Henry Ford once said “Any customer can have a car painted any colour that he wants so long as it is black!” which it’s now long gone. In the current context people seek for personalized treatment. Imagine calling customer service, every time you call you have to go through all the standard questions, and they don’t have a clue why you might be calling? Or whether you called before? and in the case of shopping, even if you are a regular customer and have a platinum or gold membership card you will not get any special treatment at the store, may be presenting the card at the cashier can get you a discount.
IoT is not a new thing, we have had lots of M2M systems that have monitored and controlled devices in the past, and when it comes to IoT we have more devices having sensors and a single device having more sensors. IoT it’s an ecosystem where IoT devices should be manufactured, app for those devices should be developed (e.g apps for phone), users
should be using those devices and finally they should be monitored and managed. WSO2’s IoT Platform plays a key role in managing and providing analytics for the IoT devices in the ecosystem.
should be using those devices and finally they should be monitored and managed. WSO2’s IoT Platform plays a key role in managing and providing analytics for the IoT devices in the ecosystem.
Data Types in IoT Analytics
Data from IoT devices are time bound because these devices do continuous monitoring and reporting. With this we can do time series processing such as energy consumption over time. OpenTSDB is a specialised DB implemented for time based processing.
Further since IoT devices are deployed in various geological locations and since some of those devices move, location is also becomes another important data type for IoT Devices. IoT devices are usually tacked with GPS and currently iBeacons are used when the devices are within a building. Location based data enables geospatial processing such as traffic planning and better route suggestion for vehicles. Geospatially optimised processing engines such as GeoTrellis are especially developed for these type of usecases.
IoT is Distributed
Since IoT is distributed by nature, components of the IoT network constantly get added and removed. Further since IoT devices get connected to the IoT network through all type of communication networks such as from weak 3G networks to Ad-hoc peer-to-peer networks, and they also use various communication protocols such as Message Queuing Telemetry Transport (MQTT), Common Open Source Publishing Platform (CoApp) and ZigBee or Bluetooth low energy (BLE). Due to these the data flow of the IoT network continuously get modified and repurposed. As data load varies dynamically in the IoT network, on-premise deployment will not be suitable, and hence we have to move towards public or hybrid cloud based deployment. IoT has an event driven architecture to accommodate its distributed nature where its sensors reports data as continuous event streams working in an asynchronous manner.
Analytics for IoT
IoT usually produces perishable data where their value drastically degrades over time. This imposes the importance of Realtime Analytics in IoT. With Realtime Analytics temporal patterns, logical patterns, KPIs and thresholds can be detected and immediately alerted to respective stakeholders, such as alarming when temperature sensor hits a limit and notifying via car dashboard if the tire pressure is low. Systems such as Apache Storm, Google Cloud DataFlow & WSO2 CEP are build for implementing such Realtime Analytics usecases.
Realtime alone is not enough! We should be able to understand how current situation deviates from the usual behaviour, to do so we have to process historical data. With Batch Analytics, periodic summarisation and analytics can be performed on historic data against which we can compare at realtime. Average temperature in a room last month, and total power usage of the factory last year are some example summarizations that can be done using systems like Apache Hadoop & Apache Spark on the data stored in scalable databases such as Apache Cassandra and Apache HBase.
Ok, with Batch Analytics we defined the thresholds and with Realtime Analytics we detected and alerted threshold violations. Notifying violations may results in preventing disasters but it does not help stopping similar issues arising again. To do so we need to investigate the historical data and identify the root course of the issue and eliminate that. This can be done through Interactive Analytics and with it’s Ad-Hoc Queries, it enables us to search the data set how system and all related entities have behaved before the alert was raised. Apache Drill, Apache Lucene and indexed storage systems such as Couchbase are some systems provides Interactive Analytics.
Than being reactive, staying a step ahead predicting issues & opportunities bring great value. This can be achieved through Predictive Analytics which helps in scenarios such as proactive maintenance, fraud detection and health warnings. Systems such as Apache Mahout, Apache Spark MLlib, Microsoft Azure Machine Learning, WSO2 ML & Skytree are systems that can help us build Predictive Analytics models.
