It is time to acknowledge the obstacles to delivering significant impact from AI in manufacturing.
One of the main causes is a common mistake: AI and data tools have been neither designed nor marketed as services for factory operators. The result is poor adoption and, in the worst cases, mistrust. Humans will play a central role in factories for a long time yet, and their needs and experience as “customers” must be carefully addressed, both when designing AI and data tools and when rolling them out in the plant. This is a key success factor for the large-scale adoption of AI solutions in the coming years. This short paper is based on Emerton Data research and analysis, and on many interviews with industry experts and solution providers.
It provides an overview of promising AI use cases and solutions for manufacturing, and detailed guidance on unlocking their value.
Introduction
“The only success of Industry 4.0 is that it is on everyone’s lips.”
This sharp comment from Ulrich Grillo, President of the Federation of German Industries from 2013 to 2016, shows how hard it is for manufacturers to benefit from new technologies, and from data and AI solutions in particular.
Despite a widespread move towards digitalisation over the last five years, and the massive data collection that came with it, progress on the production line in terms of productivity drivers (quality, time, automation, etc.) is still scarce. Solution design and industrialisation have been poorly addressed, and much of the collected data is still waiting to be processed into valuable AI-driven insights. Our analysis shows that fewer than 30% of manufacturers actually have an AI development plan for their factories, although over 85% believe they need to apply AI to their production processes.
What explains this disappointing diagnosis? Did experts focus on the right use cases? And has industry fully learned from more mature, data-centric sectors such as internet businesses, retail, insurance or banking?
One of manufacturers’ main objectives is to increase productivity and, contrary to common belief, the key to productivity is not full automation. A recent example is Tesla’s Gigafactory for the Model 3, which missed its production goals in 2018, a failure attributed to excessive automation, with human operators limited to inspection tasks. More generally, the Industry 4.0 trend has helped show that human operators are the main lever of productivity and overall value. An estimated 70% of tasks are still completed manually, generating most of a plant’s value, and experts expect this to remain stable over the next ten years. With high labour costs and a shortage of qualified operators in most Western countries, humans also account for most costs on the production line.
The human factor is therefore a decisive criterion for the design and implementation of any AI solution. Operators and plant managers should be seen as customers of AI services, and AI solutions must improve their experience and engagement as customers. Creating AI solutions for manufacturing should take inspiration from B2C product development, with strong marketing and adaptation efforts at every step of the project. This is probably the best way today to make full use of the large volumes of industrial data collected, and finally make Industry 4.0 more than a buzzword.
This paper first describes the major pain points manufacturers face. It then identifies the most promising use cases for the next decade and gives an overview of the AI-for-manufacturing solution landscape, with a strong focus on start-ups rather than established players such as IBM or Siemens MindSphere. It concludes with recommendations and best practices to help manufacturers capture the full potential of data and AI applications.
Data quality and governance, the top two pain points cited by Industry 4.0 experts
Asked about their most pressing pain points, manufacturers mention two main difficulties: human management issues and data issues.
Together, these suggest that today’s major challenge is to bridge the gap between data and AI solutions that are useful and well adopted by people. These pain points are key obstacles in many industries (assembly, process), and over 70% of interviewees rank them among their top three.
Exhibit 1: top two pain points cited by industrial experts.
Human management issues
Staff management is the most common human-factor pain point in European industry. It is driven mainly by very high staff turnover and demanding craft skills that require long training periods.
Over 40% of our interview panel cited turnover as a major pain point. A recent report by the European Commission’s Directorate-General for Employment, Social Affairs and Inclusion and the European Network of Public Employment Services (PES) confirms this, finding that craft and related trades workers (mainly in metallurgy) and plant and machine operators face shortages of high magnitude. Retaining a workforce is therefore a strategic priority for most manufacturers.
This has two consequences. First, companies struggle to find the qualified workers they need. Second, even when they do, the specificities of each industry require additional training. The length and cost of training vary across industries: up to two years in semiconductor manufacturing, for instance, which is also time-consuming for the experienced workers who train newcomers. Over 40% of the manufacturers in the study mention long, expensive training. The use cases described below, if well executed and designed as services for workers, are expected to cut training time dramatically and improve retention.
