Deliverables Ultra-low power sensor data transmitter This result derives from work performed on micro-power wireless communications in the ORESTEIA project. The overall aim of this work was to determine quantitatively the minimum power requirements for ultra-low power wireless links connecting low-level artefacts in the ORESTEIA domain. First, the electromagnetic problem of power transfer between loop antennas in a short-range wireless link was analysed in depth. Using this analysis we were able to determine the optimal transmission frequency for maximum signal power transfer as a function of range and antenna size. In a second phase of the work, a specific low-power transmitter topology - the oscillator transmitter - was identified and analysed, allowing power dissipation in the transmitter to be included in the frequency optimisation process. A key conclusion of this work was that short range (ca 1m) transmission at low data rates (ca 1kbps) should be feasible on an average power budget below 1. Phase three involved detailed design and prototyping of a 434MHz link employing a discrete oscillator transmitter and a receiver based around a commercial ASK (amplitude shift key) receiver chip. These prototypes were successfully tested, and proof-of-concept thus established. Micro-engineered vibration-powered generator This result represents the achievements of the ORESTEIA project on micro-power generators. These generators are designed to scavenge energy from the motion of the human body, for use in self-powered sensor nodes. These nodes are an important class of level 1 artefact for the ORESTEIA architecture. Consideration of the power requirements for the nodes shows that sub-microwatt levels are sufficient to power a range of sensor types, including transmission of the sensor data a short distance to a higher level artefact. Fundamental analytic studies were then carried out to show what the achievable power levels are for vibration based micro-generators, as a function of device size and the nature of the motion exploited. Project team has shown that previously reported micro-generators, which are mechanically resonant, are unsuited to body motion applications because of the low and inconsistent frequencies involved. The team has invented a new generator class based on nonlinear, non-resonant motion, the Coulomb-damped parametric generator. This device was analysed in detail, and computer models were constructed of the complete electro-mechanical system, by which the performance is assessed including the effects of parasitic components and other practical constraints. An electrostatic constant charge design was chosen. Prototype devices have been constructed based on parallel plates, which move normal to the plate. These devices are assembled in a hybrid manner from parts constructed using MEMS (Micro-electromechanical systems) technology. Novel instrumentation was also developed to assess the performance of the device. The prototypes were tested, and shown to operate as expected. Application of this generator strategy in real sensor nodes will require integration of the required power electronics for capturing and conditioning the energy. Because of the strong demands of the application in terms of device performance, construction of such circuits from commercially available discrete components is not feasible. Consequently new electronic devices were designed that are compatible with standard power electronic fabrication techniques, but which satisfy the unique requirements of this application. This work-package began with a fundamental study of how energy harvesting could be achieved from human body motion. It has advanced to testing of prototype generators that demonstrate the feasibility of the proposed novel design, and determination of the steps necessary to process the power into a usable form. Attentional agents The Attentional Agents architecture is a processing and control architecture for multiple goals inside an agent. The processing is organised in four consecutive layers. In the first layer belong inputs (and outputs) from sensors (or actuators). In the second one pre-processing operations take place. In the third and in the fourth levels local and global decisions are made. The input space is divided in a number of modalities. The level 3 goals encapsulate together perception, state evaluation, local attention control and action generation for a given modality. Fusion of local goals is achieved in level 4 goals (multi-modal decision fusers). They organise in suitable family trees other goals. Families compete for producing the required overall action, using the idea of attention indices. These indices are created as a result of special events. Attention control is used both on the local level, described previously, and in the global level of goal family competition. The architecture includes learning modes, which are smoothly integrated with processing modes. Thus the agent can be in processing mode for some goals while in learning mode for others. The framework integrates various processing styles in a coherent whole, while provides default classification and learning methods. Multi-agent extensions are supported with suitable extension of agent types to levels 5 and 6. User and internal agent goals are supported and processed in a uniform way. Self-survival is also aided by using the same mechanisms. The architecture has already appeared in the form of scientific papers in suitable conferences and journals. Prototype software systems were demonstrated in suitable events of the Disappearing Computer community. The potential for further use is very high, especially in further research and development efforts. Currently no immediate commercial exploitation is possible but development of domain specific applications is possible within one year of further developments. The Attentional Agents framework can be immediately used as an enhanced processing and control scheme and as a building block in future complex Cognitive Agents models. Many concrete systems can be based on it in a number of areas. An attention-based processing and control architecture for ambient intelligence The ORESTEIA project has developed an innovative attention-based guidance system for humans to enable a more efficient and less hazardous living. The system uses inputs from a number of sensors located suitably in the user's environment and on him/her (in the case of bio-sensors). The collected signals are processed with the goal of evaluating the state of the user, of his/her environment and of other agents in order to reach decisions as to the best action necessary to maintain the user's good state. The developed architecture is a synthesis of: 1. A general conceptual processing and control architecture (Attentional Agents). 