Original article can be found here (source): Artificial Intelligence on Medium
Data science can be applied to a lot of industrial environments to save costs and to improve processes. This industrial AI does not only include the smart algorithms and big data concepts that reside in the virtual space inside the computer systems, but it consists of the physical devices themselves too. Data have to be captured with sensors. Commands have to be sent to actuators and control systems. This whole chain and flow of information, wireless or via cables, goes through places with extreme conditions. From the points of operation inside factories, mines, or oil rigs to the big data storage and huge processing power inside data centers and control rooms there is a long way to go.
Industrial production facilities, physical transport systems, and distribution channels are complex and feature often a zoo of devices from different manufacturers. As automation is far advanced, there are a lot of existing digital control and management systems in place. Today you find data networks, Supervisory Control and Data Acquisition (SCADA), programmable logic controllers (PLC), and Heating, ventilation, and air conditioning (HVAC) in industrial settings. All these systems are looked at on different levels of abstraction. There are concepts and higher management levels of complexity, and there are lower levels stronger connected to physical challenges.
It’s getting rough
Existing industrial installations have a lot of wiring and kilometers of cables. These sophisticated networks keep operations running. You have industrial standards for digital networks that connect devices and switches, that provide gateways and that create connectivity to control rooms. These systems were built to perform well under harsh environments. With the arrival of data science and new global digital connectivity in an industrial context, the new smart devices also need to be able to perform under these extreme conditions. ML is closely connected to industrial internet of things (IIoT) to offer more wireless connectivity and links to enterprise-wide systems and the internet.
All these new systems face the challenge of nature and the powerful forces of heavy industrial machines. Production facilities may be located in faraway places, with no power grid. Imagine operations in the jungle with only one improvised road to give access. You have to bring all the energy sources by yourself via generators or batteries. The environment may be cold or hot, with nasty gases that corrode electronics or make them vanish in an explosion. But the heat and the cold could not only source from nature, but from production processes as well. Melting and freezing workpieces could be part of the production process. A lot of work takes place in the mechanical realm. You have parts moving at high speeds. You have machines that create vibrations and shocks. Data science is now facing these physical challenges mainly in the form of robotics and IIoT networks.
There are robots everywhere
Robots are a complex challenge. They need to do things, and they need to move themselves to where they’re needed. To achieve this without human interaction, and ML brain inside this robot is needed. But, we all remember robots failing in decommissioning nuclear power plants. The problem, in this case, was the radiation destroying the electronics. So, a hardened brain is needed. Further, communication with the outside world is difficult. In intense radiation wireless or wired communication is a challenge.
However, it’s not only the very intimidating nature of nuclear decay that is a problem statement for industrial data science, but also the examples of deep-sea exploration or mining. They’re very challenging too. With a lack of general infrastructure and no fixed power supply or internet available, you need to adapt existing best practices to the digital and data-driven transformation.
You may not find these extreme examples on the edge of production in your everyday production facility. But you’ll find similar challenging situations. The environment and machines in operation created threats are everywhere in industrial production. And the armada of autonomous or ready to be autonomous robots is growing every day.
You have robotic arms, walking and driving autonomous vehicles, and flying machines for all kind of tasks. They come in classes like automated guided vehicle (AGV), unmanned ground vehicle (UGV), rovers, autonomous underwater vehicle (AUV), remotely operated underwater vehicles (ROV), autonomous dump trucks, autonomous haul or mining trucks, Unmanned Aerial Vehicles (UAV), or drones.
These robots come in many forms to be more efficient and cost-effective. They can perform tasks that employees couldn’t do. May these tasks are just mundane longer shifts or may they be working in even harsher environments. They could do inspection whenever and wherever it is needed.
Everything will be connected
The second big application of industrial data science is the connection and aggregation of data throughout the enterprise. All sensor data collected will be stored in one big data lake. A network that behaves like a living organism could be created with the help of industrial IoT. Here the wireless and wired data connections are facing the physical world. Cables, switches, routers, and gateways need to be robust. They need to be able to sustain dust, and vibrations, mist and water, and more dangerous and harmful substances and physical effects. They need to be reliable, and they should be able to operate for years without help from the maintenance personnel.
The new industrial data science also faces integration safety and security challenges. A lot of legacy systems operated in a fine-tuned manner for years. They provide solutions for very good performance in the extreme conditions they face. Their optimization processes have been going on for a year. So, a new industrial data science should not endanger this achieved equilibrium. Another concern is data and operational security. Once the production facilities are hooked to the global internet, attackers have theoretical access to the systems. And as the values and stakes are very high in the heavy industry, this is another extreme reality.
Existing data are bad and broken
When setting up industrial data science, the data sources for creating the smart recommendation and predictions are facing issues. A lot of the needed data in an industrial context are time series data and data that is gathered in harsh environments. That means the single data points could be unreliable. Sensors themselves could deliver imprecise values due to environmental conditions.
We could speak of the “3B” of industrial big data. You need to be aware of these issues when you start a new industrial data science project.
The first B is bad. Most of the industrial data has clear physical meaning. It comes for a myriad of sensors that detect air, flows, or motion velocity. All kinds of noise, humidity, leaks, or levels are acquired. Motion data that reaches hundreds of terabytes per day. All this data may have poor quality due to physical measurements. Compared to data that is gathered inside digital systems, like online purchase and customer data, these data need to be thoroughly cleaned before use. Also, it’s hard to increase quality via the quantity of data.
The second B stands for broken. Data used for training data science models to come up with predictions and recommendations is not featuring clear health states. There are no fault modes or higher-level abstraction of data which indicate the working conditions. This could lead to a lot of false positives and false negatives in the implementation of the ML system.
The third and last problem B is background. In a complex industrial environment, the interpretation of sensor data requires a lot of experience of the domain experts. The emerging patterns can be highly transient, and they need expert knowledge to interpret them. It is very difficult to train the data science models solely by the collected numeric data.
Solving the physical challenges
To address the given challenges present in industrial data science, one way to start is to be aware of the specifics of industrial operations. One should not focus on off-the-shelf data science solutions, but off-the-shelf industrial solutions. Ruggedized components and devices are already available on the market for industrial IoT. As a lot of sensors are available through existing setups the gathering of the data takes place on the existing control systems level. SCADA systems and industrial robots and machines provide the data to be captured. Here specialized IIoT gateways deliver the data to the ML systems. These additional gateways do not interfere with existing control systems.
To do the digital transition talent is one of the most important factors. To keep the costs low and manage the new networks the cooperation with subject matter experts is the way to go. These professionals need to know the industry’s specifics and industrial data science. They help in selecting the right hardware and concepts.
Making the right choices
Extreme conditions are not only the physically demanding environments but also the shortage of resources abundant in other industries. Due to problems in power supply, network infrastructure and internet connectivity, industry-specific concepts need to be elaborated. When the connectivity is unstable and unreliable, solutions that are not depending on continuous network connectivity could be selected. Battery-powered IIoT solutions with battery life of several years and ultra-low power consumption are available. Edge computing provides automatic reasoning and aggregation right inside the devices with no need to uninterrupted connectivity to the data center systems.
ML models will work under extreme conditions with a good design of the industrial IoT infrastructure. The differences between general IoT and IIoT need to be addressed. Despite IoT and IIoT share common objectives, the basic requirements of the implementation strategy are very different. Industrial IoT needs to focus more on reliability and robustness due to power and connectivity challenges.
This article was originally published on AiThority.