Data Science

KI in der Produktion: Ein Sprungbrett zu Kosten- und Prozesseffizienz

BMW macht´s vor: In der Automobilproduktion setzt der Hersteller auf Künstliche Intelligenz, um eine präzise Qualitätssicherung für tausende täglich gebaute Fahrzeuge umzusetzen. Was früher Mitarbeiter mit Messtechnik erforderlich gemacht hätte, kann jetzt aus einer Kombination von hochauflösenden Kameras und Künstlicher Intelligenz erledigt werden.

Dies ist nur ein Beispiel dafür, wie KI in produzierenden Unternehmen die Qualitätssicherung sehr effizient umsetzbar macht. Die intelligenten Algorithmen erlauben Ausschuss und Nacharbeit stark zu reduzieren. Mit verschiedenen KI-Methodiken können diese fehlerhaften Teile oder Prozesse identifiziert werden und den verantwortlichen Ursachen so – auch bereits vor der Entstehung – entgegengewirkt werden.

Qualitätssicherung mit KI ist nur einer von vielen Use Cases

Doch die Qualitätssicherung ist nicht der einzige Anwendungsfall für den Einsatz von KI in der Produktion. Weitab von Buzzwords rund um Industrie 4.0 sorgen die selbstlernenden Algorithmen bereits in der Praxis für mehr Effizienz und Kosteneinsparungen – einem nicht zu vernachlässigenden Wettbewerbsvorteil, gerade in unsteten Zeiten mit fragilen Lieferketten. Das neue Schlüsselwort heißt Resilienz.

So hilft sie zum Beispiel bereits im Maschinen- und Anlagenbau Planungsdaten zu verbessern. Die Produktionsplanung dort ist durch die hohe Komplexität, bestehend aus Zukaufteilen und eigener Herstellung, meist eine Herausforderung. Der Einsatz von maschinellem Lernen verbessert die Planungsdaten wie unter anderem Wiederbeschaffungszeiten und hilft so, die Liefertreue und demzufolge die Kundenzufriedenheit zu steigern. Häufig stimmen die Planungsdaten nicht mit der Realität über und so kann es sein, dass beispielsweise mit einer bestätigten Lieferzeit von 10 Werktagen gerechnet und geplant wird, obwohl die Ist-Lieferzeit hinterher doch 20 Werktage beträgt. Dies führt zu starken Verzögerungen in der Produktion und somit zu einer verspäteten Lieferung. Die KI bereinigt diese Daten und präzisiert so die Planung. Zusätzlich wird die Stammdatenqualität langfristig verbessert.

Und: die für den Einsatz von KI benötigten Daten liegen meist bereits in verschiedenen Systemen in Hülle und Fülle vor.

Studie belegt: KI schafft realen Mehrwert in Unternehmen

Dass KI und deren Anwendung für produzierende Unternehmen eine zunehmende Rolle spielen, zeigt auch die Studie „Machine Learning 2021“ vom INFORM DataLab in Zusammenarbeit mit Microsoft und Lufthansa Industry Solutions: Bereits rund 65 Prozent der befragten Unternehmen haben bereits Machine Learning (ML) im Einsatz oder sind dabei, entsprechende Lösungen einzuführen. Die Haupteinsatzfelder sehen die Befragten klar in der IT (76 Prozent) und in Produktionsumgebungen (57 Prozent).

Die größten Hürden bei KI-Projekten sind laut der Studie Mangel an Know-how und Spezialisten. Externe Dienstleister und Berater sind somit wichtiger denn je. Doch die Ergebnisse zeigen auch, dass  KI-Projekte reale Mehrwerte für die Unternehmen schaffen. Spätestens nach drei Monaten zahlte sich der Einsatz von maschinellem Lernen für mehr als 60 Prozent der Unternehmen bereits aus. Dies zeigte sich vor allem in höherer Produktivität und geringeren Kosten.

Fazit

KI ist schon lange keine Spielerei mehr. Für produzierende Unternehmen schafft sie bereits in realen Anwendungen verschiedenen Mehrwerte. Ob verbesserte Termintreue, maximale Datenqualität, präzise Planungsdaten oder gesicherte Qualität – die Erfolge sind vielfältig, genau wie die Anwendungsfälle.

Haben Sie bereits einen Anwendungsfall im Sinn? Melden Sie sich bei uns!

Read More
Straßenschild mit der Aufschrift "Wrong Way"
Data Science, Use Case, Use Case - EN

Highway infrastructure is optimized by AI

Challenge

In order to ensure effective control of the traffic flow, rapid detection and reaction to disruptions is essential. In some cases, outdated and inflexible algorithms are in use that no longer meet modern requirements. A large European highway operator wants to design and evaluate a completely new concept for incident detection based on the existing sensor technology.

Solution

Using the raw sensor data, AI-supported systems were trained to detect failures and malfunctions of individual measuring points. Based on this, a holistic model of a section of road continuously monitors the traffic flow and makes signage decisions independently and based on past experience. Particular emphasis was placed on flexibility in the face of structural changes and continuously changing traffic behavior.

