IOT Analytics

Industry 4.0 – Quo vadis?

Industry 4.0 is the newest phase of Industrial Revolution being brought through faster wireless communications, real-time data streams, machine learning and AI driven automated decision systems and inter-connected devices and machines. It is also referred to as Industrial IoT or smart manufacturing. Industry 4.0 is revolutionizing decision making through real-time insights across production, demand forecasting, supply chain management and inventory management.

Evolution of Industry from 1.0 to 4.0

Manufacturing has evolved over the three centuries, and there have been four distinct phases of industrial transformation:

1. Industry 1.0: The invention of steam engines, and evolution of manufacturing from manual labour to optimized machines driven by steam powered engines. Happened in late 1700s and early 1800s

2. Industry 2.0: Brought about by the introduction of steel and use of electricity in manufacturing. Mass production concepts like assembly line formulated. Happened in early 1900s

3. Industry 3.0: Began with the introduction of computers in the manufacturing process. Focus shifted from the use of analog to digital technologies. Started in late 1950s

4. Industry 4.0: Started with the use of inter-connected machines (called Internet of Things or IoT). Has allowed connecting physical and digital machines. Started in 2010s

What is changing OR going to change in Industry 4.0?

Industry 4.0 is changing the manufacturing processes. We present a few use cases here:

1. Supply Chain Optimization: Manufacturers have the ability to integrate their manufacturing process with the supply chain platform. This provides them real-time insights, market trends and demands to deliver products at a faster and cheaper rate

2. Asset Monitoring: With enhanced tracking and monitoring capabilities using wireless connectivity, video based intelligent feeds and location based real-time data, manufacturers can identify supply chain issues and asset quality risks to optimize asset transfer, disposal and adjustments in real-time

3. Predictive Maintenance: Every manufacturer relies on machines, and the machines require regular maintenance to work efficiently. The different maintenance strategies are:

a. Reactive Maintenance: Repairs and maintenance done on machine failure

b. Preventive Maintenance: Failures are costly, and maintenance done at regular intervals to prevent failures

c. Predictive Maintenance: Condition based monitoring of operational parameters to predict failures, and carrying out the maintenance just before a failure is predicted to happen

Benefits of Industry 4.0

We list some (not all) benefits of adopting Industry 4.0 capabilities to your business:

1. Gives you the competitive edge

2. Increases collaboration and promotes shared decision making

3. Predictive analytics using real-time data reduces inefficient preventive maintenance costs and downtime

There are many other benefits, most of which do not fall under our skillset.

To outperform in today’s business environment, manufacturers need to embrace Industry 4.0 or risk falling behind.

IOT Analytics

Optimize the energy consumption for HVAC unit for a commercial building

The commercial real estate client owns multiple assets in Japan. They had invested in IoT based monitoring of HVAC system in their building. And were now looking to utilize advanced analytics based approach to map user preferences and optimize energy consumption in the building. The client was looking for an advanced analytics based decision system to overlay on top of temperature, humidity and external weather data to optimize the energy consumption.

Smart system for cost efficiency

Superimposing temperature and weather data from local weather station, we are able to identify ideal settings and optimized setting for the HVAC system.

Recommended temp. setting would reduce the power consumption by 1.5%

1 degree incremental temp. differential increases power consumption by 1.2%

Single AC On/Off cycle results in ~1.6 units of additional power consumption

The model had to be trained on large volume of consumption data and the results were generated using 6-month IoT data through a retrospective study.

Multiple dependencies. Complex modeling.

The real estate client has deployed sensors and IoT based monitoring of temperature and humidity for one of their commercial assets. They wanted to understand how IoT data and a machine learning based decision system could help them optimize the power consumption and reduce manual operations.

One of the key gaps we identified based on the data being Collected was non-availability of occupancy data. In the absence of this data, power wastage could not be computed. Another issue with the data being collected through IoT system was lack of integration with an outside weather data source. The third issue identified was with the data quality itself – having AC temperature observations that were practically not possible. We proceeded with a stochastic non-linear model to identify efficiency issues.

State of the art ML approach optimizes energy consumption

The study was taken up as a pilot to assess the potential of machine learning driven decision system to drive energy efficiency in smart buildings.

