With 50 billion industrial IoT devices expected to be deployed by 2020, the volume of data generated is likely to reach 600 zettabytes per year. A single jet engine produces about a terabyte of data in five hours. Given these assumptions, we need a fast and efficient way to analyze data through data analytics. In the last five years, big data technologies have been improved to scale computational capabilities. Big data analytics is about collecting and analyzing large datasets in order to discover value and hidden data, and gain valuable information. The applications of these analytics are as follows:
- Diagnostic: Understanding the cause of a fault or issue
- Maintenance: Predicting and adjusting maintenance intervals to optimize scheduling
- Efficiency: Improving the performance of the production or the utilization of resources
- Prognostic: Providing insight to avoid faults or to maintain efficiency
- Optimization: Optimizing resource consumption or compliance with local government regulation
- Logistic and supply chain: Monitoring and optimizing delivery
In the IoT, from the technical point of view, we can identify two broad categories of analytics:
- Physics-based: Based on mathematical formulas or knowledge expertise
- Data-driven: The model is built using past data
Physics-based and data-driven analytics can be combined to build a reliable hybrid model.
Recently, the introduction of deep learning (a branch of machine learning) in the contexts of image and audio processing has brought a lot of attention to data-driven technologies.
Artificial intelligence is nothing without data; the IoT is nothing but data.
We are now aiming to expand the application of deep learning to the I-IoT to improve speed and accuracy in data analysis. In addition to audio and image data, IoT data can be processed with deep learning based on learning, inference, and actions.
However, there are two drawbacks:
- The abundance of false positives that are produced by these techniques
- The fact that companies do not always understand the outcomes of these techniques
Resolving both of these issues will ensure that an abundance of caution is built into machine learning models used in industrial applications. We need to not only create better algorithms, but also make sure that people with domain expertise understand machine learning suggestions. We also need to build systems that take in feedback, and are aware of the end user and the effects of a good or bad response.
From an infrastructure point of view, we need to shift from on-premises to cloud computing, and to provide a platform for data analytics in the cloud. This is known as Data as a Service (DaaS).