The topic of artificial intelligence (AI) has been on the rise in Germany since 2018. More than two-thirds of German companies say they see AI as the most important technology of the future. And yet, only 8 percent of German companies are using AI applications so far. These are the findings of a study by Bitkom, which surveyed 600 companies from all industries at the beginning of 2021.
Data management as a prerequisite for AI
In addition to money, personnel and time, which most organizations cite as reasons not to engage in AI right now, data management also plays a major role. In fact, the Industrial Internet of Things (IIoT) is often portrayed by vendors as a kind of “magical process”. The collection of data in combination with complicated mathematics immediately leads to results that serve to improve processes. But it’s not that simple: To run a successful AI project, the following steps must be followed:
- Define goals
- Collect data
- Understand and prepare data
- Create model
- Refine model
- Operate model
Collecting, understanding and preparing data – these first three steps form the basis for a machine learning model. But it is precisely here that challenges arise that can lead to the failure of an AI project.
6 Challenges on the way to AI
- Lack of data
According to a McKinsey study, lack of data is one of the most common reasons AI initiatives fail. In order to properly program machine learning algorithms, a data collection strategy must be in place to properly classify the data. If this is not the case, data must be classified manually, which is time-consuming on the one hand and leads to increased error rates on the other. In addition, companies report that they don’t have the right data for their AI target.
- Training leads nowhere
Another problem is that companies often do not have a clear goal in mind that they are pursuing. As a result, datasets are used that contain irrelevant or poorly diverse data. If the AI is trained with this irrelevant data, it can lead to the failure of the project. You should also develop a formal process for bias-free training of AI systems and increase your data base.Otherwise, false interpretations will occur due to lack of data or low diversity of data.
- Incomplete data integration
Your organization has data that is available, but it is stored in multiple locations? In this case, the problem is incomplete data integration. Because when data is stored in too many places, it’s hard to keep track of it all. Data scientists are thus busy searching for the data and cannot concentrate on their actual work, the analysis of the data. In contrast, a singular source of data is not enough for your AI system. The solution is to integrate multiple data sources to create a foundation for a well-functioning AI system.
- Different forms of data
To train AI systems, active and transactional data (= real-time data) should be used. If, on the other hand, you are using historical data, problems will arise and the system will not work reliably in interaction with real-time data. The data used must still be “fresh” enough to fit into the production process, after the model has been deployed. Changes to sensors or products must also be taken into account.
- No inclusion of unstructured data
According to a study by Deloitte, 64% of companies avoid using unstructured data and rely solely on structured data from internal systems and resources. However, using unstructured data, such as image or text data, in combination with structured data can lead to better results.
- Cultural deficiencies
Organizational challenges also often stand in the way of AI success. Just as when starting an IIoT project, there are fundamental prerequisites to launching an AI project. Those who are involved must have the intention to change something and have ideas about what exactly can be optimized in their own company. The necessary understanding of processes must also be present and employees from different departments should always work together. At all times, the overall context must be in focus and all areas should have the goal in mind.
Running a successful AI project – with scitis.io solutions
Data is the key to a successful AI system. This is also evident in the challenges described above: the first 5 issues are directly related to data. This is exactly where scitis.io’s solutions come in: Collecting data and generating knowledge from this data, that is our job. Here we can support you in the best possible way and help you especially in the areas of data shortage, training, data integration, data forms and unstructured data. But scitis.io can also help your company with the challenges caused by cultural deficiencies – especially with regard to process understanding and target specification – and support you, for example, with workshops and experience.
Insight into one of our AI projects
But what does such an AI project with scitis.io look like? To answer this question, imagine the following scenario.
A manufacturer of paper has difficulties with the quality control of its products. The quality is tested based on some specific quality parameters. However, since the quality parameters are only measured at the end of production of each reel (paper roll) and no real-time values are available, production staff at the relevant machines can only intervene too late, which in turn can lead to a loss of quality. This is exactly where the AI solution from scitis.io comes in. The developed AI solution makes it possible to predict values for the quality parameters minute by minute during the production process on live sensor data within an error range of 5%. Now, parameters can be adjusted directly during production. This enables the customer to achieve resource-efficient production that is precisely matched to the required quality, which in turn is financially beneficial.
Here you can see the individual steps that lead to a successful AI model:
Finally, all that remains for us to say is, tackle the topic of artificial intelligence together with us, because we are convinced: AI is the most important technology of the future!