocol. In 2015, the project further developed the existing methodology, resulting in improvements in efficacy and discovering a number of vulnerabilities in web browsers.

Internet of Things studies security and privacy issues in large-scale sensor networks. Topics of interest are alternative ways of authentication, such as proof of work and cryptocurrencies, secure update mechanisms, software defined networking and related service-level agreements for data centres. The work brings together our research themes of Quality, Complexity and Awareness to an application area where resource limits are combined with global connectivity.

Privacy and Security and Online Social Networks

Exploiting Social Structure for Cooperative Mobile Networking (SOCRATE), a two -year (2015-2016) Tekes funded project under the Wireless Innovation between Finland and U.S. programme WiFiUS [http://wifius.org/], is a collaboration between the University of Oulu (co-PI Dr Ulrico Celentano), VTT, Aalto University, Arizona State University, and University of Nevada. During 2015, BISG has focused on architecture and on privacy and security issues in online social networks data mining.

Knowledge about the social structure of network users can be exploited to optimise radio network operations (Figure 3).

Figure 3. A conceptual framework supporting optimisation of mobile networks by exploiting social structure of physical users. (Celentano et al. 2016.)

To this end, personal and sensitive information about users should be collected. Clearly, this needs to be done in the fullest respect of the users’ privacy. Privacy protection is important not only to protect individuals’ intimacy, but also to prevent identity theft: related threats may finally expose the system to denial of service or sabotage. The solution to overcome the above and minimise the risks associated with potential threats, includes properly specifying the data structure by partitioning attributes, managing access rights and the related keys, and designing the topology of repositories (Figure 4).

Figure 4. Data distributed across repositories and partitioned into domains, with access and content management. (Celentano and Röning 2015.)

Results in this area have been presented at WF-IoT 2015 (Celentano and Röning 2015). Furthermore, the University of Oulu was the editor of a project-wise joint magazine article (Celentano et al., Manuscript) currently under review.

Intelligent Systems Incorporating Machine Learning and Data Mining

GlobalRF: Two-year Tekes funded project, the Global Spectrum Opportunity Assessment (GlobalRF) finished at first quarter of year 2015. GlobalRF was a collaborative research project with a joint effort undertaken by WiFiUS (Wireless Innovation between Finland and U.S.), leveraging research and education sources in Finland and the U.S. in the area of wireless communications. The collaborating institutions have been the Illinois Institute of Technology (IIT) and the Virginia Polytechnic Institute and State University (Virginia Tech) in the U.S., and the VTT Technical Research Centre, Turku University of Applied Sciences (TUAS), and the University of Oulu in Finland. All these institutions have ongoing research and education programs in wireless communications, and have brought significant expertise and resources to the proposed project.

During 2015 we have continued bi-monthly teleconferencing between participants in Finland and US. The data analysis and inference group has been working with large-scale statistical analysis in the GlobalRF project where a fundamental radio frequency (RF) spectrum shortage problem is tackled by developing methods for understanding the current and evolving use of the spectrum in various environments. We have been concentrating on modelling of human behaviour aspects of radio spectrum usage during mass events such as football and baseball games, athletics competition, and music festivals. More specifically, we try to find different variables and their impact explaining the sudden and normal changes in spectrum measurements.

We have been using Bayesian hierarchical regression where both individual (e.g., event on/off, time of day) and group level variables (e.g., day of the week, frequency band, measurement site) as well as uncertainties related to them can be modelled jointly.

Furthermore, the research in the area has included the development of big data management, processing, and visualization tools, as well as building predictive models to realize novel ways and guidance for dynamic sharing of spectrum usage. Conversely, RF spectrum measurement (and open datasets), intelligent data analysis and machine learning algorithms could provide novel ways to model environmental and human related contextual variables in urban city areas. Several RF measurement units are installed in downtown Chicago in US (see Figure 5), Blacksburg in US, and Turku in Finland (see Figure 6), as well as several mobile measurement units are used, all producing ongoing data for analysis.

Figure 5. WifiUS project RF spectum observatory in Turku (courtesy of TUAT).

Figure 6. WifiUS project RF spectum observatories in downtown Chicago (courtesy of Dennis Roberson, IIT).

Data mining methods for steel industry applications: BISG is a member of the Centre for Advanced Steels Research - CASR, which is one of the interdisciplinary umbrella organizations of the University of Oulu. Year 2015 was the second year in participation to a large national research programme System Integrated Metals Processing – SIMP.

One of the main goals in SIMP programme has been the development of an innovative supervisor system to assist the process development personnel and the operators of a steel production line over the whole production chain, and to help discover new alternative solutions for improving both the products and the manufacturing process. The tool is based on statistical models that predict different quality properties and rejection risks in several process steps, and it provides also model visualization. Currently, models predicting profile properties for steel strip and roughness properties for stainless steel strips have been implemented into the tool. During 2015, the online tests of the tool were started at Outokumpu, Tornio and offline tests at SSAB, Raahe.

The tool was developed in co-operation with VTT, where our contributions include creating predictive models and analysis tools for the steel production process, and VTT is mainly in charge of developing a platform to access industry databases and represent the modelling results on a web-based user interface. In Figure 7, three different operation situations of the tool have been presented. The user can observe the production in different process states and select one or several quality predicting models for viewing. Good quality is indicated with green and alarms can be seen in yellow or red colour depending on the probability of the rejection.

Other research topics active in our research group during 2015 were optimization of the steel plate tempering process based on the rejection probability models for strength and toughness, and steel plate shape analysis, where the goal was to find a measure that characterizes the concavity of the sides of the plate.

SIMP programme will continue for another 18 months, and we will present our research results annually on SIMP and FIMECC seminars around Finland as well as at international publishing venues.

The book covering the research history of the SSAB Europe in Raahe (formerly Rautaruukki Oy and then Ruukki Metals, Raahe Works) was published in 2015. BISG participated in the writing process based on our over 15 years’ long co-operation with the company. The book ”Rautaa ja terästä, 50 vuotta terästutkimusta” (in Finnish) was edited by Veikko Heikkinen (Figure 8).

Figure 7. Quality monitoring tool was developed in co-operation with VTT.

Figure 8. The research history of SSAB Europe was presented in a book edited by Veikko Heikkinen.