An Integrated Solution for IoT Analytics
From the about technologies by selecting WSO2 Siddhi, Apache Storm, Apache Spark, Apache Lucene, Apache HBase, Apache Spark MLLib and with many other open source softwares WSO2 has built and integrated Data Analytics solutions that support Realtime, Batch, Interactive and Predictive analytics solution called WSO2 Data Analytics Server.
Issues in IoT Analytics
Extreme Load
With compared with the scale of the data produced by sensors, distributed centralised analytic platforms cannot scale and even if they can - it will not be cost effective. Hence we should look at whether we need to process and/or store all the data produced by the sensors? In most cases we only need the aggregations over time, trends that exceed thresholds, outliers, event matching a rare condition, and when the system is unstable or changing. For example from a temperature sensor we only need to send readings when there is a change in temperature and its no point periodically sending same value. This directs us to optimise sensors or data collection points to focus on doing local optimisations before publishing data. This helps in quick detection of issues as part of the data is already processed locally and instant notifications since decisions are also taken at edges. Taking decision at the edge can be implemented with the help of complex event processing libraries such as WSO2 Siddhi and Esper.
Uncertainty
Due to the distributed nature of IoT, data produced can be duplicated, arrives out of order, missing or even be wrong.
Redundant sensors & network latency can introduce duplicated events and out of order event arrival. This can impose difficulty doing temporal event processing, such as Time Windows & Pattern Matching. These are very useful for usecases such as Fraud detection, and Realtime Soccer Analytics (based on DEBS 2013 dataset) https://goo.gl/c2gPrQ where we build a system that monitors the soccer players and the ball and identified ball kicks, ball possession, shot on goal & offside. Algorithms based on K-Slack can help to order events before processing them on realtime.
Due to network outages data produced by the IoT sensors can go missing, and in these situations using complimenting sensor reading are very important where one of those sensor value will be some sort of an aggregation done at the edge which will help us to approximate the missing sensor values based on its aggregation. Such as publishing Load and Work reading when monitoring electricity where when some events are missed, from a later Work Event reading we will be able to approximate the Load reading that should have arrived during the outage. The other alternative is using fault-tolerant data streams such as Google Millwheel.
Further at times sensor reading won't be correct, this can be due to various reasons such as sensor quality and environment noise, in such situations we can use kalman filtering to smoothen consecutive sensor readings for better approximation. These type of issues are quite common when we use iBeacons for location sensing.
Visualisation of IoT data
Visualisation is one of the most important aspect of effective analytics and hence with Big Data and IoT visualisation becomes even complicated. Per-device & Summarization Views are essential and more than that users should be able to visualize as device groups based on various categories such as device type, location, device owner types, deployed zone and many more. Since these categories are dynamic and each person monitoring the system have various personal preferences, therefore composable & customisable dashboard is essential. Further charts and graphs should be able to visualise the huge stored data, where sampling & indexing techniques can be used for better responsiveness.
Communicating with devices
In IoT sending a command/alert to a devices is complicated, to do so we have to use client side pooling based techniques. Here we store the data the need to be pushed to the client in a database or queue and expose them via secured APIs (through systems like WSO2 API Manager).
Reference Architecture for IoT Analytics
Here data is collected through message broker such as MQTT, immediately written to the disk by WSO2 Data Analytics Server (DAS), at the meantime the collected data is cleaned in realtime, this cleaned data is also persisted, and parallely the cleaned data is fed into realtime event processing which in deed sends alerts and provides realtime visualisations. Stored clean data is used by WSO2 Machine Learner (ML) to build machine learning models and deploy them at WSO2 DAS for realtime predict ions. Further the stored clean data is also used by Spark to run Batch analytics producing sumarisation, which are then visualised in dashboards.
It’s a pleasure for me presenting “Sensing the world with Data of Things” at Structure Data 2016, San Francisco. Please find the slides below.
