Predictive maintenance and decision support systems in heavy industry
Digital transformation is one of the top priorities for industrial companies. The largest players are already moving in this direction, for many years continuously working to improve production efficiency and launching large-scale optimisation programs. The adoption of new analytical systems for industrial enterprises ships under a lot of different names. They’re called advanced analytics or digital innovation, and at their core, the technology could be summarised under artificial intelligence. In all cases, the efforts to utilise AI models or data analytics systems are part of a bigger digital transformation effort of the progressing companies.
In an industrial context, such strategies for cost-saving and process optimisation often start from pilot projects, or top management directives for digital change guide them. In general, changes in processes or investments in capital-intensive and competitive industries require large sums of money. Traditional capital expenditures usually stretch over a long period, so a current financial standing may not allow for a complete physical overhaul of the plants or facilities. These high costs lead to the search for cheaper alternatives. Investments in digital and artificial intelligence solutions could be a magnitude smaller.
When creating a transformation strategy, you can find a place to use AI in all operations. Some of the scenarios provide the opportunity for a quick return on investment, such as decision support systems that accompany technological processes, while others require lengthy testing. In this article, we will look at popular and promising use cases of machine learning applications in industry.
Independent of the specific industry, a trend for more cutting-edge artificial intelligence is showing in the last three years, especially in the field of deep learning. Collaborations with technology providers and machine learning startups are more often found. The focus of researchers has shifted from big data analysis to an edge computing. However, industrial companies are still reluctant to adopt advanced data analysis methods.
Benefits from AI implementation can be found in general market trends. In a new report, Meticulous Research notes that the AI in manufacturing market is expected to grow at a CAGR of 39.7% from 2019 to 2027 to reach $27 billion by 2027. The AI in Oil&Gas market was valued by Mordor Intelligence at $2 billion in 2019 and is expected to reach $3.81 billion by 2025, at a CAGR of 10.96% over the forecast period 2020–2025.
Demand prediction
We start with a look at Big River Steel from Arkansas, USA. This steelmaking startup teamed up with an AI consulting firm to improve its profit margins and processes. They have chosen a data-centric approach that utilises historical demand for steel, macroeconomic data, and activities of their customers. They developed a demand signalling application which enhances traditional big data application with machine learning algorithms. This system consists of three parts. Demand sensing, statistical demand, and consensus demand lead to better supply management and a decrease in safety stock. Due to wide fluctuations of demand in the steel industry, new analytics offer great potential.
More insights are given by the automotive industry. A case study of the automotive industry explains soft-computing approaches for demand prediction. The used models are fuzzy logic and Delphi method, plus time-series ANN (artificial neural network). This structure combines expert opinions with artificial intelligence.
Automotive is strong in demand prediction, as the Volkswagen AG Data:Lab Munich shows. They initiated over 100 dedicated projects to forecast sales for different products and regions. The development of the necessary tools that utilise context-based data like growth forecasts, economic sanctions, and weather conditions is one of their key efforts. According to the estimates of the consulting company Capgemini, large manufacturers of structurally complex auto components can increase profits up to 16% due to the large-scale implementation of artificial intelligence.
Another vendor, Predictive Layer offers a solution for power consumption prediction. They developed a dynamic pricing engine that includes demand and supply elasticity analysis of the client company. According to the company, it was possible to form a fairly accurate forecast of consumption for the next day. The declared savings of the annual electricity purchases in the national electricity markets of the EU is more than $45 million.
Oil & gas industry is using conventional and artificial intelligence-based models for energy consumption forecasting. And they use a broad spectrum of models for forecasting different horizons. A review from 2019 showed that conventional models are preferred for the yearly energy consumption forecasting at the national level. AI-based models could be applied in full-scale in all forecasting horizons and areas. However, in the short-term forecast regression models and time series models can compete in forecast accuracy with AI-based models. As these are white-box models, they are superior in explainability of the relationship between consumption data and influencing factors. The review also showed deep learning models were only used sparsely so far, and their performance and robustness need to be further validated. This lack of use is that even though some researchers are indicating the superiority of deep learning models.
Risk management and predictive maintenance
The review shows a steep increase in publication on risk management AI in the last ten years with doubling it from 2016 to 2018. The leading industries for the application of AI for risk management are automotive and construction. The Oil&Gas, Mining, and Energy industries are on par. A safety science review from 2020 gives a lot of insights into the use of artificial neural networks in the risk identification phase during risk assessments.
