The new technologies of artificial intelligence, automation, Internet of Things and machine learning are already being used in many companies. With machine learning, IT systems can learn patterns and connections from data and improve themselves. Classic fields of application range from the detection of possible incidents to data-supported decision-making and the interpretation of large amounts of data.
In general, companies can Machine Learning in Industry (ML) Reduce costs and use your employees more effectively. However, the implementation brings with it challenges – including preparing the data, providing the concept or estimating the ROI. the complex technology However, it requires internal experts for implementation, project experience and realistic goals. Therefore should reliable partners be brought on board.
Application fields of machine learning in industry
Industry and manufacturing can benefit from ML in many ways, for example in asset management, supply chain management and warehousing. In particular, through demand forecasts, automatic inventory control and optimization of procurement processes. Intelligent route planning and loading of vehicles reduce their consumption. In addition to the planned and optimal use of all resources, costs can also be reduced through predictive maintenance of vehicles, digital system monitoring and reduce the prognosis of anomalies. The same applies to personnel planning that takes into account performance-reducing factors such as high temperatures and the resulting lack of concentration in good time, thus helping to avoid mistakes in advance.
In quality assurance, video monitoring of package or container contents and automatic recording of dimensional deviations in mass-produced products are also possible. Utilities can benefit from ML, for example, by automating routine activities such as transmission control in the power grid. This releases a lot of engineering potential and at the same time ensures fewer incidents.
Data quality is the basis for the success of ML
Projects with machine learning in industry often fail due to the quality of the available data – they are the only source for the automated process. Inferior data then inevitably lead to unusable results. An ideal database is large (with 10,000 to 100,000 facts), correct, sufficiently comprehensive and appropriately identified. Although this ideal state rarely exists in practice, it can be achieved with the following measures:
Incompleteness: If it is not possible to provide sufficient input data, the problem must be redefined. The same applies to too small amounts of data in certain situations. For example, for automatic route planning, not only the data from a certain region must be available, the special pattern of which cannot be transferred to others.
Wrong labeling: All ML algorithms need correct answers to the question asked for training. The data records must therefore be marked accordingly. However, this should not be left to the technology partner alone, as they often do not have the necessary expert knowledge. What is known as labeling works faster and more cost-effectively with the participation of in-house specialists.
Data errors: There is a risk of a larger amount of systematic errors if the basic data of the ML model was collected manually. Experts should therefore check and correct the data in advance. Because an ML model only forgives individual errors without a common pattern.
Machine learning in industry: It is best to start with prototypes
When introducing machine learning, it is possible to determine before the development phase how much money can be made from the collected data and how much an ML project will bring to the company. By creating a pilot project, one can test the technology and see the potential for applications and tasks. This approach does not make the project a risk, but a very conscious investment with a projected ROI.
There is therefore a lot to be said for initially approaching ML projects as a proof of concept so that they can be thoroughly examined for advantages and disadvantages. What data best represents or interferes with the relationships? What hit rate can be achieved with the model and how can it be improved? According to which criteria can the success of a project be evaluated in a meaningful way?
A prototype also gets by with anonymous data sources, i.e. without passing on sensitive data to the technology partner. Partner companies offer the opportunity to carry out such pilot projects free of charge.
Realistic expectations of fault tolerance and success
The expectations of time savings, hit rate, quality of results and payback time of ML projects are often unrealistically high, while at the same time the problem description and data are far too imprecise. Experienced partners and prototyping can help set expectations at a realistic level.
The error tolerance of the model can also be better calculated if the return on investment is estimated in advance. Every percent more accuracy can cause a multiple in costs. It is therefore often wiser to miss a false negative result than to have to expensively evaluate many false positive reported results. Incidentally, the non-occurrence of an expected event can also provide valuable information about the model.
Every project also needs the right algorithm. However, the creation of expensive and computationally intensive neural networks or the purchase of market research data are often unnecessary. Here, too, one should trust in the experience of the technology partner. The implementation of ML projects in manufacturing can easily fail if it is driven by rapid success, the wrong partner is chosen or the necessary insight into business processes or data is denied. On the other hand, clean data, prototyping, realistic expectations and trust are the prerequisites for the success of a machine learning project in industry.
The author Stanislav Appelganz is Head of Business Development, Consulting & Smart Customer Solutions at WaveAccess.
Also read: Industry 4.0 – How do you actually implement it?
#Machine #Learning #Industry #Opportunities #Technology