Uncertainty of classification results caused by missing data: Many real-world data sets contain missing data values. These might be the result of e.g. malfunctioning sensors or some measurements being too expensive to measure from every sample etc. Having missing values when classifying a sample means that there is an increase in uncertainty in the final classification result. Knowing how uncertain the result can sometimes be as important information as the classification result itself. In an Infotech doctoral program project, we are quantifying that uncertainty so that interpreting the classification results becomes easier.

Classification algorithms have traditionally been developed using complete data sets and most require values for all variables to be present to work. Many real world data sets are, however, cursed with missing data. To tackle this problem, we developed an algorithm that uses multiple imputation to handle the missing values. The algorithm can be used with any classifier that supports estimation of class posterior probabilities. The developed algorithm performs as well or even better as a benchmark algorithm (see Figure 7) and it does not require the classifier to support handling of missing values.

The uncertainty does not, however, behave consistently across different data sets. In a follow-up work we are addressing this issue and first results of this work are now under review

Data mining methods for human activity recognition: Wearable sensors based activity recognition is a research area where inertial measurement units based information is used to recognize human activities. The overall activity recognition process includes a data set collected from the activities wanted to be recognized, pre-processing incl. labelling, segmentation, feature extraction and selection, and classification. The activity recognition approaches can be used for entertainment, to give people information about their own behaviour, and to monitor and supervise people through their actions. Thus, it is a natural consequence of that fact that the amount of wearable sensors based studies has increased as well, and new applications of activity recognition are being invented in the process.

In year 2015, Pekka Siirtola successfully defended his PhD thesis, called “Recognizing Human Activities Based on Wearable Inertial Measurements - Methods and Applications” against his opponent Professor Barbara Hammer from University of Bielefeld, Germany. Thesis introduces new methods to recognize human activities, especially, when recognition is done using the sensors of a smartphone or when recognition is done in industrial context. Human activity recognition using wearable sensors has been one of the BISG’s main research interests for years and this thesis continues that work.

Moreover, in year 2015, the Academy of Finland funded postdoctoral research project of Dr. Heli Koskimäki, called MOVE (Mobile Sensors and Behaviour Recognition in Real-world) proceeded. The main interest of the project is on finding continuous, unvarying activity chains to discover performance of certain tasks (behaviour recognition). Another significant goal is to study the possibility of end-users to automatically or semi-automatically increase the amount of the behaviours to be recognized. This is a major question in many practical applications while, for example, in exercise area the initial recognition cannot include all the possible sport types. In addition, the behaviour recognition can be used for segmentation of data making possible to study the performance of tasks in more detail, for example, if the performance is comprised of repetitions of certain actions. This all is approached from the aspect that the methods can be used also online in real world.

In the year 2015 the project concentrated on the bias of classification accuracy caused by the back propagation step in leave-one-person-out cross-validation (see Figure 9). It was shown that the bias do not just effect to the classification accuracy itself but the selection of the optimal features as well as classifier and its parameters can be effected. This can be a major drawback when developing models into real world applications. Moreover, this similar effect can be anticipated also other areas where signals measured from humans are used.

Figure 9. Principles of basic leave-one-person-out cross-validation in activity recognition and three ways for train-validate-test approach: single division, 10-fold division and double leave-one-person-out cross-validation. The bias in basic leave-one-person-out cross-validation is due the back propagation marked with slashed red arrow. If there are no back propagation the basic scheme is adequate.

Foundations of knowledge discovery and data mining: Knowledge discovery in data (KDD) was defined in 1996 by Fayyad et al. as “the nontrivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data”. Although this definition still has its merits, it represents a rather narrow interpretation of the concept of knowledge that may prove a hindrance to the development of more advanced KDD tools. Dr. Lauri Tuovinen posited in his dissertation in 2014 that the data model underlying the KDD process should provide a formalization of the concept of knowledge that enables a computer to apply it autonomously, allowing the computer to perform KDD tasks traditionally reserved for humans. In 2015, work on this idea continued and new results were submitted for review.

Another aspect of the KDD process discussed in Dr. Tuovinen’s thesis is the actors of the process and the interactions between them. Under this theme, a long-term research interest of BISG that also ties in with information security is the ethical implications of KDD, such as the potential threat to privacy when mining personal data. This research area was active also in 2015, and a case report of recent work addressing the use of intelligent systems in health promotion interventions was submitted for review.

Model selection in time series machine learning: In autumn of 2015, Eija Ferreira defended her doctoral thesis "Model selection in time series machine learning applications". Her opponent at the defense was Professor Daniel Roggen from University of Sussex, UK. In the thesis, Ferreira discussed model selection in the context of three different time series machine learning application areas: resistance spot welding, exercise energy expenditure estimation and cognitive load when starting to solve a new machine learning problem. She also considered the special restrictions and requirements that need to be taken into account when applying regular machine learning algorithms to time series data. 

Intelligent Systems Incorporating Robotics and Cybernetics

euRathlon Summer School

The euRathlon SHERPA summer school 2015 was organized from the 1st to 5th of June mainly by the robotics group members at the department of Computer Science and Engineering. The summer school was attended by 42 students, mostly doctoral, originating from 10 different countries (Figure 10). Also, eight invited lecturers, mostly from SHERPA, held lectures during the summer school. Overall, based on the satisfaction survey after the event, the participants were especially pleased with the overall organization of the event.

Figure 10. Early attendees in the euRathlon SHERPA 2015 summer school organized at the University of Oulu.

This year, the euRathlon summer school was a joint operation between different autonomous mobile robotics fields, the aerial (SHERPA, Smart collaboration between Humans and ground-aErial Robots for imProving rescuing activities in Alpine environments), the land and marine robots (University of Oulu, Probot Ltd. and Aquamarine Robots Ltd., Coppelia Robotics GmbH). In total, the summer school lasted for four and a half days consisting roughly 50% of lectures and 50% of practical exercises. At the beginning of the summer school, the students could choose their preferred field to study; aerial, marine or land. The students following the land robot exercises needed to implement control algorithms for the land robots with the aid of the intuitive Coppelia Robotics V-REP simulator before testing the algorithms in the real world environment. The robot server interface, where control algorithm software client implemented by the students in Python language connected, was the same for both the simulator and the real robots. The drag and drop robot in the simulator world and the real robots could then be controlled by the same implemented client side control software (Figure 11–12).

Figure 11. V-REP Simulator environment running on Porteus Linux and the Navigation UI client program connected to the virtual robot.

Figure 12. The outdoor land robots that were assembled for testing the control implementations made by the students; Mörri on the left side and C-frame built from modular robot building blocks, provided by Probot Ltd, on the right side.

The quadrotor aerial drone exercises consisted of implementing control algorithms for the drone (Figure 13), which was operated through software running on ground station PC. The control code was implemented in C# programming language, using predefined primitives, such as Goto, RotateYaw and TakePicture. The ground station communicated with the aerial drone via a Wi-Fi link, using MAVlink protocol. The students also had access to high level information of the quadcopter, such as drone altitude, position in the NED (North-East-Down) frame, status of the flight battery and more.

Figure 13. The flying drone utilized in the aerial robot exercises.

In the aquatic scenario, four teams were instructed to design path tracking and station-keeping algorithms for an unmanned surface vehicle. The challenges for the aquatic control algorithms followed from an unknown control model and varying environmental conditions (wind). In order to obtain successful implementations and to pre-test their algorithms, after some literature review to the state-of-the-art control algorithms, students started to implement their algorithms in a PC class room with simulator environment. In simulations, students were using Aquamarine Robot’s GUI to follow how their algorithms perform with different wind conditions. After their implementation was accepted in simulator environment, each team tested their implementation and made some modifications with marine robot Dolphin in a small lake (Figure 14).


Figure 14. The UI of the Aquamarine Robots Dolphin robot utilized in the marine robot exercises and the image of the Dolphin captured from above autonomously by the SHERPA flying drone (bottom picture).

The exercises were aimed to implement control algorithms for the robots in order to execute a co-operation scenario where the robots could perform joint actions using a central communications server. The scenario area was selected to be the nearby bay area (Kaijonlahti), where the marine, land and aerial robots could operate. The area used was approximately 350 * 350 meters, consisting mainly of a water area and of a grassy field with no trees (Figure 15).

Robotics Research

In 2015 BISG had a wide range of research in the area robotics, including industrial safety, aerial data gathering, battery life management and control of complex wheeled land robots.

ReBorn

In the EU funded ReBorn project, having participants from 17 industrial and academic institutions from 10 different countries, research related to improving the recyclability of industrial robots has been carried out.

Figure 15. The outdoor robot exercise area viewed from the Google Earth application.

The main focus of the University of Oulu in the project has been in the identification of typical use cases, the related standardization and the identification of potential improvement areas in the existing standardization. In industrial manufacturing, the work scenarios are becoming progressively more human-robot interaction and collaboration oriented. Therefore, the safety requirements in human-robot collaborative scenarios must be clearly defined and standardized to minimize the possibility of user injuries (Figure 16). Especially in collaborative scenarios, proper safety equipment and user aware control algorithms must be in place to eliminate the possibility of injury to humans operating inside collaborative work spaces. In addition to enhancing safety equipment, future product design applications require new virtual, cloud based and robot capability aware technologies to be employed for improved manufacturing interaction. These technologies can also enable faster product development cycles and facilitate higher level of product customization options on the production line.

Figure 16. One possible risk potential evaluation scenario of collaborative work spaces, each object is assigned a risk potential based on possibility and severity of injury. The total spatial risk potential is used in evaluation of the constraints used during robot movement planning and operation.

Matine

Study of ionizing radiation detection and sample collection with a DJI Inspire 1 quadcopter is being carried out in co-operation with STUK (Radiation and Nuclear Safety Authority), University of Helsinki and Finnish Defence Research Agency in a MATINE (Maanpuolustuksen tieteellinen neuvottelukunta) funded project. In initial phases, the quadcopter has been equipped with a Kromek GR1-A gamma-ray spectrometer (Figure 17–18) which is connected to a smartphone via USB. This smartphone is equipped with an application that was developed to collect radiation data from Kromek GR1-A, and send it to database via WLAN/3G (Figure 19). This combination is used for identification of radiation sources in outdoor environments. Initial testing has been performed for testing the feasibility of the suggested approach and the field tests may begin in the early 2016.


Figure 17. 3D printed casing implemented for attaching a gamma-ray detection sensor onboard a DJI Inspire 1.

Figure 18. Test spectrums collected with Kromek GR1-A spectrometer from low radiation (10 μCi) sources from a 1 cm distance. The main photon energy spike (662 keV) detected from a Cs-137 sample is shown in 1. The photon energy spikes at 1.17 MeV and 1.33 MeV detected from Co-60 sample are seen in 2. and 3.

Battery Management

First functional prototypes of the modular intelligent battery modules are being tested with mobile modular robot units developed with Probot Ltd. With hot-swappable, modular intelligent battery energy storages

Figure 19. An Android application was developed to gather data from Kromek GR1-A gamma-ray spectrometer and send it to database.

 (Figure 20–21), the energy can be transferred to the mobile platform modules from the switched payloads also during operation, enabling continuous uninterruptible operation. Battery state awareness and automated operational efficiency optimization are being developed related to these platforms for implementing more energy efficient mobile robot units for challenging environments. A general artificial neural network based Li-ion battery State-of-Charge estimation model has been developed to be used in battery energy management applications (Figure 22–25). Research is also being carried out for utilizing deep-learning neural networks in conjunction to predictive route planning for the mobile robot mission execution algorithms. Deep-learning based environment classification could be especially useful in co-operation applications of mobile ground platforms and unmanned aerial units that can survey and classify the ground robot operating area from above.

Figure 20. Insides of the intelligent battery module with a separable DC-DC converter module used for battery charging that can be located inside or outside of the battery module. Internal charger is useful for improving system flexibility while external charging schemes are more cost efficient solutions.

Figure 21. Real-time measurements done by the intelligent battery module onboard a mobile robot.

Figure 22. Basic normalized measurement inputs obtainable from a Li-ion battery system; voltage (blue), temperature (red), power (green). This data was collected from a 2012 model Mitsubishi I-Miev full electric car. Later, dataset 1. was used as training data and 2. as validation and testing data.

Figure 23. Artificial neural network based State-of-Charge (SoC) estimation system that uses basic measurement inputs available from most battery measurement systems. In addition to a neural network, input space is expanded with battery voltage behavior related operators f.