Risk assessment is defined here as “the overall process of risk identification, risk analysis and risk evaluation.” The two main data sources for the application of artificial intelligence are learning from textual data and numerical data. The textual data for risk management consists of written reports of safety incidents or accidents. Large structured questionnaires or unstructured free text are the two forms in which the data are available. So, this kind of data source is ideal for enterprise AI analytics. The numerical data comes in the form of accident frequencies or other time-series data. Here also a wide variety of models are applied. An emphasis lies on classification and regression trees (CART), decision trees and SVM.
A practitioners perspective shows a Petroleum System Risk Assessment as an available service from a technology company that leverages exploration workflows. They help robust decision making with artificial neural networks. They focus on enhancing existing numerical modelling with the determination of the uncertainties, and they create maps of risks at exploration scale and in time frames compatible with operations. They use expert knowledge in combination with continuously updated regression and clustering models.
Predictive maintenance is a hot topic now, but development of this area requires high quality data and a systemic restructuring of maintenance and repair processes. In the Oil&Gas industry, these anomaly detection systems are significant applications of machine learning. The widespread sensors control hardware fosters the use of this kind of analytics. Defect on turbo machines, pumps, and motors could be recognised early, and so it’s possible to prevent further losses by converting unplanned repairs to scheduled ones.
Benefits applying data analysis were clearly demonstrated in creation of oil and gas platform management system by BCG. Obtained operational data helped to prevent shutdowns of dynamic equipment, and assess the achievable potential of its operation. In addition, process control opens up new possibilities for dealing with production bottlenecks.
Early anomaly detection in a technological process
Undoubtedly, one of the key priorities for all industries is safety, exclusion of humans from hazardous environment, online operation monitoring and tracking violations of equipment operating modes, as well as prompt response to key risks.
One of an artificial intelligence system’s inherent capability is anomaly detection. It could be seen as one of the main objectives of the Industrial IoT. The identification of abnormal behaviour within the pool of collected data. Known process patterns could be interrupted by rare events that are usually undetectable by a human expert. As an example for the industrial application of such an anomaly, a leaking connection pipe could be representative. In an unfortunate case, this defect could lead to the shutdown of an entire production line. With the huge data streams inside production plants, the hunt for anomalies by manual inspection seems unreasonable.
Ukrainian technology provider Sciforce shares a case study of their manufacturing client who wanted to speed up regular processing algorithms and to increase the system stability. They set up a commercial anomaly detection process which includes the anomaly detection itself and the prediction of future anomalies. As a model for detection, they used autoencoders. These encoders map given data into a hidden representation. Then they attempt to restore the original input from this internal representation. For the prediction part, a recurrent neural network (RNNs) was in service. This model could make accurate predictions for up to 10 minutes ahead.
Pricing
The epidemic-related crisis has further exacerbated the long-overdue need for industrial digitalisation. Due to negative oil prices, oil companies found themselves in an extremely unstable environment, forcing them to rethink their strategies and operations. When we look at the supply chain management of Oil&Gas industry, there are several points where artificial intelligence could be applied. AI helps to predict market prices of crude oil and finished products giving more insight to adjust a pricing policy.
The research on AI-based pricing models is an ongoing effort. A review from 2015 shows that all classical artificial intelligence models are used for oil price forecasting, from artificial neural networks, support vector machines, wavelet, genetic algorithms to hybrid systems. The researchers concluded that due to the complexity of the oil price influential factors, the models are very limited. There are no single indicators driving oil prices. And the needed input data is still to discover.
Despite the unpredictability of oil prices, the slightest advantage needs to be used in a fierce competition. The expensive arms race forces traders to keep up without clear benefits of the solutions. An interesting development comes from the commodities market. There, the price risk management could be changed soon through new technology that fundamentally alters how small and mid-size enterprises handle their commodity risk. AI trading systems with autonomous farming and mining operations could minimise price risks.
Logistics
The future management of raw materials, logistics, and transportation are closely connected to the concept of the smart factory. When looking at the industries that lead the digital transformation process towards complete digital production, the automotive industry is again the digital master. They are capable of increasing their profits through digitisation efforts at double the margin of beginners. For logistics, the connected plants and factories are key to successful projects. At the scale of global acting companies, the existing communication channels need to be used for a unified digital logistics system.