2. Pre-processing algorithms for extraction of the most relevant features for bio-signal based state estimation and classification. 3. Special methods for state classification, based either on fuzzy rule-based systems or extracting the necessary rules from the data. 4. Enabling technologies for the low level artefacts present in (1) in the form of: - Ultra low power antenna design for wireless communications - Power generators and accompanying electronics structures for powering the low-level artefacts for processing and communication. Collection of a large body of data in the quasi-realistic setting of a car-driving simulator was performed for the development of the overall system. The scenarios used aimed in collecting data for users' (drivers) reaction under stressed conditions. The results produced in the project have already appeared in the form of scientific papers in suitable conferences, journals and books. Prototype systems were demonstrated in suitable events of the Disappearing Computer community. The potential for further use is very high, especially in further research and development efforts. The current state of affairs does not allow immediate commercial exploitation but component technologies can reach this stage in 2-4 years of further developments. The project is are primarily interested in applications in the following areas: - Ambient Intelligence - Health Monitoring - Car Hazard Avoidance - Robotics The current results will be further developed for application to the aforementioned domains. Of particular interest is the area of safe driving and health-monitoring of aged, very young or chronically ill people. The power generator and accompanying ultra low-power wireless communication technologies are expected to find very wide use in future "autonomous" artefacts. Thus they enable the vision of Ambient Intelligence of a more natural Human Computer Interaction paradigm. The Attentional Agents framework can be immediately used as an enhanced control scheme and as a building block in future complex Cognitive Agents models. Finally the classification methods provide a way to implement state classifiers for open domains. Advanced methods for physiological data analysis The physiological signals that are used in the framework of ORESTEIA are, with the exception of ECG, simple signals providing information about the vital signs of humans. Thus, for blood pressure, heart rate, breath rate and body temperature the features that medical experts use in their everyday practice are more or less the instantaneous values and their variations inside a time window. The situation for ECG is rather different. ECG is a complex signal and there are several time-templates for which medical experts are searching for in order to provide their diagnosis. It should be noted, however, that even the experts face difficulties in interpreting ECGs in several cases. This is mainly due to the fact that the time-templates/characteristics are strongly influenced by noise and rarely have the ideal form that medical handbooks describe. Apart from the analysis of the aforementioned simple signals, the basic aim is to identify predictive variables that will be used to detect abnormalities in the form, duration, and amplitude levels of ECGs sub-components. The project has implemented a fast and robust method for QRS detection on ECG signals that works reliably on real data. It should be noted that this method works quite robustly for noisy data. Allan factor was also chosen as a frequently used feature for possible irregularities in the heartbeat sequence. The wavelet transform standard deviation was also used as a way to detect irregularities at the ECG waveform. Emphasis was also given to the ECG spectrum. The idea here was to find correlations between various pulses of ECG and corresponding bins in ECG power spectrum. The algorithms that were developed have been presented in events of the Disappearing Computer Community and are extensively described in the project deliverables. They have been included in the prototype demonstrations as parts of the whole system. The potential for use in the healthcare sector for fast, real-time surveillance and diagnosis is very high but still further development is needed. A hybrid intelligence architecture for health status monitoring (CAM-SPM system) The CAM-SPM system (Connectionist Association Module Symbolic Processing Part) is a hybrid intelligence architecture that has been designed mainly for modelling physical phenomena for which a priori knowledge is available, either in the form of linguistic rules or in the form of known symbol values. In the framework of ORESTEIA project this architecture has been considered as the most appropriate to implement the state mapping in the biometric artefact. The generic CAM-SPM architecture consists of a connectionist (subsymbolic) association part, with a Numerical Data component, and a symbolic processing part, with a Semantic Knowledge component. In this modular architecture the Connectionist Association Module (CAM) provides the ability of grounding the symbolic predicates (associating them with the input features), while the Symbolic Processing Module (SPM) implements a semantically rich reasoning process. The Semantic Knowledge module corresponds: (a) to linguistic rules that describe a particular physical phenomenon, and/or (b) to rule databases that may adapted by researchers as more rules describing the new particular physical phenomenon made available. The semantic knowledge component may not exist at