The second issue is the unpredictable quality of operators’ output, one of manufacturers’ main concerns about human work. Human operators can perform tasks that still cannot be automated, but their consistency is weak compared with robots. Over 90% of the manufacturers interviewed cite product quality as their top concern, and half mention the difficulty of maintaining consistent quality standards among their top issues. Many factors cause variability, from external conditions (temperature, pressure) to site specifics (machine age and type), but the quality of human work is the main source of variability in many industries, especially those requiring highly skilled operations such as metallurgy. This variability means many tests along the production line to check product quality, which are costly and slow production down significantly. Sometimes quality cannot even be tested along the line, because major defects only appear at the final stages (e.g. in tyre or aluminium strip production). The result is high defect rates and significant raw material losses, which explains the strong emphasis on quality in the use cases below. A key success factor for controlling the quality of human work is to avoid AI solutions that are too intrusive or controlling, and to focus on formats that workers perceive as saving time or improving performance.
Data issues
Factory data is seen as a core business asset, and over 60% of interviewees mention data issues. Two main concerns arise: data quality and data governance.
Data quality is the most immediate obstacle to a successful data solution. Digitalisation brought huge progress in data collection, but its lack of focus often makes the data unusable for AI. The variety and volume of IoT data illustrate this well: in much of the feedback we collected, raw data was unsuitable for specific AI use cases (for example, algorithms need continuous data while many IoT sensors output binary variables). Data lakes that took months to build can end up useless, and several manufacturers believe they may have to redesign them from scratch, often adding new sensors to the production lines.
Data collection, visualisation protocols and algorithm replicability are also hard to standardise across sites, because industrial sites vary so much.
Data governance clarifies ownership, confidentiality and privacy. Manufacturing generates less personal data than consumer businesses, so regulations such as the GDPR may seem to have less impact. Yet governance still has to address two challenges in manufacturing: integrating sensitive workforce data, as operators are increasingly observed by machines, and protecting core business know-how. The first is gradually leading to automated monitoring and assessment of operators, which creates strong fears on production lines and could easily lead to massive rejection of the solutions. The second was raised by many interviewees: data-driven monitoring of production lines reveals valuable insights into industrial recipes, threatening highly strategic trade secrets, especially in industries such as tyres or defence.
Beyond predictive maintenance, three main use cases are emerging: AI industrial robotics, production process optimisation and vision control
To address these pain points, AI can bring value through three generic use cases:
- AI can greatly empower the robots that have started to equip factories in recent decades.
- Vision control can benefit from the tremendous recent progress of deep learning in processing any type of image.
- Production process optimisation is a natural next step after the first wave of digitalisation.
Although less mature than predictive maintenance, these use cases were consistently ranked first or second in the expert interviews conducted for this analysis.
AI industrial robotics
AI industrial robotics combines machine learning and industrial robotics in a single system used for manufacturing. Among the start-ups screened for this analysis, it is the leading use case in terms of total venture capital investment in AI for manufacturing, with 30% of total investment (see the overview of VC activity below), partly because of the high cost of materials and hardware. One of the fastest-growing players is Brain Corp, which has raised $125 million to develop AI that turns existing manual equipment into intelligent robots. AI embedded in industrial robots addresses the main limitations of the current generation: safety close to humans, quick reconfiguration and the ability to grasp delicate items. Emerton Data analysis identifies three main focus areas today.
The first is communication, to develop cobotics (collaborative robotics). To improve adoption, workers need to interact seamlessly with machines, through better communication with robots, using voice or gestures to control robots or guide human operators. Better communication can reduce the cost and duration of training, which matters given the cost of turnover. It can also reduce quality issues caused by human error at non-automated stations, and increase the productivity of operational teams and equipment use. Manufacturers mentioned many unsuccessful projects with poor interaction design, such as tablet-based dashboards and controls for operators who cannot easily use tablets. Simsoft Industry, for instance, addresses this with intelligent voice assistants for industrial technicians.
The second is navigation and picking. The objective is to let robots optimise their movements in the factory and navigate non-standard environments. The main challenge is handling non-rigid and moving objects. These capabilities increase collaboration between robots and humans, which improves safety in the plant and reduces inspection costs, and therefore overall productivity.