Figure 24 Battery state estimation model output êSoC (blue) vs. the post calculated real SoC (green).

Figure 25. Error between model estimated and real SoC.

Control of Complex Wheeled Robots

Pseudo-omnidirectional robots with individually steer-able wheels offer a good balance between payload, robustness and mobility.  However, the non-holonomic nature of the regular wheels and the often redundantly actuated structure of these robots make their control a complex issue. This complexity of control is further exacerbated when the wheels are not rigidly connected to the robot body but are instead connected via actuated chains which allow the wheels move relative to the body. BISG has developed control algorithms for such Articulated Wheeled Vehicles (AMW). The control algorithms are mathematically simple closed-form analytical functions and are thus computationally light but are currently limited to planar cases. The computational load is only linearly dependant on the number of wheels making the developed control algorithm suitable for multi-wheel configurations and/or low-powered embedded MCUs. The control algorithms synchronize the rolling and steering velocities of complex planar robots (example in Figure 26) with freely located wheels forming fixed or variable (Figure 27) footprints. The rolling and steering velocities remain synchronized even with very complex motions of the robot (Figure 28). With the developed control algorithms, the traversable path, robot’s heading on different points of the path and the path velocity can be controlled separately, thus offering great freedom on how to control the robot on a given practical task. The control algorithms do not in practice suffer from representation singularities which are a common problem in wheeled control. The control algorithms also compensate for the proximity of mechanical singularities by adjusting the robot’s path velocity according to the maximum capabilities of its wheels’ steering and rolling actuators. In fact the developed control algorithms are time optimal in a sense that at any given moment the robot is either traversing with maximum allowed path velocity or at least one of its steering or rolling actuators is turning at its maximum velocity (Figure 29), i.e. the robot traverses the given path in the given way with the given velocity restrictions as fast as it possibly can.

Figure 26. Example of complex wheeled planar robot.

Figure 27. Example of wheels’ motion with respect to the robot body. The colored lines depict the wheel paths as the robot body traverses a straight line.

Figure 28. Simulation run of Figure 26’s robot traversing a given path (blue dots) while keeping its front directed at all times to a point of interest (green larger dot). The path was 82 meters in length.

Figure 29. (Top) wheel rolling speeds, (Middle) wheel steering speeds and (Bottom) robot path velocity for the first 30 seconds of Figure 19’s simulation run.

In summary, the developed control algorithms can be used in a wide range of robot configurations and scenarios with low computational cost. The control algorithms are currently limited to planar surfaces and can cause sudden and large changes in velocity (Figure 28 bottom) and the control algorithms are being extended to work also with uneven surfaces and limited accelerations.

Magnetic Field Localization and SLAM

The objective of our indoor localization research is to develop methods for exploiting indoor magnetic field variation in positioning and mapping. The idea is based on analysis with various indoor magnetic field datasets showing that indoor magnetic fields provide sufficient spatial variation and temporal stability to permit inference about sensing locations, given noisy measurements. In recent years, we have published various papers studying magnetic field localization and related methods in robotic and human contexts (Figures 30 and 31).

In 2015 we continued our magnetic field SLAM exploration studies. SLAM exploration refers to the methods where we try to find optimal ways to collect magnetic information for mapping purposes, meanwhile, simultaneously use this information for localization purposes. We have developed new ways to model magnetic field information using spectral Gaussian processes. We also have developed methods for efficient action selection using environmental partition based on information similarities. The results will be published on autumn 2016.

Based on the magnetic field localization studies, a new start-up company, called Indoor Atlas Ltd., was founded in 2012. This company offers indoor positioning technologies for various application areas. The company has generated high interest in international technology magazines.

In our research on magnetic field simultaneous localization and mapping (SLAM), we have put a strong emphasis on light-weight methods running entirely on mobile platforms, such as Android smartphones and tablets. Compact map representation and effective algorithms are essential when using devices with very limited resources, and we have developed methods to tackle the problems arising from very sparse data and high uncertainty levels produced by low-cost and noisy smartphone sensors. Our work is continuing toward an autonomous mobile robot system based solely on smartphone sensors that is able to intelligently build a map of the magnetic environment (Figure 30).

Figure 30. Magnetic landscape (bottom) of University of Oulu Discus entrance hall (top-left). The landscape illustrates the spatial variation of the magnetic field that allows magnetic field-based localization and mapping.  The map is created by an iRobot Create robot (see Figure 31) in a simultaneous localization and mapping (SLAM) experiment.

Figure 31. An iRobot Create collecting data for simultaneous localization and mapping (top-left) and the resulting map of the magnetic field (top-center). The bottom row visualizes that the robot can correct its path by using the magnetic information (bottom-left). Without magnetic data, the odometry is highly erroneous (bottom-center). If equipped with similar hardware, such as smartphones or tablet computers, both humans and robots can utilize the same magnetic maps (right).

Social robot scenarios are particularly difficult because of the dynamic (often crowded) environment. Magnetic field localization is not affected by surrounding people like laser scanners and cameras for example, and it is therefore very promising in these kinds of scenarios. While our method is used to localize the robot, the other sensors can be assigned to handle the social tasks. We have also developed localization methods that are usable by both robots and humans equipped with similar mobile devices (Figure 31).

Efficient Systematic Sampling from a Discrete Distribution

In order to improve the sampling process in the magnetic field SLAM motion model, we have developed an efficient and low-variance Systematic Alias Sampling (SAS) method based on the alias method by Walker. The method produces batches of samples from an arbitrary discrete distribution by systematically drawing from the alias table structure (see Figure 32). SAS produces samples up to 20 times faster than the compared sampling method in a popular Java mathematics library (Apache Commons Math). In addition, the Cramér-Von Mises statistic shows that SAS produces much more fit empirical distributions than i.i.d. sampling, if the problem of almost divisibility between bin and sample count is carefully taken care of.

Figure 32. Top: Discrete 101-valued approximation of the standard normal distribution with support of [−4,4] denoted as N101. Middle: Alias table structure generated from N101. The probabilities of selecting the aliased values are depicted as the upper part of the bins. The colors correspond to values in the topmost figure. The black dots are an example batch of 16 systematic samples. Bottom: The empirical distributions defined by the 16 systematic alias samples (SAS) and 16 i.i.d. samples compared to the true cumulative distribution function.

