Technical challenges for AI in heavy industry

Original article was published on Artificial Intelligence on Medium

Technical challenges for AI in heavy industry

and how to overcome them

Illustrations are made by Giovanna Conforti

The decision has been made. The company we’re looking at has decided to implement a new digital strategy that saves costs and lays the foundation for competitive business in the future. With the help of artificial intelligence and industrial internet of things (IIoT), new CEO is committed to digital change. In heavy industries, there are many challenges to consider. The exemplary mining company we are investigating here is facing the same issues a lot of enterprises from the industrial sector face when going digital.

The initial situation in the industrial sector

Most of the companies are struggling to attract and retain the right technological capabilities. This not only affects investments in new systems and procurement of new technology but also affects the hiring of new high potentials. The main barrier for graduates entering these industries is outdated proprietary technologies, which demands time-consuming and expensive training and limits future job prospects. As a result, the job market stagnates, and older generations are the only people with the knowledge of how specific systems work.

Further, a culture of traditional industrial engineering prevails, which is quite different from the computer science vibe. In mechanical and electrical engineering patents and secret sauces are the core values to protect. In data-driven companies, a lot of the valuable tools are shared and even given away freely under an open-source license. From their viewpoint, the benefits of sharing outweigh the problems for the whole industry keeping data in silos. They make their money with other unique business models and their expertise.

The AI use brings results, which significantly differ from expectations. It is hard to harvest a 50% improvement of some process, as most of the metrics are already high. There are simply not many things to try. This has happened because laws of physics are constant, standard optimization methods are already in place, and further equipment modernisation is expensive. Thus, 2% optimisation is significant progress for the company, which can be achieved by implementing an artificial intelligence solution.

Scepticism towards other industries

A common saying in the industrial sector is that it is unique and very different for all other industries. “You can’t do that here!” is an often-heard quote. Independent of the comprehensible pride of real masterpieces of engineering, heavy industries actually have a lot in common with other industries. The heavy industries’ companies optimise their business in the same way as they do. Even though in other industries, the safety concerns were lower, and high business values were not as strongly at risk as inside the heavy industries, the same basic economic principles apply.

Artificial intelligence and data science approaches were tested in other industries and delivered extraordinary results. These technologies saved money and opened new business opportunities. At the beginning of their journey, they also faced huge technical problems they had to overcome. When you look at the Amazon 15 years ago and examine AWS systems, they were crooked and improvised. Not suitable in any way for a safety-concerned industry. But nowadays, these globally installed systems show reliability unseen before.

The nature of data may be a problem as well. It is quite complex in processing and analysis since the industrial sector has machine-generated data characterized by high frequency and created in a fully automatic fashion. On the one hand, machine learning algorithms may be trained better having a large amount of data. On the other hand, companies have to overcome very specific data science challenges and limitations, while having a lack of skills for processing this kind of data.

Illustrations are made by Giovanna Conforti

Challenges for new technologies in heavy industries

When a CEO wants to apply AI and IIoT solutions in remote and harsh environments, there are some unique realities to be aware of. Mines and other production facilities could be in faraway regions where access is limited. Oilfields might be offshore and only accessible via ship or airlift. Weather conditions are often poor. Networks and internet access are limited in bandwidth and reliability. Harsh environments present specific challenges and limitations. They could create an explosive environment, produce gases and vibrations or be underwater or underground with high or low temperatures.

Further to notice are the machines and systems used in heavy industry. These costly setups were created to sustain and be active for a long time. Often the systems have grown over time. Mines or other installations feature heterogeneous machine parks and setups, and they don’t have a harmonised infrastructure. These machines were not meant to connect to the internet. They were designed to deliver a lot of local data for planning and control. Most of these systems require real-time control. They don’t have dedicated interfaces to extract data for more general purposes.

Some technical requirements for the existing machines also apply to new IIoT and artificial intelligence applications. They need to be able to operate with no light source or no power grid, have to sustain vibration and be reliable. These applications should not create a threat to life, and be able to cope with dropouts and interruptions of all different kinds.

The problem of gathering the needed data

Artificial intelligence systems need the right data sets, with high quality and integrity to create valuable predictions and to optimise operations. So, how to get started with setting up new sensors and refitting old machines to get this data? How could you extract data streams from existing systems, and how do you clean the existing data streams? While digital twins are attractive for the design of mass-manufactured and expensive products, it can be harder to scale the technology beyond manufacturing facilities. Heavy industries operate under a lot of different configurations, and they use equipment from different suppliers. So, a fitting technology for IIoT is needed.

One example of available technology is IIoT provider relayr. The company developed sensors in the form of a chocolate bar to refit all kinds of machines. They were bought by Munich Re for US$300 million and were transformed into an equipment-as-a-service solution provider. However, the problem with these new technologies is still the systems integration. The client still needs proof if the provided piece of technology fits the needs and could be integrated into existing processes.

Illustrations are made by Giovanna Conforti

It often happens that data needed comes from process control and MES systems that were not originally intended for the purpose of analysis. Outdated archive systems often store data in unsuitable formats. At the same time, the level of structuring and systematization of information significantly is poor. Sometimes the data may not be accurate, due to some humans in the loop. Knowing and understanding the process to a certain extent is essential to account for all such obstacles.

The problem of transmitting the data to machine learning units

A lot of technical processes need real-time decision-making and control. Business processes need fast decisions in cases of trading at spot markets or for procurement in time-critical cases or accident handling. How can you continually power the sensors that communicate mission-critical data on the condition and performance of isolated assets that are beyond the reach of the electricity grid? How to transmit all the needed data without existing mobile networks? How to make reliable decisions in unreliable networks that could break down in the most unexpected moment?

These problems are not unknown to industrial companies that were built and operated by engineers. In data engineering, the same principles that were used in traditional engineering apply. Only the goals that have to be achieved are different. Here the flow of uninterrupted data is a priority. The approach to achieve this high-quality data stream is different. Traditional engineers and software/system engineers have very different training and philosophy. Here are cultural differences to overcome.

Sometimes some data with key information can come to the system with a significant delay. For instance, the results of the analysis made in the laboratory become known only after a few hours. Some mistakes also accumulate over time. If some sensors were broken there is still a need to determine a pattern and take them into account, restoring the missing data. Even in production use, models need to be able to operate in real-time and return the recommendations or predictions in a situation when not all parameters of the process are known with precision. There is a need to use a number of tricks to reconstruct and make virtual measurements to ensure that machine learning models are reliable even when the company is operating under uncertain conditions.

A solution to address the challenges

One of the main problems of setting up an IIoT infrastructure is the power supply. With the fast-paced progress in IIoT devices with low power consumption, already available today, it is easy to find of the shelf components to make a start. A compromise is needed between the amount of energy harvested and the amount of data demanded by rapidly developing IIoT applications. For the connectivity needs, there are emerging edge computing solutions helping to overcome this. IIoT and self-driving cars rely on computing power built into the device itself, rather than the bandwidth of and access to the network.

The professional engineers are always at the origins of industrial companies. They are experienced in using numbers and performing experiments. In this way, the process of choosing the success metrics, defining features for the models or designing A/B tests is easier. There is no need to explain why testing is done before going live. Engineers are used to basing their decisions on data, not on gut feeling.

Data sharing and protection

Once the physical infrastructure is set up, data handling has to be discussed. Data has a monetary but intangible value that companies don’t want to share with others. Companies see the risk of losing competitive advantage by sharing data insights. There is a concern that data could get cross-contaminated. The benefits of sharing data, especially in the case of supply chain integration, often surpass the fears of insufficient data protection. AI processes improvement leveraging more available data allows getting more value to the whole industry as well as individual companies. This approach has been proven in a lot of other sectors. This data pool improves AI-powered systems learning results leading to the improved operating model and reduced operational costs. Concerns that data might be lost as a result of a cyberattack is still alive, influencing the willingness to share data in general, neither working in the cloud. Security and privacy are on the top list of technical challenges faced by IoT projects. According to Gartner, this is due to a lack of skilled staff, a complex vendor landscape and immature standards.

Illustrations are made by Giovanna Conforti

Despite all these concerns, innovations are fostered, and new product ideas emerge. Testing ideas on a broader base accelerates development and acceptance by a larger number of companies inside the given industry. The risk is shared, and development effort for each enterprise is reduced. Shared data is anonymous, so it can’t be traced back to the data owner. Research-led initiatives should have a look into differentially-private decision forest algorithms. These algorithms minimise the number of queries on the given data. So less computing power is needed, and the sensitivity of those individual queries is lower. We also see that companies are becoming more open to exchange data with the original equipment manufacturer (OEM), because it is more profitable for operators (for example, in the manufacture of aircraft engines).

Business perspective and scope

The industry-wide saying, “Everything is too expensive! Everything is too complicated!” is not valid anymore. There are good ways to start the digital transformation and integrate valuable AI and IIoT applications to every company. However, there are some hurdles to overcome. The specific challenges of the used machinery and used systems in heavy industries need to be addressed. CIOs and other digital missionaries must prioritise opportunities within the organisation through the lens of business return and technical feasibility — in terms of both absolute value and payoff period. The global research and advisory firm Gartner recommends that the majority of AI and IIoT projects should target financial payback in less than one year.

To make a start, the CEO of the mining company could initiate a first project and pilot phase. Starting projects doesn’t have to involve high levels of skills, people resources, or cash spend. The first steps could include building AI solutions using open source tools. From that experience and the proof of concept, the project could quickly scale up to deliver instant value to the whole company. However, sometimes technical PoCs do not convince executive leadership to spend money on IIoT, so consider allocating some time to prove value via IIoT implementation. It might also be an option to reach out to existing vendors for AI IIoT solutions and go down the rabbit hole with a reliable partner.

Illustrations are made by Giovanna Conforti