We took up a retrospective study to analyse the energy wastage over a 6-month period and demonstrate the insights from an ML driven decision system. We were able to identify the data gaps, and also make recommendations, including:

a. The optimal temperature to be maintained in a room based on outside temperature

b. Ideal temp. readjustment frequency – as it lead to additional power consumption

The real power of Machine Learning based HVAC load management system can be realised if it is integrated with the IoT platform

IOT Analytics

Estimate the economic changes using night-time illumination data

The PE firm makes investment decisions based on growth prospects of an economy. For this, they use economic performance indicators published by the countries. However, these reports are subject to manipulation e.g. China and data quality issues in developing countries where agencies do not have a lot of pressure on collecting accurate data. They wanted to look at a measurement methodology that is available faster than official releases and cannot be manipulated easily.

Identify activity clusters

Data gaps in developing countries are constraining and impacts the accuracy of the economic performance forecasts.

Economic activity estimates predicted within a week’s time

Accounted for the impact of cloud cover and pollution levels in estimates

Accounted for moon phase in illumination observations

The PE firm used the estimates from the model as an important metric to make investment decisions and have a more thorough analyses of investment decisions.

Data and measurement gaps

The economic activity data in developing economies suffer from two challenges – the first being availability of reliable data. This has more to do with enough investments in improving the data collection process and training the staff.

The second problem lies with the pace at which the data can be collected. Developing economies, with generally higher density of population (esp. China, India, Indonesia) make it difficult to collect data at a rapid pace. The staff take their own time collating data from far flung locations and coming up with the estimates, which need to be vetted by the respective government agencies. We proposed an inexpensive, open data source – NASA nigh-time illumination data that is available from their website on request and test if it can be used to estimate economic activity on ground.

A proxy for changes in the economy

In order to develop an effective economic activity estimation solution, the first thing we need is the ability to handle the large volume of data being generated by the satellites on a daily basis.

Once the above has been addressed, the external impact on the illumination readings should be taken care of through data cleaning process to:

a. eliminate the impact of cloud cover

b. eliminate the impact of cloud cover

c. eliminate the impact of pollution and smog

Time series decomposition of the albedo data yielded seasonality – for cloud cover, cyclicity – for moon phase, external air quality data for pollution effects. The final component is the proxy for economic activity.

IOT Analytics

Classification using drone imagery to identify crops under cultivation

Africa as a continent has been struggling with poverty. The political system is most of the countries are not stable. The high population growth rates are not helping in reducing the absolute number of people living below extreme poverty line. The rise in food imports in Africa has exacerbated the problem. Agriculture in Africa provides employment to two-thirds of the population and contributes between 30-60% of GDP for each country.

Locate, track & improve growth of crops

One of the avenues to boost agricultural output for small scale farmers is machine learning driven crop and growth monitoring for guided improvements in productivity.

Identified 10 crop types with average accuracy of 92%

Enabled an improvement of 6% in yield estimates

Opened doors for scaling the solution across a wider region

Crops under cultivation and crop-farm mapping improves food security through better yield estimates, crop rotation, and soil productivity.

Africa’s struggle with poverty

Agriculture in Africa is dominated by smallholdings, constraining use of technology and other interventions to boost productivity (and income) in a sustainable manner. The struggle is to minimize the gap between actual and potential yields, enabling smallholders grow sufficient crops to feed their families, and a surplus to sell. This would help them meet food security needs and generate income, helping them move out of poverty.

Drones capture data from farms at regular intervals, that can be analysed for identifying area under cultivation, predict expected yield and help farmers with optimum planting and harvesting strategies. The hope is that technology driven solutions can provide scalable solutions that would allow us to create a data-driven approach towards crop management.

Machine learning driven fight for food security

Harnessing machine learning to help the most vulnerable people is the best way to use it.

Machine Learning solution over farm imageries captured through drones is an important step towards a bigger data-driven approach towards food security and income growth in agriculture. The farm to crop mapping exercise opens doors to:

a. Faster and more accurate prediction of yields – at farm, village and municipal area level

b. Efficient monitoring of crop growth, and addressing the risks

c. Better soil productivity and crop rotation

Machine Learning solutions might change the way we farm and grow crops – reducing poverty in the process.

IOT Analytics

Predictive Maintenance in Industry 4.0

Predictive maintenance is a strategy that tries to predict when a machine is at high risk of failure and proactive maintenance activities can be performed just before it is predicted to happen. These predictions are based on machine learning algorithms utilizing data from condition based monitoring of machines and their critical components, made possible through inter-connected systems and sensors on these machines that form a part of Industry 4.0

Imagine a factory floor supervisor who is working on a normal day – having all machines running efficiently. The sensor data is regularly monitoring the running parameters & conditions of each critical component. If a machine part or the machine as a whole is about to develop a snag, one or more sensors monitoring the running conditions send out an alert – saving the manufacturer from unplanned downtime and saving time & lost revenue. This was possible as the manufacturer had deployed a machine learning based predictive maintenance solution that was able to predict a failure before it happened.

Maintenance Strategies for Asset Management

Maintenance has evolved from corrective interventions to condition based monitoring and prediction of failures before they happen:

Algorithms for Predictive Maintenance

The physical assets, including sensors, connected devices to collect real-time data etc. is one part of achieving predictive maintenance capabilities.

The second equally critical component is to train and develop the right machine learning algorithm to predict these failures. Our team at Xtage Labs is equipped to train and deploy machine learning algorithm for predictive maintenance. We follow the following steps for a machine learning based predictive maintenance algorithm development:

Remaining Useful Life (RUL) estimation models

We classify different remaining useful life (RUL) prediction approaches based on the following three business possibilities:

a. Similarity Model Algorithm: If we have the complete history of the running conditions of a machine, including normal working and failure data available

b. Survival Model Algorithm: In this case, we only have the failure information available and normal working conditions data is not captured

c. Degradation Model Algorithm: Here, we do not have failure data available, as the failure cases are very rare and critical and the machines are not allowed to fail (e.g. Aircraft Engine Failure). In this case, we collect data for the safety thresholds which should not be exceeded, and we use it to predict the RUL of a machine and when maintenance is required

Benefits of Predictive Maintenance are wide and far reaching. This includes

a)     Reduce Potential Downtime

b)     Improve inventory and maintenance schedules

c)      Estimate (and minimize) losses due to inefficient machine parts

d)     Improve component life

Real-time dashboard

We build and deploy predictive maintenance solution through interactive web based dashboards, having inbuilt alert system – to collect, process and predict RUL based on real-time data inputs from sensors and critical machine components. The real time dashboards is where the sensor data can be streamed and regularly monitored

Beyond Predictive Maintenance?

Prescriptive maintenance looks to build upon predictive maintenance. However, it can only be achieved once predictive maintenance solution has been developed, deployed and refined. Prescriptive maintenance would suggest possible resolutions to a predicted failure.

The usage of prescriptive maintenance is linked to the progress of machine learning & AI technologies, and hopefully we need not wait for long before prescriptive maintenance becomes the solution of choice in the maintenance industry

IOT Analytics

Predictive analytics to move from reactive to proactive maintenance

The utility provider has a network of about ~4200 distribution transformers located across the service areas. Maintaining these number of DTs is an operational challenge for them due to breakdowns, high scheduled maintenance cost and public inconvenience. They wanted to build a more robust, scientific and predictive solution – to forecast distribution transformer failure with high degree of confidence – therefore also improving customer satisfaction with their services.

Risk of failure

The utility provider was look at a machine learning driven solution that could predict the chances of failure with a high degree of reliability.

A failure prediction accuracy of 75% over validation dataset

Pre-emptive scheduling of maintenance, thereby reducing downtime

Better planning & scheduling of maintenance

A web based interface with automated reporting was provided to the utility services provider at every 24 hour intervals.

Random breakdowns. Difficult to predict.

The utility services provider had spread across teams. The operations team held some data, the maintenance team had few other data sets and so on. They did not have a centralized database to dig into. The distribution transformer (DT) to meter mapping had a lot of data quality and consistency issues. And since DTs have a life of more than 10 years on an average, the data was not available for the lifetime for most of the DTs.

In addition to the above, scheduled maintenance data was not properly maintained, with a high percentage of missing values. The biggest challenge however was the low percentage of failure cases – creating a highly imbalanced class with ~.025% failures. Because of this failure rate, standard Classification algorithms would not work, and we proposed a boosting approach to create more balanced classes.

Automated, accurate prediction of failure risk

The predictive maintenance (PdM) solution deployed for the client had an accuracy of 75% on the hold-out sample.

With the web based interface having integrated reporting, the client now gets reports at desired intervals. This enables the client to:

a. Identify DTs that are at higher risk of failure and proactively address them before a breakdown

b. Improved and planned maintenance scheduling, reducing inconvenience and revenue losses

c. Improved customer satisfaction levels with better, uninterrupted supply of power

Predictive Maintenance (PdM) solution enabled the client to serve customers with the desired efficiency.