Evolutionary Robotics

During 2015 we published the results of study considering the Lego Mindstorms NXT platform’s suitability for evolutionary robotics by using a genetic algorithm to evolve a neural network-based controller for Lego sumo wrestling. The genetic algorithm was able to evolve a controller with simple but effective strategy for the task. The emerged behavior is illustrated in Figure 33. A video of the evolution can be found at: https://www.youtube.com/watch?v=PaOADqerpWU.

Furthermore the Lego platform has been utilized in collaboration with local schools and students, e.g. in the form of workshops. The experiments on the platform show that the tools already used in school education can be utilized to create hands-on experiences e.g. on the principles of evolution for the students.

Figure 33.  Evolutionary robotics on Lego NXT platform. A typical example of the emerged behavior during generation 100. Both robots start scanning the environment by turning left. Soon the red one gets the yellow one in sight and is able to charge towards the opponent, pushing it outside the ring, resulting in an easy victory.

The Evolutionary Active Materials

The Evolutionary Active Materials (EAM) project, which is funded by the Academy of Finland, is a joint effort between the Computer Science and Engineering laboratory (CSE) and the Microelectronics and Materials Physics laboratories. The aim of the EAM project is to develop novel, evolutionary computation (EC) based design methods for active and versatile materials and structures. The first components are being developed through a novel holistic design process utilizing constantly increasing computation power, the development of multi-physics simulators, and EC techniques, such as genetic algorithms.

During 2015, the height and the top diameter of Cymbal type piezoelectric actuator were optimized by genetic algorithm and FEM modelling. From the optimized results, maps of electromechanical capabilities of different structures were generated. The blocking force of the actuator was maximized for different values of displacement by optimizing the height of the cap and the length flat region of the end cap profile. By using values obtained from a genetic algorithm optimization process, a function was formulated for design parameters. Using the function, a map of displacement, the steel thickness and the height of the end cap the optimized length of flat region was constructed (Figure 34). A similar map with the length of the flat region for the optimized height of end cap was created. The results will be published at 2016.

Figure 34. The top diameter of the steel cap as a function of steel thickness and displacement for Cymbal.

New type of actuator called Mikbal (Figure 35) was invented, optimized with genetic algorithm and realized. Mikbal was developed from Cymbal by adding additional steel structures around the steel cap to increase displacement and save the amount of used piezoelectric material. The best displacement to amount of used piezo material ratio was achieved with 25 mm piezo material diameter in the case of 40 mm steel structures, and lower height and top diameter of the cap increased the displacement. The results will be published during 2016.

Figure 35: The von Mises stresses in Mikbal actuator under 500 V voltage.

Also optimization of the end cap structure of the Cymbal type energy harvester was done with genetic algorithm and FEM modeling software Comsol Multiphysics. The aim was to improve harvested power levels from human walking (Figure 36).  The power produced by the energy harvester was increased by allowing the algorithm to modify thickness in certain regions as grooves in the end cap. By evolution of the structure, power produced by the harvester increased by 38 % compared to traditional linear type Cymbal harvester which was also optimized by the algorithm. Increase in power was obtained by change of mode in mechanics of the harvester by grooves.

Figure 36. Cymbal type energy harvester in a shoe and an optimised profile for the harvester. In the profile piezoceramic disc is depicted in yellow and steel cap in grey. The grooves shown in the left side of the profile have been found by the genetic algorithm.

Intelligent Systems Incorporating Bio-IT Solutions

Developing novel real-time biosensors for glucose monitoring. For developing “second generation biosensors” we have taken use of our skills to purify and culture the skin derived progenitor cells that are responsible in skin renewal and regeneration. We obtained for the project a Tekes strategic opening funding. With this support we have advanced the work to develop of a novel biosensor strategy (Figure 37).

Figure 37.  Novel biosensor strategy. Donor skin renewing cells are set to culture and a specific responsive component is engineered to target a tag to the 3´end of the coding sequence in the genome. Such a cell is then implanted to the donor to serve as a measure for a given physiological parameter. 

By now we have been able to conduct the proof of principle set up in the sensor construction.  These indicate that the skin is indeed responsive to the changes in certain serum constituents. The data also indicated that the cells with in the skin can also be engineered and be converted genetically to serve as biosensors. We have screened in selected biological phenomena with the proteomics and transcriptomics the respective mediator signal transduction pathways serving as a read out. We identified wealth of candidate factors whose genes are currently being engineered to convert the respective protein into an isoform whose activity can read with an external device.

We have also tested the capacity to culture of FACS purified cells of the skin and also if such cells can be transplanted with a fluorescent tagged cells to the donor so that the cells indeed become incorporated. We assayed the stability of the sensor transplant. The data suggest that a syngeneic host suggesting that the aimed biosensor strategy is feasible accepts the skin progenitor graft.

In collaboration with VTT we have also developed the electronic unit that has the capacity to measure the changes in the skin basal progenitor cell integrated sensor. We are currently in a process of filing a patent of these biotechnological avenues with VTT.      

Developing an ex vivo supernatural personal mobile biosensor device. To advance the goal to develop wearable sensory devise we started to assemble first via a HILLA funded project a micro fluidistic set up that will be converted to a bio recognition tool. During the research period several micro fluidistic prints were planned, made and tested. Out of these a configuration was obtained that collected successfully the skin associated fluids as depicted by the presence of color dye in the fluidistic chamber (Figure 38). A patent search of the strategy has been conducted. 

We also developed capacity to the micro fluidistic set up to monitor specific biomolecules present in the skin fluids. We are currently advancing the research line and aim to obtain more throughput capacity. To achieve this we are planning and printing array format fluidistic champers. In parallel to these developments to be able to read the fluorescence that is revealed by specific antibodies against certain factors is currently developed in collaboration with VTT. A mobile phone based micro fluidistic reader was achieved and the reliability of the phone based reading with associated programs was developed during 2015 there in. 

Screening of electromagnetic and optogenetic responses in organs generated from stem cells. The genetic engineering offers opportunities to developed technologies where the cellular in or output signals can be regulated by certain wavelengths in the electromagnetic spectrum. Alternatively the cellular actions can be genetically constructed so that a signal will be transmitted to a biosensor that will convert it to a form readable by an electric device. 

