gefördert vom AVS-Extrem Vernetzte Sensorsysteme zur Lokalisierung und Überwachung unter Extrembedingungen Norman Dziengel 1/25
Area Surveillance (e.g. Construction Site) Distributed Event Detection Reliable event detection within the network Hardware development and fence integration Event/Status-Packet (FU) Request-Packet (FU) Ranging/Localization (Nanotron) GPRS (Ezent) Data Aggr./Manag. (Ezent) Communication Localization TOA-based Chirp Ranging 2.4 GHz Localization Data Management/ Visualization Visualization and aggregation of Event-Data Time Schedule-Information Weather-Data Network Management Distributed Event Detection Event Tools Base Station Data Management/ Visualization Web- Server Localization Control Station 2/25
Localization-Demo LNDW 2011 Remote controlled wheeled loader Four anchor nodes Localization of wheeled loader during movement 4 3 2 1 3/25
SDS-TW Ranging Symmetrical Double-Sided Two-Way Ranging Node A Node B Data TOA TOA Propagation TOA Propagation Processing Ack Initiating device start round trip counter Responding station replies after constant time delay Propagation Initiating device stops round trip counter Processing Propagation Roles of Initiator and Responder is exchanged Calc mean of both round trip values for clock offset compensation 4/25
Steps to improve localization No Filter, No Border Penalty Border Penalty Walking from corner to corner Heavy Multipath Environment Ranging Filter Ranging Filter + Location Filter Location Filter Ranging Filter + Location Filter, Border penalty Multiple Filter Options Plausibility constraint Ranging filter Location Filter Combined filter: Ranging/location Border penalty 5/25
Control Station / Visualization Alarm types are visualized different colors depending on the alarm relevance In case of intrusion => short term notification additional mail/sms-notification Possible sensor node requests: position and status Configurable Keep-A-Live messages Visualization of multiple base nodes fence sensor nodes device sensor nodes 6/25
Database Main Functionality Maintenance of User- and Network data Logging and Notification Event evaluation (weather data) Maintenance of GI-data and objects (vehicle, fence sensor, sectors, base nodes) Normalized scheme for maximized extendibility and flexibility SQL Server 2008 Low latency for server based alarm 7/25
Event Aggregation/ Webservices/ Interface Communication between WSN and base nodes: ASP.net based SOAP-Webservice Scheme uses XSD with scheme validation Available WSN-Data: Status information temperature, humidity battery-status distances, position Events (shake, climb, lift, etc...) 8/25
WSN System Architecture Sensor node development Integration of hardware + energy supply in fence System Application Extreme environment (weather und massive shocks) Thread based OS Dynamic routing Multiple base stations Encrypted communication Distributed event detection Data quality estimation Self calibrating acceleration sensor WSN System Architecture HW: AVS Sensor Board OS: FireKernel Distributed Event Detection Data Quality Estimator ACC-Sensor Micro-Mesh Routing Preprocessing Energy Management MCU: ARM7 Multiple Base Nodes Feature Extraction ACC- Calibration Housing Security Prototype Classifier Energy Management (ACC-Logic, WOR, PD, In-Network Evaluation) 9/25
AVS-Extrem Platform Hardware ARM7-MCU based architecture 3D-Acceleration sensor (±2g, ±4g, ±8g), 10bit resolution 868 MHz CC1101 transceiver Coulomb Counter to monitor energy consumption External modules Software Priority driven OS Kernel supports power saving ACC-Data is buffered to enable time-delayed access e.g. for algorithms for analysis or classification allows higher priorities for communication and routing Integrated Distributed Event Detection Pattern recognition and in-network evaluation Event distribution dependent of relevance ISM antenna TI CC1101 Bosch SMB380 3D-ACC sensor NXP LPC2387 ARM7 MCU External pins for research JTAG connector Linear Technology LTC4150 Coulomb Counter FTDI FT232R USB UART IC USB connector AVS-Extrem Sensor Node 3 cm 10 cm Socket for optional FSA03 GPS module Connector for ext. modules (e.g. nanopan) Reset button Sensirion SHT11 RH/T sensor LEDs SD card socket Power supply 10/25
Packaging Diameter in the fence: Ø 38 mm Energy supply within the fence: 4 D Cells (Ø 33 mm, 16Ah) Board size: 30*100 mm Setup: simple plugging system Housing: Ø 36 mm (wall thickness 1-2 mm) Weatherproof material: Makrolon and Brass Board-Layout AVS-Extrem PCB Additional modules Skeleton Housing 11/25
Housing Details Installation in the fence is simple Alternative energy technologies possible Mechanical reverse polarity protection Capacitor Extendable through modules Reverse Polarity Protection 4* 1,5V D-Cells (16Ah) Battery Case (Brass) Makrolon -Tube Skeleton AVS Sensor Board Fence Installation Lock Complete Module 12/25
Former Solution: AVS-Extrem Distributed Event Detection scenario analysis In >98% of the time the system is idle (no pending events) active components High-cost software based acceleration data analysis Few pending events imply low radio traffic Former implementation Poll Sensor Data Transceiver uses CRX (Constant Receive Mode) Approximation of current time No multithreading OS Solution reduces lifetime and flexibility! LPC2387 block diagram 13/25