all. The Numerical Data module provides the means for (a) the initial training of CAM module, (b) the adaptation of the CAM module to specific contexts. In the case of the biometric artefact features derived from the heart rate, respiration rate, ECG signal, systolic and diastolic blood pressure are fed to the CAM. The symbolic predicates (intermediate representation) correspond to semantic evaluations of users heart rate, blood pressure as well as of its ECG-pulses form that should be able to provide some indication about the health status of its user, which may fall into four categories: normal, slightly abnormal, abnormal, dangerous. The rules connecting the symbolic predicates with the World Representation are inserted in the system through the SPM module. Aspects of the CAM-SPM architecture have already been presented in several international scientific papers. A prototype has been demonstrated in suitable events of the Disappearing Computer community. The potential for further use is very high, especially in further research and development efforts. The current state of affairs does not allow immediate commercial exploitation. However component technologies of this architecture can reach this stage in 2-4 years of further developments. We are primarily interested for applications in the areas of Ambient Intelligence, Health Monitoring, and Car Hazard Avoidance. The current results are expected to be further developed for other applications. Of particular interest is the area of safe driving and health-monitoring of aged, very young or chronically ill people. Prototypal system for driver attention detection during a drive This part of the project was about monitoring the awareness state of a driver focused on demanding manoeuvres. A Boolean normal form launches a flag when the driver is paying special attention to his driving. The contrasting analysis of these flags with the physical parameter of the car may alert a decision system whenever the driver's awareness is judged unsatisfactory. The task is carried out through a system capable of learning the flag computing procedure directly from a set of examples. Henceforth we relate to figures of the MS PowerPoint file reposited on the web site http://laren.dsi.unimi.it/PAC-DEMO-2.ppt. In particular a synopsis of the whole procedure is in page 1. Input signals We work on a quadruplet of tracks of physiological data like in Figure 1(a): electrocardiographic (ECG), respiration (RSP), galvanic skin response (GSR) and skin temperature signals (SKT) of the subject at the sampling rate of 200 Hz (pg. 2). The data are collected through a Biopac device during driving sessions held at the driving simulator of the Department of psychology of Queen University of Belfast. Besides we store the main cinematic car parameterssuch as speed, steering angle, brake strength, etc. The device architecture It is a hybrid multiplayer perceptron where first layers compute mainly subsymbolic functions like conventional neural networks producing propositional variables from the signal features, while the latter are arrays of Boolean gates grouping these symbols in formulas of increasing complexity. (pg.3). The learning procedure We specialize the two modules of our architecture to the following tasks: From features to symbols Starting from a set of features extracted by the above physiological tracks, we give the neural network the task of producing symbols from them with the key commitment of getting a syntactic mirroring of these data. We look for a vector of Boolean variables whose assignments reflect the relevant features of the original (possibly continuous) data, where assignments may coincide when they code data patterns with the same values of the features of interest to us. This amounts to computing Independent Components of the signals with the specialfeatures of being Boolean (BICA) (pg.4). This task is fulfilled by training the neural network through a special back-propagation procedure having the goal of minimizing peculiar entropic costs. Building rules Given a set of examples in the Boolean hypercube we compute a pair of rough sets respectively included and including the goal formula. Namely, we compute a minimal DNF (union of monotone monomials m) D including all positive examples (instants requiring attention) such that no other DNF with same property exists included in D. Analogously, we compute a maximal CNF (intersection of monotone clauses) C excluding all negative (not requiring attention) points such that no other CNF with same property exists including C (Pg. 5). The computation is very fast. Then we simplify these formulas in the aim of gaining them a better understandability at the expense of broadening their contour. Namely, we endow the formulas with membership functions spanning the gap between them. An optimal simplification (balancing length and fuzzyness of the formulas) is obtained via a simulated annealing procedure. The rendering procedure We are just interested in discriminating between attentive and non-attentive state of the driver in connection with driving episodes requiring or not attention. We define these episodes with a series of tentative rules such as: Overtaking -> trajectory far from the correct lateral line position, high wheel angle and high acceleration; Both car and driver states evolve along substantially continuous processes. This means that we are looking for blocks of attention requiring and attentive response pulses during the driving story. The standard rendering of the procedure consists of a pair of graphs: one reports the physical curve parameters (STISIM parameters) recorded by the simulator, denoting car acceleration, braking strength, wheel angle, distance from the center line). The other shows two Boolean variables: red bars are up when an attention requiring episode occurs according to our above empirical rules; green bars are up when the system detects a driver attentional state. A set of frames of a driving simulation session and related graphics are reported in Pg. 7 to 17 of the above presentation. Searching for OpenAIRE data... There was an error trying to search data from OpenAIRE No results available