Figure 38. A micro fluidistic print design is able to collect the skin-associated fluids as depicted by the accumulation of a blue indicator dye in the chamber.

During 2015 we developed novel tissue engineering technologies that do enable introduction of specific gene expression constructs to individual cells of the model organ such as the mammalian kidney. Here the organ primordia is dissociated to single cells, the genetic construct encoding the protein of interest such as the optogenetic or the radio responsive component is transduced to such a cell with a reporter for read out. There after the organ is let to self-assemble and placed for a long-term culture (Figure 39).

Figure. 39. An organ primordia can be dissociated to single cells, the constitute cells transduced with a genetic construct to acquire optogenetic and radio genetic guidance capacity to the morphogenetic cells ex vivo.    

With the developed model systems we have taken use of the image analysis technologies to visualize how the morphogenetically active cells behave in three dimension (3D). To achieve this we applied defined pressure to the assembled organ primordia in ex vivo setting.

We found that under a defined pressure the organ flattens towards two dimension (2D) but yet morphogenesis progressed (Figure 40). This novel set up has made it possible follow the fate of individual cells is the cells are constricting a detailed manner while the natural form.

Figure 40. The 3D kidney organ primordia develops under a mild pressurize in ex vivo conditions converted towards the 2D configuration.  The setup has enabled to identify pressure sensors in the cells and also to develop novel organ pressure monitoring tools. 

To target the detailed dynamics by which the form is assembled in a model organ we took use of the genetically engineered Wnt4CreGFP knock in mouse model. This was crossed to the floxed Rosa26 Yellow Fluorescent Protein (YFP) transgenic mice. In this genetic crossing the stem cells that generate whole of the nephron will become labeled with the YFP.

We have captured 3D movies from the developing kidney with the confocal microscope in a time-lapse setting.  We are in a process of analyzing the detailed cell behavior via the machine learning/computer based image analysis with Prof. Janne Heikkilä. With Dr. Jari Juuti we aim to construct a specific device that allows detailed measure of the pressure forced encountered by the tissue undergoing morphogenesis.  These novel capabilities now allow analysis in great detail the mode by which the spatial and temporal organization of the cells go on to construct natural form that is open at present in any developing organ system. We will use models to identify the pressure sensors from the cells with the OMICS technologies. 

Developing high throughput robotic aided platforms to screen complex cellular responses to magnetic/electric fields via signaling pathway reporters. To advance the strategies to measure in a high through put manner the cellular responses to stimuli we have assembled a bio robotic workstation. Here an Operetta confocal microscope was obtained and this was coupled to a hood that contains an automated plate-cargo arm, a rack for the plates with a bar code reader and incubator for long term exposure of the cells to compounds such as drugs or specific electromagnetic spectral radiation (Figure 41). The Operetta confocal microscope has machine learning/image analysis capacity for wealth of measurements to be conducted from the cells.

Figure 41. Operetta confocal workstation coupled to a robotic set up and an incubator was assembled. A) A holder for plates and transported by the robotic arm (B) and the cells with in will be transported to an incubator (C ). The whole set up is inside a hood (D) and the robotic arm transports the plates to the Operetta confocal semi-high throughout microscope fluorescent reader. The data is analyzed by wealth of machine vision/image analysis programs present with in the assembled bio robotic set up. The bio robotic core facility will be used to screen with a library of live indicators cellular response to specific frequencies in the electromagnetic spectra.       

To take use of the set up a yeast cell library was obtained and three replica clones from it was generated and stored for later use. The library is composed of cell where each of the 3´end of each of the yeast gene was targeted by a green fluorescent protein (GFP) tag. The next goal is to obtain capacity to start to use the set up to define the oscillating properties of the cellular genes and to use it as live measures for screening responses to stimuli such as those mediated by the opsins for the visible light frequencies. 

Intelligent Systems with cohort data sets: Cohort data set is a special data set from the medical domain, which has not been studied with a machine learning approach before. The data set, Northern Finland Birth Cohort 1966 (NFBC 1966), is a unique data set with over 14 000 original variables in various yet heterogeneous formats (numerical, ordinal, categorical, images, text etc.) from a population of over 12 000 mothers and their children without any complete data points. The amount of variables raises to millions if genetics and epigenetics are considered (p >> n).

There are two extremely important aspects of modeling this type of data: confidence of the predictions made with the model and model interpretability. Steps towards instance level confidence estimates have been made in our previous work (see above) and we will continue to pursue this goal, along with keeping model interpretability in focus also, when we start digging into this fascinating data set. Our goal is to use a machine learning approach to make novel discoveries from the data that traditional data analysis approach has not yet uncovered.

Elders are an increasingly large fraction of the population in developed countries. From one hand people expect an independent life also in presence of more or less important diseases. On the other hand the treatments to care those diseases, often together with co-morbidities, imply larger costs. To respond to both these goals, the disease progress should be kept as low as possible (see Figure 42), which means early disease detection, deinstitutionalisation and personalised medicine, striving to allow a better quality of life, a more cost-efficient healthcare system and a more inclusive access to healthcare both in developing countries and in remote areas in developed countries.

Figure 42. Progress of a disease (left), outcome (right) and access to healthcare. (Celentano and Röning 2015.)

Novel Bio-ICT technologies are needed to achieve these targets and BISG is active in this area in many fronts summarised below.

By tracking health status of large groups and including in the analysis a wealth of metrics and parameters, large amounts of data are generated. On the other hand, by downscaling biology-based technologies down to the nanoscale including sensing biological parameters directly from living cells, potential security threats are correspondingly moving into human bodies, but promising tools are offered for personalised medicine and treatments, including tight biological interaction, prostheses and their control (Celentano and Röning 2015). BISG is strong in all these areas (data analysis, security and robotics) and it is therefore pushing itself among the world leaders in this growingly important area.

Towards a Holistic Self-awareness in Humans and AI

Self-awareness is involved in a number of cognitive functions of the human brain and correspondingly its disturbances are part of a number of disease of various statistical relevance. Self-awareness may also improve the efficiency in robotic systems by a more conscious execution of tasks (Celentano and Röning 2016). Interaction among concurrent cognitive entities further expands the model in Celentano (2014). For multi-robot systems, various cognitive and social inputs can be used for self/nonself discrimination, including the observation of the self, of the environment and of neighbour entities, as well as the exploitation of social interaction among agents (Figure 43).