The third is learning: programming robots through physical or video demonstration, and enabling trial and error and group learning, like parallel computers. Better robot learning reduces programming time and costs, and ultimately makes these technologies accessible to small and medium-sized companies.
Overall, AI embedded in industrial robots is expected to directly improve uptime, productivity and safety, and to reduce menial labour by limiting non-ergonomic, repetitive tasks. Here manufacturers could benefit greatly from emulating winning models in the consumer space. Autonomous cars and voice assistants such as Amazon Alexa show how AI can unlock productivity, engagement and collaboration with hardware, and we believe this can be replicated in many manufacturing use cases.
Production process optimisation
Production process optimisation gives manufacturers full visibility of their plants, lines, machines, products and process data, improving productivity, throughput time and quality. The industry leaders we interviewed always cite it as their first or second priority for future implementation, starting with quality improvement. It is a master use case, expected to have a direct impact on the heart of the factory, with two main sub-use cases.
The first is production settings optimisation. Processes are often complex, with many steps and parameters such as raw material quality, pressure settings, weather conditions or output quality. Machine learning applied to this information uncovers correlations and optimal combinations of variables that no human could find alone. The expected impact on productivity, time and quality is huge. Capturing the real value of process optimisation is a challenge, though: full digitalisation of the production line and strong business involvement are key success factors.
To ensure proper industrialisation and return on investment, leading players such as Seebo or Sight Machine build AI algorithms on top of digital twin platforms. The digital twin is the only way to make sure these solutions provide contextualised recommendations and rich investigation environments that drive adoption. Sight Machine says its platform has helped manufacturers reduce scrap costs by 30% within three weeks.
The second high-impact application is production planning and scheduling. Complex processes must be managed against the availability of raw materials, production capacity and demand. AI finds the best optimisation strategy from machine data and supply and demand data, and recommends optimal production and maintenance schedules. This improves production, maintenance costs, scrap rates and customer service. Flexciton, a start-up in the field, offers technology that it says can save manufacturers up to 20% of operational costs with no capital expenditure.
Process optimisation can directly improve production capacity and on-time delivery, but it leaves some points out. As noted above, 70% of tasks in a plant are still performed by humans, and human management is among the top three concerns of the industrial leaders we contacted. This source of variability must be taken into account in the optimisation process. Companies must integrate humans, either quantitatively, by turning human actions into data with computer vision for example, or qualitatively, by understanding how operators interact with the plant. People still govern processes, so process optimisation products must be customer-centric and perceived by operators as tools that make their work easier.
Vision control
Vision control is a booming use case that combines vision hardware and sensors on the production line with computer vision algorithms. The main objectives are often to automate the visual detection of quality defects and identify their root causes, to prevent low-quality items from being produced in the future. Manual inspection can sometimes be fully automated at every stage of the line. Thanks to recent progress in deep learning for vision, this is the fastest-growing use case in AI for manufacturing.
Vision control makes exhaustive inspection possible, instead of inspecting only a small sample of products because of time constraints. Quality control use cases also drastically reduce low-value, repetitive manual inspection. Engineers on the line spend about 40% of their time looking for the root causes of quality issues, so these use cases would greatly improve how they use their time.
To exploit the full potential of the technology, vision control products should provide a collaborative platform that lets engineering teams find issues easily, investigate failures and implement corrective actions immediately. A key success factor is ergonomic human-machine interaction, with a seamless experience for generating and labelling data, because deep learning algorithms need to learn from a huge number of images of good parts, non-conformities and failures. Instrumental, a leading vision control start-up, reports that its quality engineer clients are twice as efficient during their time in the factory and find on average 20% more issues that would otherwise only have surfaced in production.
Funding of AI-for-manufacturing start-ups in Europe and the US is finally intense
Exhibit 2: overview of start-ups providing AI solutions for manufacturing.
Awareness of manufacturers’ pain points is growing, and start-up activity and funding in AI for manufacturing are intense, with several major start-ups having recently raised capital.
The start-ups in this analysis focus only on manufacturing and have AI at the core of their products. From an initial Crunchbase extract of over 400 start-ups in Europe and the US, 65 met these criteria and were analysed.
They are distributed across seven major use cases, including the promising ones detailed above: industrial robotics, production process optimisation, vision control, manufacturing supply chain, predictive maintenance, facility management and product R&D.
Total investment over the last three years is approximately $1 billion for this cluster of start-ups. In average funding, industrial robotics ($50 million) and predictive maintenance ($20 million) lead the way, driven by big players such as Uptake and Brain Corp. These are also the oldest use cases in terms of average funding date.
Predictive maintenance appears to be the most mature use case in terms of fundraising and competitive intensity. Vision control and production process optimisation are much younger, with medium and high competitive intensity respectively. The process optimisation start-ups that are winning the AI race are building powerful data integration and digital twin platforms with a suite of AI algorithms. Product R&D, manufacturing supply chain and facility management show fewer start-ups and lower average funding, suggesting lower attractiveness in the manufacturing space specifically. However, these use cases are not specific to the core of manufacturing, and therefore attract fewer start-ups in this cluster.
Exhibit 3: average funding to date ($ million) and number of start-ups per use case.
Key takeaways to unlock the full value of data and AI applications
Our study of both manufacturers’ pain points and the use cases offered by innovative players shows that, as stated in the introduction, human operators are not yet considered a real market for AI applications.
Even with many challenges remaining (specialisation, diversity of industries, etc.), we believe that approaching operators as consumers and end customers of AI services is the key to adoption and value delivery. Any game-changing project needs to be well structured and delivered with the right partners, those able to bring four core competencies: UX, data and AI knowledge, business knowledge and change management.
Here are a few recommendations to create more value with AI in manufacturing.
Use AI applications to improve workers’ environment, by targeting stress factors and offering user-friendly interfaces
Turnover is manufacturers’ main human resources concern. Yet the most popular and best-funded AI use cases, predictive maintenance and robotics, have little to do with human operators. Researching the stress factors of workers in a given industry will help decision-makers prioritise the AI applications that address them. Some will not even require AI and will be very simple to set up: in the tyre industry, connecting machine data to smartwatches has already significantly reduced stress for production line supervisors.
Using AI to improve the working environment, however, requires designing the tools together with the workers who will use them. The worker experience is critical. It demands ergonomic interfaces and a variety of human-machine interaction options, as simple as voice commands for workers who cannot easily use a tablet.
Given the cost of turnover in European industry, human operators can be considered a market for AI solutions and should be addressed as such. This means using marketing techniques to understand the real “buying” and adoption criteria of operators, or of their managers, and matching them as closely as is customary in B2C markets.
Use AI applications to accelerate training and improve the monitoring and optimisation of human operations
The unpredictable quality of operators’ output reveals an important gap in traditional Industry 4.0 systems: no data is collected on tasks performed manually without machines. Since 70% of tasks are performed manually, almost no data is collected on that share of the process, and its optimisation and quality rely entirely on training. Most of the production process is in fact ignored by Industry 4.0.
Capturing data on human operations, using vision systems for instance, would help understand technical know-how and share it with new operators, shortening training and reducing the unproductive teaching time of experienced operators. It would also help trained operators improve, reducing variability in output quality. As mentioned earlier, these tools must be developed carefully and with respect for workers’ privacy, just like B2C AI applications in private life.
Bring all actors on board from the start, but only collect data once business experts have precisely defined the use case and the relevant data
The digitalisation frenzy of the early 2010s proved inefficient and caused many of today’s difficulties in building AI applications. Given the complexity and diversity of data sources and the cost of storage, massive data collection is both costly and inefficient. Data and business expertise are too often separated when planning an AI solution. The key to success is to scope the use case precisely beforehand with everyone involved, including IT departments and industrial specialists.
Industrial expertise should lead the definition of the use case and the selection of data, while looking at the data itself. It is critical to make the necessary hardware additions as early as possible.
Finally, one of the major challenges is making the right IT choices to enable scalability and industrialisation. IT and business specialists must work together from the very beginning. Otherwise, the result is very good proof-of-concept algorithms that can neither be industrialised nor bring real value.
To conclude, AI and data-driven applications have been designed around IoT and machine data, while human behaviour has often been forgotten. There are still many people in factories, and they will remain there for quite a while. Let’s make sure we take all the implications of this into account when designing future AI solutions.
