Determine your company's AI readiness
Ready for a new era of AI
It has been possible to use artificial intelligence (AI) on the part of companies for a long time. Nowadays, this option is becoming more and more necessary. But what exactly is AI, how can companies use it and where does it even make sense? Dario Luipers from Mittelstand-Digital Zentrum Rheinland provides answers in this guest article.
Thanks to the latest achievements in AI research and development, several AI tools are available that can be integrated into everyday working life at any time. AI-based language applications based on so-called large language models (LLM) automatically generate texts based on questions or bullet points, abbreviate them or formulate the content in the desired language style. AI-generated texts can even be used for initial consultation with customers. The current example of this is ChatGPT , which many companies are currently implementing as a professional text creation tool. AI-based image applications, such as Stable Difusion, are used to produce new images. Companies like to use such visual innovations for advertising purposes or to create presentations.
From a value creation perspective, data-driven analyses and process support are the far more sensitive areas of application for AI, especially in production. Of course, this requires data that teaches machine learning algorithms to recognize patterns in the data structures. What is generally referred to as AI are mostly so-called deep learning models or deep neural networks, which were constructed based on the model of neuron connections in brains. In some cases, however, simpler algorithms, such as linear regression, can also be used. The goal of these algorithms determines the area of application.
Train machine learning models
A distinction can be made between regressions, classifications and clustering. Regressions or predictions can be applied to various scenarios. This includes predictive maintenance, which can predict machine downtime, or, in the case of predictive quality, the product quality of the component to be produced. Classification is increasingly used in the automatic inspection of components using machine vision algorithms. In the case of clustering, the algorithm attempts to create data categories. This makes it possible, for example, to determine combinations of machine parameters which lead to increased waste. All of these applications need well-trained machine learning models. In this case, “good” means that the statements, or the conclusions drawn from the input data, are consistent. There are two criteria to achieve this:
Data quantity and data quality
The quantity of data for learning AI should be as high as possible. The more, the better is the motto here. The higher the variation in the data basis, the more different examples are available to the algorithm to learn from them. It is important that no outdated data is used, which may no longer reflect the current state of the process. This aspect is directly linked to the issue of data quality. The data that the algorithm uses to learn should be as “clean” as possible. This means that there were no gaps in data recording due to failed sensors, for example. The data must not be too “noisy.” Here, the standard deviation is a good indicator, it shouldn't be too high. It is best to remove outliers from training data. Non-scalar values should be uniformly defined. This is the case, for example, when evaluating product quality.
Standardised procedure
If all of these parameters are correct, it should be possible to train a machine or deep learning algorithm on this data. It is worthwhile for the company to proceed methodically in order to Really use AI in a targeted man . AI models can help here because they provide concepts for targeted implementation of AI In the company. The so-called CRISP-DM (CRoss Industry Standard PRocess for Data Mining), a standardized process model for data mining projects. With CRISP-DM, the individual work steps can be precisely defined. The main phases are summarized below.
Business Understanding
AI is not an end in itself, but must be tested for the appropriate use case before it is used.
Data Preparation
This step ensures data quantity and quality. It requires the most working capacity and forms the basis for a functioning AI model.
Evaluation
After the AI model has been trained, the results of the model must be examined in detail. Does the model not produce the expected results? What could be the reason for this? By answering this question, important insights can be gained about the process that is to be supported by AI.
Deployment
If the previous steps were successful, it is time to use the AI model. It is important to record this process with easy-to-understand and concise documentation. This also includes the new and AI-based process that is now in force. This must be maintained and lived.
People and processes
From the time of deployment, the process must be maintained and maintained. The algorithm must now be continuously supplied with data. For this purpose, an appropriate data pipeline should be implemented. This ensures that the algorithm receives data in real time, for example from a machine, in order to be able to make forecasts about product quality or the time of maintenance. Just like when training AI, the data must be of high quality. Incorrect values or faults in the data collection process make the expensive algorithm unusable. The AI system must be regularly retrained with new data. Otherwise, at some point, fluctuations in the process will no longer be correctly interpreted by the algorithm. The results of AI should be regularly evaluated in order to identify problems in the process at an early stage.
The human factor should also not be forgotten. In the end, processes are still implemented and lived by people. It is therefore important to bring those responsible for AI projects and local employees together at an early stage and to consider all interests and aspects relating to the process to be optimized through AI.
An implemented project involving the author of this article involved monitoring the painting of a component using AI-based image recognition. Defectively painted areas should be identified and therefore rework or scrap should be automatically recognized. The outlined steps were all carefully implemented. A short time after implementation, each component was marked as scrap. The reason for this was the change in position of the installed camera to record data. Employees felt that they were watching their work, even though the task area was not directly related to quality control in the painting area and the camera system did not film a human working area.
AI and communication
This experience shows how important early communication is in AI projects. In the best case, the goal, as well as the affected work areas and processes, must be communicated with the entire workforce. Fears, such as losing a job, must be taken seriously and discussed. These are usually unfounded or the company is generally at risk because it does not sufficiently prioritize the issue of AI and data generation and utilization (AI readiness).
AI will have to be part of every company. In order to be able to really use AI effectively, data is the basic building block, which is of high quantity and quality. In addition, an infrastructure must be available to be able to record and store data in real time during operation. From this point on, even without AI, purely through analyses, there can usually be enormous added value for companies. All this is often not possible without external help. In addition to many service providers on the market, there is also vendor-neutral and free initial support through funded projects. One of them is the Mittelstand Digital Center, which provides initial assistance in the area of AI through funding from the Federal Ministry of Economics and Climate Protection.
This article was first published in a modified form in issue 02/23 of our magazine data! You can find all issues and articles here:
