Artificial intelligence in manufacturing: applications and implementation tips

According to a recent study by Deloitte, the manufacturing sector is ahead of other industries when it comes to data generation. Artificial intelligence technologies possess an unmatched ability to analyze large amounts of data, so it is only natural for manufacturers to adopt this technology.

Predictive maintenance is one of the most funded applications of AI in the manufacturing industry. Equipment fault can cause significant disruptions, delays on production lines, and increase production costs. One minute of downtime at large factories can cost as much as $20,000. Additionally, regular diagnostics by human experts are relatively expensive. AI-powered solutions analyze equipment’s historical performance data to spot anomalies and predict when it will need maintenance before it malfunctions or comes to a halt. This allows employees to choose a suitable time for fixing the device instead of stopping everything in the middle of the production process when this machine is out of service. General Motors gives one example of AI implementation in manufacturing. The company mounted cameras on its assembly robots and trained AI algorithms to analyze the data streaming from these cameras to identify signs of component malfunctioning. In a pilot test of this solution, it worked on 7,000 robots and identified 72 instances of component damage before they resulted in an unplanned outage.

Forecasting prices of raw material Raw materials costs are volatile in nature. When manufacturers have this information in advance, they can adapt their operations to minimize expenses. A UK-based startup ChAI uses machine learning to forecast price fluctuation of raw materials, such as aluminum, oil, and copper, among others. The company was founded in 2017, and it secured €1.5 million in seed financing in 2020. ChAI targets Fortune 100 companies, including manufacturers, who rely on these materials as a part of their supply chain. Predicting demand AI analyzes behavioral patterns, socioeconomic data, location, and weather forecast to determine which products will be in demand, allowing manufacturers to focus on what matters and cease producing items that no one would purchase. AI can even predict which product will be a hit before they go to the market. Danone deploys machine learning in manufacturing to foresee variability in demand and adjust its production plan accordingly. Thanks to this approach, the company decreased its lost sales by 30%.

Generative design is a program that relies on AI technologies to mimic a human engineer’s approach to designing products. Engineers feed different design parameters, such as size, materials, and cost constraints, into generative design algorithms, which generate different design options for one product. This method allows manufacturers to create hundreds of alternative designs for one item and experiment with how adjusting parameters reflects on the outcome. A human designer would not be able to come up with so many ways of building one item. The resulting designs can be further tested using machine learning to determine which options work best. Considering AI’s recommendations, a specialized workforce will select the design they want to pass to the development stage. For example, Nissan experimented with letting AI propose car designs hoping it will come up with something different. According to the company, their algorithms put forward a design that no one has ever seen before. It was not perfect, but it’s a good start. AI and ML in manufacturing can also assist designers with user experience. Typically, designers try to imagine possible ways the user might use a particular product. With its learning potential, AI can analyze data on how people utilize such products historically to come up with optimal designs.

Products IBM defines a digital twin as a virtual representation of an object or system that spans its lifecycle, is updated from real-time data, and uses simulation, machine learning, and reasoning to help decision-making. To set up a digital twin, you need to collect data from sensors attached to the physical item and project this data onto the digital twin. This way, when you look at the virtual item, you can see what is happening to its real-world counterpart. For example, if you create a plane engine’s digital twin, it will receive data from the real engine upon landing and takeoff. You will be able to evaluate the condition of the actual engine by examining the digital twin. Researchers can use this technology to conduct simulations and anticipate malfunctioning. Manufacturers can also use digital twins to make design modifications tailoring to customer preferences. Manufacturing processes Digital twin technology is not limited to products. You can create a digital twin of the whole production line to optimize the manufacturing process. You will need to position sensors along the production line and use the generated data to analyze performance indicators. Unilever partnered with the Marsden Group and used Microsoft Azure to set up eight digital twins of its factories. Algorithms embedded into the digital twins can suggest production improvements based on the data they receive. For instance, one of the digital twins analyzed the shampoo production process data and could predict the correct order of processes to get the best batch time. Also, using this technology, Unilever reduced the number of production-related alerts by 90%, freeing up operators’ time.

To make sure that products are up to par with quality standards, manufacturers use in-line visual inspection. However, it is time consuming for human employees to examine all products manually. Cameras, computer vision, and other AI technologies for manufacturing can perform a fast inspection in real time, detecting flaws at the earlier manufacturing stages allowing engineers to make adjustments before the product can cause further delays. Audi installed an image recognition system at its Ingolstadt press shop to capture and evaluate the quality of pressed sheets. This AI-powered system was trained on millions of test images and can identify even the finest cracks that could easily escape the human eye. Another example of AI in production comes from a large food processing organization, which produces over 200,000 eggs per hour. Human operators used to inspect these eggs employing the sampling method, but it was prone to errors as inspectors couldn’t spot every damaged egg. Realizing this problem, the company switched to an AI-enabled quality control system. It was trained to identify several defects, including holes, leakage, and cracking in eggshells. This innovative solution can scan one egg in less than 40 milliseconds and spot any of the classified defects.

Recently, Deloitte surveyed the manufacturing sector. The respondents confessed that 91% of their AI projects failed to meet timely expectations. There are things that you can do to minimize the chances of your AI project joining the deck.

It is best if the AI applications you are planning to adopt are in line with your business goals, be it cutting down costs, finding new revenue streams, increasing operations efficiency, etc. This tactic will ensure that business units are involved. AI efforts also need to match your established business goals timeline. Before employing a more advanced AI in the manufacturing industry, check if your schedule can handle the likely delays. Highlight the business goals you want to achieve with AI in manufacturing and specify how to measure improvements. For example, increasing operations efficiency by reducing equipment downtime by 20%. It can help compose a roadmap with the business applications where you want to use AI in the short, mid, and long terms.

Even if you have ambitious plans regarding AI, it is a good practice to start with a few carefully selected use cases. As the company’s capabilities and experience grow, it can expand its AI in manufacturing efforts to more applications. You can prioritize use cases based on their feasibility, total value, and time needed to achieve this value. In his interview with Capgemini, Luis Miguel del Saz Rodriguez, Head of Digital, Design, and Manufacturing Services at Airbus, explained how he approaches use case selection at his company: “First, we organize a team workshop where we discover the pain points and the opportunities. We also consider the scale and impact in the business. Next, we take these pain points, or opportunities, and work on the digital solutions, analyze budget and the associated business case.”

Data is the main foundation of any AI-related endeavor. Your system needs to be able to capture data from different sources in various formats. It must be clean and accessible.

You probably have some legacy manufacturing systems, such as enterprise resource planning (ERP) and product lifecycle management apps that can generate valuable data. Discuss with your vendor the possibility of integrating such software in your AI solutions. You can consider placing a standardized equipment purchase policy. Neeraj Tiwari, Director of Manufacturing JV Organization at Fiat Chrysler, explained how this is done at his company: “We have a centralized process for purchase of equipment, their subsystems, and associated software. This brings a level of standardization and makes integrating AI applications much easier and results in far fewer issues.” It is also a good practice to examine your manufacturing devices and attached sensors. Some of them might be generating data in formats that you cannot use. Forrester Research Analyst, Paul Miller, spoke about such equipment: “Many [devices] may have been in use for a decade or more, and they either have no sensors at all or they have proprietary sensors that send commercially sensitive data in proprietary formats, which can be hard to decode.” Miller also added that such problems have a solution. Some companies sell specialized sensors that manufacturers can fit into their old devices if they know what they want to measure.

If your data is not at the maturity level you need to support AI, it is worth investing in a reliable data foundation. It is paramount for AI long-term success and will allow you to roll out new AI-powered applications in the future. Furthermore, you might want to establish strong data governance practices. This includes determining:

When moving towards machine learning and AI in manufacturing systems, you will need to hire people with specific analytical skills. Limiting talent search to data scientists might not suffice. Your organization will need other specializations, such as data engineers and data stewards. Also, make sure your data experts collaborate with internal domain experts who have a deep understanding of the business problems AI in manufacturing is intended to solve. Some companies initiate upskilling programs for their in-house employees by teaming up with academia and startups. Manufacturers typically begin with fragmented uses of AI experts and slowly move to more coordinated centralized efforts. Some end up establishing AI labs or centers for excellence, which will define best practices of using AI in the company.

When your data is at the desired maturity level, run a proof of concept with your vendor of choice. This will help you better understand what to expect and what you still can fix before a large-scale adoption. Do not forget to integrate AI solutions into the end users’ workflow. According to McKinsey's research, overlooking this step is one of the major obstacles to AI adoption.

When your AI solutions are fully up and running, it is advisable to keep monitoring the results. Assign dedicated staff members to make sure that ML in manufacturing is delivering on expectations, and if not, find out why and what to do to improve the situation. Also, someone will need to adjust AI to any change in your operations. AI algorithms will need retraining with new data categories. Or if you installed the AI system in a different location, it might need to be retrained with location-specific data.

Both your employees and AI need to learn how to do their job together optimally. There is a good possibility that this technology will produce false results frustrating everyone involved. Especially people who were not that excited about adopting artificial intelligence in manufacturing. Siddharth Verma, Global Head and VP – IoT Services at Siemens, shared his AI adoption experience with Capgemini. Here is what he said: “In the early days, when the accuracy of the system was low, it predicted a few failures which turned out to be false alarms. At these points, it is important to remind everyone that it is a prediction which has a probability of being right or wrong. As accuracy improved, the system was able to predict many failures in advance and saved a lot of cost and downtime, proving its worth.”