Industrial IOT and Predictive Maintenance

Industrial IOT

IOT enables collection of huge amount of data with all sort of connected devices. This in combination with industrial manufacturing opens up a plethora of new possibilities. The industrial IoT adds new capabilities to operational technology including remote management and operational analytics. But the major value-add so far can be predictive maintenance.

Predictive Maintenance

Predictive maintenance involves combining machine learning and artificial intelligence (AI) with the deep pool of data generated by the flood of newly connected devices. This opens up the opportunity to deeply understand the way complex systems work and interact with each other.

Data Collection

The journey of preventive maintenance with industrial IOT begins with the collection of data coming from numerous sensors connected to the devices. Some of the common data sources are

  • Vibration: Monitoring the vibration of equipment, usually bearing vibration
  • Temperature: Monitoring the temperature variation.
  • Oil Levels: Measuring the variation in oil levels of equipment.
  • Acoustics: Using ultrasound to detect changes in sound made by the equipment.
  • Motor voltage and current: Monitoring for speed, torque, arcing.

The data can be analysed in runtime using either Spark streaming or storm based solution. Alternatively, the data can also be stored using Flume for analysis at later times.

Building Failure Models

This is the area where data analysis and data science comes in. The model building exercise can be divided into two main categories

  1. Conventional rule based – This form of analysis relies on human intelligence and empirical models built over time by understanding the system. One needs to specify a set of rules or failure conditions to the streaming data. If these are violated, the machine is likely to fail. This is fairly simple and has been in practice for a long time.
  2. Machine Learning based – This form of analysis comes in when there is no idea what’s happening with the system. The data has some patterns hidden in some files but requires sophisticated algorithms to decipher the patterns about which no prior knowledge exists within the realm of the system.

Machine Learning Algorithms

Machine learning algorithms are classified into two broad categories.

  1. Unsupervised learning – Algorithms that run on a data set with no human intervention. The result is a set of automatically identified patterns from your data that can be mapped to equipment failure. The key algorithms are – clustering and dimensionality reduction algorithms.
  2. Supervised analysis – These are algorithms that can be explicitly trained to detect the failure. This requires a subset of the data, which is already classified as a failure/not a failure. The algorithm learns from that and can then be run on the complete data in real time to pick out equipment failure. Algorithms include – Neural networks, deep neural networks, gradient boosting, support vector machines.

Typical Machine Learning Workflow

Offline analysis and model building

Online Application of the model


Data Storage (on cloud)


Online Model Update


For online model update, feedback data based on human judgement can work wonders in improving the performance of the model over time. This will result in an adaptive model that will mature over time.


Supervised vs Unsupervised Machine Learning Model

Supervised models typically outperform their unsupervised counterpart because of the availability of the ground truth. For image recognition tasks, proper labelling of the data samples is a necessity for supervised algorithms to run to their full potential. Further, these models could be improved if we channelize the feedback of the quality check data as it is the updated ground truth.


Comparison of In-house Analytics and Storage vs Cloud Solution

  1. Technical complexity – Open source programming frameworks such as Python, Scikit-Learn, Hadoop, Spark are available which require an amalgamation between Data Scientist, Data Architects and Data Engineers to implement the above analytics pipeline. These complexities are virtually handled on cloud platforms such as GoogleCloud, Amazon EC2 or Microsoft Azure.
  2. Cost Implication – The charges for off the shelf cloud solutions are typically made on pay per use bases. The cost incurred for in-house solution includes the hardware cost (servers) and additional engineers to develop and maintain the server. The net amount paid for inhouse solution can buy cloud solution for a period of 2 years.

Suggested Approach: Design and develop cloud solution and then replicate the setup in-house. This should be done if the projects are planned for a period of 1 year where the models are first deployed on the cloud and then replicated on in-house servers within the same period.


Step to solve an Analytical problem?

  1. Identify the key business problem to be solved. This will involve identifying the problem areas and analysing the cost benefit analysis of the solution provided on the table.
  2. Once the key problem area is defined, eg. Predictive maintenance using IOT, start identifying the data sources from which the data collection would begin. Some of the data sources might be getting collected; some other data sources would require new sensors to be installed.
  3. A data storage pipeline has to be established such that all the necessary data is available for further analysis and deriving insights.
  4. Special efforts and guidelines have to be placed to collect ground truth. Ground truth can involve feedback from the operator, quality check results and so on. Please remember, the ground truth is mandatory to install Supervised Machine Learning setup.



Predictive maintenance using IOT can predict machine-specific failures with high degree of accuracy by leveraging the power of data it is accumulating. It can diagnose problems and deploy maintenance technicians much more quickly – and only when actually needed, reducing costs associated with unplanned downtime. It can also predict the likelihood and the timing of an event type and uses that data to optimize the maintenance schedule and prevent unpredictable breakdowns. Overall, predictive maintenance can be much more valuable than preventative maintenance!