Added value

  • Concept for the renewal of the current system

  • More efficient congestion avoidance through early detection of traffic incidents

  • Robustness against single sensor failures

  • Alleviation of the maintenance situation

More information
Read More
Geschäftsleute bei einem Meeting
Data Science, Use Case, Use Case - EN

Natural language processing supports investment bankers

Challenge

At a major German bank, thousands of final reports, each with hundreds of pages, were to be examined for risks / anomalies. The goal was to optimize the portfolio and provide a better basis for investment decisions.

Solution

Using a natural language processing framework that performs feature engineering based on financial statement text, our Data Scientists were able to perform anomaly detection and supervised machine learning using historical company information such as sales in subsequent years or bankruptcies.

Added value

  • With a probability of 70%, the bankruptcy of a company within the next 5 years can be predicted

  • Identification of conspicuous text passages from current annual reports, as an additional basis for decision-making for investment bankers

Highlights

  • Thousands of annual reports, each with hundreds of pages, were examined using natural language processing

More information
Read More
Auto wird mit einem Tuch poliert
Data Science, Use Case, Use Case - EN

Machine learning gives indication of soiling and damage in new vehicle logistics

Challenge

One of the largest German car manufacturers was faced with the problem that damage regularly occurs during the transport of new vehicles. These must be detected as quickly as possible and repaired before delivery. Current prevention measures are not appropriate, especially with regard to sustainability, and are therefore to be abolished.

Solution

Throughout the logistics chain, drones regularly fly over the vehicles to document them with high-resolution images. With the help of machine learning methods, these images are examined for damage and contamination and appropriate measures are initiated.

Added value

  • Resilient detection of damage and contamination with over 90%

  • Reduction of repainting

Highlights

  • GPU accelerated processing of 4K video footage

  • Autonomous drone flights based on GPS waypoints

  • State-of-the-art object recognition

  • Vehicle identification via QR code as a by-product

More information
Read More
zwei Personen besprechen etwas am PC
Data Science, Use Case, Use Case, Use Case - EN

Machine learning supports fraud prevention in the telecommunications industry

Challenge

A top German provider of telecommunications contracts and hardware regularly has to contend with unpaid invoices from its customers for contracts involving hardware, as this causes severe financial damage to the company. The goal and challenge was to predict these fraud cases as best as possible and prevent them accordingly.

Solution

With the help of a supervised machine learning pipeline including customized feature engineering, it is possible to predict whether future invoices can be paid or not based on various customer information in combination with payment history.

Added value

  • 90% of all unpaid contracts are reliably detected.

Highlights

  • Data from six different online stores from one year as well as other data sources, including device, product and customer information such as past purchases.

More information
Read More
Hand stoppt Domino-Effekt
Data Science, Use Case, Use Case - EN

Speech recognition helps insurance company detect fraud

Challenge

Insurance companies must prioritize when investigating claims because not all submitted cases can be verified. The damage report of the reporting party is a valuable resource for the assessment of the case, and the assessment usually requires many years of experience. A large German insurance company faced the challenge of needing to process claim reports more efficiently.

Solution

To make the processing of claims more efficient, a voice-based tool was developed to assist claims handlers. For this purpose, a large volume of claims reports was combined with the assessments of experienced claims handlers and the outputs of investigations looked into. Using this data, a Natural Language Processing model was trained to detect fraud attempts with significant precision. The damage reports marked in this way are processed with special attention in the further process.

Added value

  • 70% of fraud cases were identified in the test set

Highlights

  • Data from damage reports from over 15 years

  • Extensive validation against retained data

  • Tuning of the classification model according to the wishes of the claims handlers

More information
Read More
Flugzeug im Steigflug
Data Science, Use Case - EN

Process mining improves plan adherence

Challenge

In the past, a full-service partner for services at airports was often unable to meet plans. At the same time, there was no transparency and information about where delays, changes and deviations from plans, and thus costs, occurred.

Solution

Our team of experienced data scientists were able to troubleshoot numerous data quality issues using a process mining framework, distinguish between different perspectives such as per aircraft or service, among others, and then calculate process models. This involved using two years of flight and process data from a major international airport to develop a new algorithm to merge similar activities automatically or manually.

Added value

  • Process models with which bottlenecks and bottleneck resources become directly apparent and process complications can be identified

  • Basis for predictions about, among other things, process duration and the next process step

More information
Read More
Frau, die im Kundenservice arbeitet
Data Science, Use Case, Use Case - EN

Intelligent audio recognition leads to higher customer satisfaction

Challenge

A large German health insurance company found that it was only able to ensure customer satisfaction for a very small proportion of its call center calls - and that only through costly callbacks. There was no overview of how satisfactory the calls were for the individual customer.

Solution

A framework was developed that does not analyze the spoken words, but the properties of the audio tracks (e.g. spectrum, reaction times) - so-called prosody. This should make it possible to understand the mood and potential annoyance of a customer.

Added value

  • Very good detection of "escalated" calls

  • