The session was about open architecture using IoT Edge, Azure Cognitive Services, Mosquitto MQTT, Influx DB and GraphQL web services to develop advanced architecture for early detection of forest fires that integrates sensor networks and mobile (drone) technologies for data collection and processing. Unmanned air vehicles (UAVs) will allow coverage of larger areas to raise the percentage of forest fires detections, monitor areas with high fire weather index and such already affected by forest fires. All information is forwarded and stored in cloud computing platform where near real-time processing and alerting is performed.
4. Objectives
Early detection of forest fires
Open and interoperable
Wide range of interfaces, protocols and devices
Existing Crisis Management Systems (National, EU-level)
Continuous monitoring of disaster related data
Retrospective disaster assessment
New methods for fire detection (AI, drones, sensors)
Command devices in surrounding area (i.e. barriers)
Automatic processing and alert generation
Decision making support
Cost efficient monitoring
4
5. Platform Benefits
Open source and free license components
Deployment on-premises and public cloud
Adaptable – multiple abstraction points
Cutting edge technologies
AI, Machine Learning, Time Series data, Drones support
High performance
30’000+ simultaneous agents/things on node, 1M sensor parameters
Built with security in mind
TRL 6 (test in relevant environment)
With sensors, thermal cameras and state of the art AI algorithms,
the platform aims to combine the best approaches
and achieve 10% better fire assessment and prevention.
5
6. 6
ASPires
Distributed & Open
IoT Platform
Funded by
EUROPEAN CIVIL PROTECTION AND
HUMANITARIAN AID OPERATIONS
ECHO/SUB/742906/PREV03 (ASPIRES)
7. Ingestion Scope
Field gateway (On-premises)
Local aggregation point for range of sensors (on-site, drone mounted)
Features: Device specific protocol conversion (i.e. Lora-Modbus TCP)
Cloud gateway (On-premises)
Cross-platform SW gateway based on Azure IoT Edge with 2-way communication
Features: Buffering, Processing, Protocol adapters (i.e. Profinet, Profibus, OPC, ZigBee…)
Ingestion hub
Entry point of the cloud platform (MQTT broker)
Features: Secure 2-way communication over MQTT and HTTP
7
8. Cloud Scope
Stream Processor
Real-time data simple analytics (i.e. threshold, anomaly detection, data conversion)
Features: Device specific protocol conversion (i.e. Lora-Modbus TCP)
Data Persisters
Data storage engine abstraction (RDBMS, Time Series)
Features: Read/Write operations on persistence storage
Data Storage
Asset Configuration, Raw Data, Time Series Data
Features: Performance (40x-60x faster than SQL and NoSQL) and TS data functions
Analyzer services
Generic and specific pluggable custom logic components
Features: Asynchronous data processing 8
9. Consumer Scope
Platfrom Micro-services
Open format scalable APIs (GraphQL)
Features: Identity Management, Data Retrieval & Manipulation, Alerts
Configuration Portal
Web application with limited access (i.e. organization administrators)
Features: Platform configuration and asset management (i.e. GW, sensors, users)
Presentation and Visualization
Mobile apps, Web apps, Hybrid apps
Features: Responsive, administration, dashboarding
Third Party Platform Integration
Crisis Management Systems, Recourse Management
Features: Raw data, trends, analytics and alerts visualization
9
11. Drone Support
Complement other infrastructure (i.e. towers)
Rapid deployment almost anywhere
Sensor data collection (fixed, mobile)
Visual image and thermography
Near real-time transmission
Low cost
Protocols (ZigBee, WiFi, 3G/4G)
11
12. Computer Intelligence
Seas of data are generated by sensors
How to gain insight on hidden relations?
How to get actionable results?
How to make platform provide business value?
Open issues
Cameras are diagnostic and not predictive approach
Thermal cameras cannot detect fire behind hills
FWI is not able to model hidden relations and thresholds
Human operators are required
12
13. Computer Vision (Positive Sample)
13
Computer intelligence and computer vision could be used for
automated alerting for camera capture
Works on predefined description tags
Less operators could control larger area
14. Computer Vision (Negative Sample)
14
Computer vision can process low resolution images (VGA)
Computer vision is capable to distinguish clouds from smoke
Alerts could be raised based on confidence level
15. Forest Fire Prediction ML Model
15
ML model could be used for prediction of
fire state with certain confidence level
Model features are result of in-depth
analysis of other crisis management systems
Model consumed as web service from Azure
(ML Studio & Azure Experimentation Service)
16. Platform Openness
Open source technologies
Azure IoT Edge
Influx DB
Mosquitto MQTT
Open protocols
Inbound & Outbound Interfaces
Data & Alerts Services
CMS Systems: EFFIS, MKFFIS
16
18. Validation Setup
InfluxDB v1.3.0 single node SQL Server 2016 Standard
CPU: Intel Core i7 3.40GHz (quad, 8MB cache)
Memory: 16GB DDR3 1333MHz
OS: Windows 10 Enterprise, 64 bit
SSD: Samsung 850 EVO 250GB
HDD: Seagate 7200RPM, 16MB cache, 500GB
Name Type Indexed
ID (PK) int Yes
TagID int Yes
Value int No
Timestamp DateTime Yes
Name Type
tagID Int Tag Key
value Int Field Key
time DateTime TimeStamp
InfluxDB schema: SQL Server schema:
19. Write Performance (writes/sec)
0
50000
100000
150000
200000
250000
0 20,000,000 40,000,000 60,000,000 80,000,000 100,000,000
Influx
Insertspersecond
• Influx average write speed of 200,000 writes/second
• SQL Server average write speed of 5,000-8,000 writes/second
• Memory optimized tables increases write performance by 1.4x
• Write trend unaffected by the amount of data (in this volumes)
• Influx outperforms SQL Server 40x
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
0 20,000,000 40,000,000 60,000,000 80,000,000 100,000,000
SQL In-Memory OLTP
20. Disk Storage
• 27% improvement is SQL Server page compression is enabled (1883MB)
• Influx has 19x-26x less disk requirements for the same functionality
2579MB
98MB
0
500
1000
1500
2000
2500
3000
On-Disk Storage
SQL Server
Influx
DiskSpaceUsed(megabytes)DiskSpaceUsed(megabytes)
21. Read Performance (queries/sec)
0
500
1000
1500
2000
2500
3000
3500
4000
1 2 4 8
InfluxDB
SQL Server
QueriesperSecond
Concurrency
Benchmark query:
Aggregate samples of 10,000 points, average over 1,000 runs.
• SQL Server mean query response time – 78ms (13 queries/sec)
• Influx mean query response time – 1,3ms (769 queries/sec)
• Influx has 59x better query throughput than SQL Server
22. Read Performance (rows at a time)
• SQL Server relatively unaffected by SSD (due to caching)
• Influx performance improves up to 100% on SSD, chunk size 10’000
• SQL Server is up to 2x better on HDD (Influx is better up to 600K)
• Relatively equal on SSD (Influx is better up to 1.6M)
0
500,000
1,000,000
1,500,000
2,000,000
2,500,000
3,000,000
3,500,000
0 5 10 15 20 25 30
HDD Performance
SQL Influx
Queriedresults
Seconds to execute (lower is better)
0
500,000
1,000,000
1,500,000
2,000,000
2,500,000
3,000,000
3,500,000
0 5 10 15 20 25 30
SSD Performance
SQL Influx
Seconds to execute (lower is better)