One example of a smart factory solution already in service is general electric’s brilliant factory program. GE powers 20 Brilliant factories worldwide and links real-time data from design, engineering, manufacturing, supply chain, distribution, and service activities. They altogether build one interconnected intelligent system. DHL and IBM describe the function of AI for logistics as an interconnection to robotic process automation (RPA) where AI learns to copy and improve processes based on data provided by RPA. Artificial intelligence systems become the assistant of logistics based on human decisions. They can make fast judgements and interact with humans. Beyond RPA the AI systems help logistics to shift its operating model from reactive behaviour for forecasting and proactive operations with predictive analytics.
Production processes
From robots to cobots, smarter automation, with artificial-intelligence-driven management and control systems AI, is changing the processes in the factory. An example of more process-driven AI is the optimisation of additives consumption is a lot of industrial processes. Despite the fact that most processes have undergone a year’s long traditional optimisation process, machine learning can push these optimisations even further. Complex processes feature still a degree of uncertainty. So expensive additives were used for a broad safety margin. AI can not inspect deeper initial material composition, quality of raw materials, readings from hundreds of sensors and come up with meaningful reasoning. After training the AI model, it can accurately predict the parameters of the final product.
An example of the application of such systems is the optimisation of the additives use in many industrial processes, for example, the use of ferroalloys in metallurgy. The case of NLMK is distinctive, revealing the capabilities of machine learning in predicting the chemical composition, when adding certain materials. People are often reinsured and use expensive additives in excess to ensure certain properties of the product. The use of machine learning methods helps to carry out a deeper analysis of the initial composition of the material, the quality of raw materials, streaming data from hundreds of sensors, to predict the properties of the output product and recommend the necessary actions.
In 2019, Accenture presented a financial model for a typical chemical plant AI implementation. In their example, a business with $11.3 billion in annual revenue could increase its bottom line by 10%. The technology provider bitrefine offers a general solution for a broad range of processes. With a prepared set of machine learning tools, ranging from classification, cluster analysis, regression analysis, structured prediction to reinforcement learning, they serve whole process industries. They promise to optimise ore beneficiation, gas processing, oil refinery, chemical industry, plastics industry, glass manufacturing, and semiconductor manufacturing processes. When, through digitalisation, we can increase the speed of steel casting or reduce the amount of scrap, this is a direct income increase.
Quality control
Decision support systems also include quality control, which help to visually control the properties of products and build the logic on working with anomalies after. Here, AI is again the core of such systems. Using deep learning techniques, AI can detect defects better and more accurately than humans. According to McKinsey, AI can be used for automated quality testing. So, productivity could increase up to 50%. Advanced image recognition techniques for visual inspection and fault detection can increase the defect detection rates by up to 90%. This increase is a significant improvement over manual inspection.
The main approach for visual inspection is a supervised learning algorithm-based system. Here deep neural networks learn to differentiate between good and defective products. The visual system is trained with views from different perspectives, so previously unknown defect types can be identified by employing semi-supervised learning. In this approach, both labelled and unlabelled data is used.
This visual inspection and quality assurance can be used for a large variety of products. The range reaches from machined parts, solar panels, painted car bodies to or textured metal surfaces. Practical examples for automated visual inspection are the approaches by DevisionX. They offer solutions for the automotive industry. Here an industrial camera or sensor is detecting the defects. It can be used to inspect for welding, stamping, assembly, or size conformity. The captured image is analysed through machine learning algorithms, and the system can take action and delivers real-time and customised reports.
New practical applications outperform traditional artificial intelligence
The use of new deep neural networks and other new machine learning concepts have emerged in recent years, and their application is thriving. But they’re widespread in other industries than the heavy industries. Here in Oil&Gas, mining, manufacturing, and metallurgy and metalworking more traditional AI still prevail. And the automotive industry drives the adoption of the newest models.
The different applications show that all existing artificial intelligence models are used over all the industries. However, most of them are still in pilot phases, or they were promoted by technology providers. The vastly available stockpile of data is only slowly becoming available for new artificial intelligence projects.
Big companies like Royal Dutch Shell PLC, Tata Steel Europe Ltd., General Electric Company, Volkswagen Group, and Bayerische Motoren Werke AG are promoting huge enterprise-wide digital transformation. These strategies include a lot of smaller AI projects that try to grab the low hanging fruits first.
Academic research overviews show the maturity of the new AI technologies and the reluctance of the industry for adoption. Some startups are showing the successful implementations of cutting-edge technology and are so leading the way.
_____
This article was originally published in AI almanac.