Energy Concerning Advantages Priority based Kernel supports power saving ACC-Logic/RTC/Radio wakes up MCU only if necessary IDLE: active components PD: active components PD: cyclic active components (integrated motion detection IDLE-mode (int. Interrupts) internal clock of ARM7 disabled (MCU stop) Wake-Up via IR Power Down (ext. Interrupts) Periphery=OFF MCU=OFF Wake-UP via ext IR (e.g. ACC, RTC, CC1101) Wake On Radio (WOR) for transceiver Duty Cycle of 542 ms LPC2387 block diagram 14/25
Failure Safety: Redundant Base Stations Web- Server 15/25
Distributed Event Detection In-network data processing is a key feature of Wireless Sensor Networks (WSNs) Time: In-network evaluation of condensed data, transmitting results ONLY if needed => In-network decisions are possible Energy: Reduce communication with base station => Extend network lifetime Centralized approach Distributed approach Passive, Active Node Radio Link Data Traffic Event Base Station 16/25
Principle of Distributed Event Detection 1. Preprocessing: Sample raw data Filter and smoothen data 2. Feature Extraction: Assessment & Selection of appropriate features is part of training (cross-validation) Extract application-specific set of features from raw data 3. Feature Distribution: Send features to local neighborhood 4. Classification: Fuse received and own features to distributed feature vector Classify the feature vector (prototype classifier) Report to base station, if event is configured as relevant node n+1 node n Event node n-1 0. Event 1. Preprocessing 2. Feature Extraction 3. Feature Distribution 4. Classification 5. Report 5. Report: Possible alarm or event feedback 17/25
Experiments / Location 18/25
Results Metrics 100 80 93 93 80 82 87 60 % 40 Shake Kick 20 0 Shake Kick Lean Climb Average Sensitivity Specificity PPV NPV Accuracy Lean Climb Sensitivity (recall) = TP / (TP+FN) Proportion of correctly detected events Specificity = TN / (TN+FP) Proportion of correctly ignored events Positive Predictive Value PPV = TP / (TP+FP) Probability that correctly detecting an event reflects the fact that the system was exposed to a matching event Negative Predictive Value NPV = TN / (TN+FN) Probability that correctly ignoring an event reflects the fact that the system was not exposed to a matching event Accuracy = (TP+TN) / (TP+TN+FP+FN) Proportion of true results in the population: the sum of all correctly detected and all correctly ignored events. 19/25
Results Feature Fusion 100 80 93 93 80 82 87 60 % 40 Shake Kick 20 0 Shake Kick Lean Climb Average Sensitivity Specificity PPV NPV Accuracy Lean Climb Detection of Shake and Kick Events All Metrics above 80%, Accuracy above 90% Detection of Lean and Climb Events: lower accuracy Sensitivity is comparatively low while specificity stays high Too many events have been rejected because of small prototype region Run of training was too similarr which leads to a too small prototype region Accuracy is about 87.1% after feature fusion 20/25
Energy-Awareness 500 400 Energy-Awareness Techniques (EAT) No Event Detection Distributed Event 456 Detection Application (DED) 419 360 Centralized Event Detection Power input in mw Lifetime in days 300 200 100 0 288 206 162 96 39 8 10 23 9 18 13 - - IDLE IDLE + PD - ACC-Logic No EAT WOR WOR WOR WOR + IDLE + PD No EAT: WOR: IDLE: PD: DED: ACC-Logic: no Energy-Awareness Techniques Wake On Radio internal clock of ARM7 disabled(mcu stop) Periphery on, Wake-Up via IR Powerdown (Periphery=OFF, MCU=OFF, external IR-Wake-UP e.g. ACC, RTC, CC) AVS-Extrem application using Distributed Event Detection MCU for sensing not necessary 21/25
Energy-Awareness & Latency 500 Energy-Awareness Techniques (EAT) No Event Detection Distributed Event 456 Detection Application (DED) 419 Centralized Event Detection 2.800 2.400 400 360 2.000 300 288 1.600 Power input in mw Lifetime in days 200 100 0 206 162 96 8 10 23 9 18 13 39 - - IDLE IDLE + PD - ACC-Logic No EAT WOR WOR WOR WOR + IDLE + PD latency in ms 1.200 800 400 0 1 2 3 1 2 3 1 2 3 Indoor Outdoor Indoor WOR Hops CRX Hops Worst Case: one week lifetime => Short term (full logging) apps are possible Lifetime increases by 3 times for DED compared to Centralized Event Detection Internal ACC/Logic increases lifetime about 16 times Latency increases, because WOR uses duty cycle of 542ms 22/25
Future: Bridge-Event Detection 1. Deployment Technical Initialization 868 MHz Collection + Exchange of data GSM-Modul Upload of Reference Data 2. Surveillance with Feedback 868 MHz Exchange of data Classification Event-Feedback via GSM-Modul 23/25
Future: BAN Rehabilitation/ Sport Rehabilitation 1. Supervised Training Extract Features 2. Setup Calculation & Upload of Prototype Vectors 3. Event Detection with Feedback Exchange & Fuse Features Classification of Events Supervisor Training Wireless Sensor Network Control Station Training Wireless Sensor Network Feedback Sport: Stick-Fight Device 24/25
Thank You Questions/Comments gefördert vom 25/25