Figure 43. Self-awareness through observation of the self and of the environment, top, and communication among agents, bottom. Even if the final outcome (e.g., relative positions) may look similar, it is important to discriminate what I am doing from what the others are doing. Sharing knowledge among agents further improves awareness. (Celentano and Röning 2016.)

In line with BISG strategy, cognitive functions in humans and in artificial entities are in this research collaboratively studied side by side to allow exploiting the results in both domains bridging neuroscience and artificial intelligence.

Exploitation of Results

Within the framework of its Bio-IT theme, see above, BISG has recently started a cooperation with the SpAtial, Motor & Bodily Awareness (SAMBA) research group:[http://dippsicologia.campusnet.unito.it/do/gruppi.pl/Show?_id=hhuv] at the Department of Psychology of the University of Turin, Italy. The cooperation with Prof. Raffaella Ricci and colleagues, focuses on bridging neuroscience and artificial intelligence. This research aims at cross-fertilising the two scientific domains, continuing and strengthening the research paths currently active at respective sides

Several initiatives for EU projects including Horizon 2020 are ongoing and these efforts will be continued.

BISG will continue the cooperation with current US partners: University of Nevada, Arizona State University and Carnegie Mellon University. In particular, as a follow-up of SOCRATE cooperation, with the University of Nevada can be exploited complementary expertise in the area of multi-layer security.

The results of our research were applied to real-world problems in many projects, often in collaboration with industrial and other partners. Efficient exploitation of results is one of the core objectives of the national Digile and FIMEC ICT SHOK projects like SIMP, IoT and Cyber Trust; in these projects we work in close collaboration with companies throughout the projects.

During the reporting year, the group continued utilizing outdoor robotic systems. Development and utilization of Mörri, a multipurpose, high performance robot platform continued. More focus was put on perception in natural conditions, representation of detections, knowledge, and an environment model of the operating environment. The software architecture further developed the earlier work on Property Service Architecture, and the Marker concept as general purpose representation was further developed.

Future Goals

The partnership in the SIMP programme that belongs to the SHOK concept of Tekes enables us to continue our steel research into new areas. The new goals are in quality prediction at different process stages and for more challenging properties. As a result more advanced expert systems can be developed to aid the operators with different roles in steel making.

We will continue to strengthen our long term research and researcher training. We will also continuously seek opportunities for the exploitation of our research results by collaborating with partners from industry and other research institutions on national and international research programs and projects. The University of Oulu is a founding member of euRobotics. Juha Röning is a member of the Board of Directors of euRobotics.

We will strengthen our international research co-operation. With the University of Tianjin in China, we have a joint project in which methods and a system will be developed for vision-based navigation of Autonomous Ground Vehicles, which utilize an omni-directional camera system as the vision sensor. The aim is to provide a robust platform that can be utilized in both indoor and outdoor AGV (Autonomous Ground Vehicles) applications. This co-operation will continue.

In the USA, we will continue to co-operate with the Human-Computer Interaction Institute in Carnegie Mellon University with Assistant Professor Anind K. Dey. The research is on human modelling in the area of human-machine interaction. We continue and strengthen US-Finland co-operation through an NSF grants. Shorter research visits to European partners in EU-funded projects are also planned.

In 2016, the aim is to utilize more widely the know-how from sensor technology and data mining. New application areas will be studied, including rehabilitation, exercise motivation and energy efficiency in households, and the benefits of our expertise will be highlighted to actors in the areas.

In human-environment interaction and sensor networks, our research will continue. Our main goals are to develop analysis methods for sensor network data and to develop applications utilizing physical user interfaces. Research on novel software architectures, reasoning and knowledge representations will continue as well. Field trials in realistic settings, and close collaboration with research groups (national and international) and companies will be emphasized.

Personnel

professors

2

postdoctoral researchers

7

doctoral students

14

other research staff

7

total

30

person years for research

25

 

<--break->

External Funding

Source

EUR

Academy of Finland

201 000

Tekes

639 000

domestic private

96 000

international

206 000

total

 1 142 000

 

Doctoral Theses

Siirtola, Pekka (2015) Recognizing human activities based on wearable inertial measurements : methods and applications. Acta Universitatis Ouluensis, Technica C 524.

Ferreira, Eija (2015) Model selection in time series machine learning applications. Acta Universitatis Ouluensis, Technica C 542.

Selected Publications

Alasalmi T, Koskimäki H, Suutala J & Röning J (2015) Classification Uncertainty of Multiple Imputed Data 2015 IEEE Symposium Series on Computational Intelligence: IEEE Symposium on Computational Intelligence and Data Mining (2015 IEEE CIDM).

Bibikova O, Popov A, Bykov A, Prilepskii A, Kinnunen M, Kordas K, Bogatyrev V, Khlebtsov N, Vainio S & Tuchin V (2015) Optical properties of plasmon-resonant bare and  silica-coated nanostars used for cell imaging. J Biomed Opt. 2015 Jul;20(7):76017. doi: 10.1117/1.JBO.20.7.076017. PubMed PMID: 26230637.

Berry RL, Ozdemir DD, Aronow B, Lindström NO, Dudnakova T, Thornburn A, Perry  P, Baldock R, Armit C, Joshi A, Jeanpierre C, Shan J, Vainio S, Baily J, Brownstein D, Davies J, Hastie ND & Hohenstein P (2015) Deducing the stage of origin of Wilms' tumours from a developmental series of Wt1-mutant mice. Dis Model Mech. 2015 Aug 1;8(8):903-17. doi: 10.1242/dmm.018523. Epub 2015 May 14. PubMed PMID: 26035382; PubMed Central PMCID: PMC4527280.

Celentano U, Röning J, Ermolova N, Tirkkonen O, Chen T, Höyhtyä M, Yang L & Zhang J (Manuscript) Secure cooperation in social-aware cognitive D2D networks.

Celentano U & Röning J (2015) Framework for dependable and pervasive eHealth services. IEEE World Forum on Internet of Things (WF-IoT) [http://www.ieee-wf-iot.org/], Special session on Dependable IoTs for eHealth and management of chronic conditions. 14–16 Dec 2015, Milan, Italy.

Celentano U & Röning J (2016) Multi-robot systems, machine-machine and human-machine interaction, and their modelling. International Conference on Agents and Artificial Intelligence (ICAART) [http://icaart.org/], 24–26 Feb 2016, Rome, Italy.

Daniel E, Onwukwe GU, Wierenga RK, Quaggin SE, Vainio SJ & Krause M (2015) ATGme: Open-source web application for rare codon identification and custom DNA sequence optimization. BMC Bioinformatics. 2015 Sep 21;16:303. doi:10.1186/s12859-015-0743-5. PubMed PMID: 26391121; PubMed Central PMCID: PMC4578782.

Ferri, G., Ferreira, F., Sosa, D., Petillot, Y., Djapic, V., Franco, M. P., Wineld, A., Viguria, A., Castro, A., Schneider, F., & Roning, J (2015) euRathon 2014 marine robotics competition analysis. Eurocast 2015 Workshop on Marine Sensors and Manipulators, Las Palmas de Gran Canaria (2015).

Juutilainen I, Tamminen S & Röning J (2015) Visualizing Predicted and Observed Densities Jointly with Beanplot Communications in Statistics - Theory and Methods, 44:340-348.

Juutilainen I, Tamminen S & Röning J (2015) Density forecast based failing probability predictors in manufacturing European Journal of Industrial Engineering, 9(4):432-449.

Koskimäki H (2015) Avoiding Bias in Classification Accuracy - a Case Study for Activity Recognition IEEE Symposium on Computational Intelligence and Data Mining

Krause M, Samoylenko A & Vainio SJ (2015) Exosomes as renal inductive signals in health and disease, and their application as diagnostic markers and therapeutic agents. Front Cell Dev Biol. 2015 Oct 20;3:65. doi: 10.3389/fcell.2015.00065. eCollection 2015. Review. PubMed PMID: 26539435; PubMed Central PMCID:PMC4611857.

Maezawa Y, Onay T, Scott RP, Keir LS, Dimke H, Li C, Eremina V, Maezawa Y, Jeansson M, Shan J, Binnie M, Lewin M, Ghosh A, Miner JH, Vainio SJ & Quaggin SE (2015)  Loss of the podocyte-expressed transcription factor Tcf21/Pod1 results in podocyte differentiation defects and FSGS. J Am Soc Nephrol. 2014 Nov;25(11):2459-70. doi: 10.1681/ASN.2013121307. Epub 2014 Jun 5. PubMed PMID: 24904088; PubMed Central PMCID: PMC4214535.

Naillat F, Yan W, Karjalainen R, Liakhovitskaia A, Samoylenko A, Xu Q, Sun Z,  Shen B, Medvinsky A, Quaggin S & Vainio SJ (2015) Identification of the genes regulated  by Wnt-4, a critical signal for commitment of the ovary. Exp Cell Res. 2015 Mar 15;332(2):163-78. doi: 10.1016/j.yexcr.2015.01.010. Epub 2015 Jan 30. PubMed PMID: 25645944.

Naillat F, Veikkolainen V, Miinalainen I, Sipilä P, Poutanen M, Elenius K & Vainio SJ (2015) ErbB4, a receptor tyrosine kinase, coordinates organization of the seminiferous tubules in the developing testis. Mol Endocrinol. 2014 Sep;28(9):1534-46. doi: 10.1210/me.2013-1244. Epub 2014 Jul 24. PubMed PMID: 25058600.

Pietilä I, Prunskaite-Hyyryläinen R, Kaisto S, Tika E, van Eerde AM, Salo AM, Garma L, Miinalainen I, Feitz WF, Bongers EM, Juffer A, Knoers NV, Renkema KY, Myllyharju J & Vainio SJ. Wnt5a Deficiency Leads to Anomalies in Ureteric Tree Development, Tubular Epithelial Cell Organization and Basement Membrane Integrity Pointing to a Role in Kidney Collecting Duct Patterning. PLoS One. 2016 Jan 21;11(1):e0147171. doi: 10.1371/journal.pone.0147171. eCollection 2016. PubMed PMID: 26794322; PubMed Central PMCID: PMC4721645.

Pietilä I & Vainio SJ. Kidney development: an overview. Nephron Exp Nephrol. 2014;126(2):40. doi: 10.1159/000360659. Epub 2014 May 19. Review. PubMed PMID: 24854638.

Prunskaite-Hyyryläinen R, Skovorodkin I, Xu Q, Miinalainen I, Shan J & Vainio SJ. Wnt4 Coordinates Directional Cell Migration and Extension of the Müllerian Duct Essential for Ontogenesis of the Female Reproductive Tract. Hum Mol Genet. 2015 Dec 31. pii: ddv621. [Epub ahead of print] PubMed PMID: 26721931.

Rajaram RD, Buric D, Caikovski M, Ayyanan A, Rougemont J, Shan J, Vainio SJ, Yalcin-Ozuysal O & Brisken C. Progesterone and Wnt4 control mammary stem cells via myoepithelial crosstalk. EMBO J. 2015 Mar 4;34(5):641-52. doi: 10.15252/embj.201490434. Epub 2015 Jan 20. PubMed PMID: 25603931; PubMed Central PMCID: PMC4365033.

Siirtola P & Röning J: Reducing Uncertainty in User-independent Activity Recognition - a Sensor Fusion-based Approach. accepted to ICPRAM 2016 conference.

Tuovinen L, Ahola R, Kangas M, Korpelainen R, Siirtola P, Luoto T, Pyky R, Röning J, Jämsä T (2016) Software design principles for digital behavior change interventions: Lessons learned from the MOPO study. 9th International Conference on Health Informatics (HEALTHINF 2016), accepted.

Tuovinen L, Kahelin R & Röning J (2015) A conceptual framework for middle-up-down semantic annotation of online 3D scenes. In Proc. Ninth IEEE International Conference on Semantic Computing (ICSC 2015), 464–469.

Junttila S, Saarela U, Halt K, Manninen A, Pärssinen H, Lecca MR, Brändli AW,  Sims-Lucas S, Skovorodkin I, & Vainio SJ. Functional genetic targeting of embryonic kidney progenitor cells ex vivo. J Am Soc Nephrol. 2015 May;26(5):1126-37. doi: 10.1681/ASN.2013060584. Epub 2014 Sep 8. PubMed PMID: 25201883; PubMed Central PMCID: PMC4413750.

Last updated: 21.4.2017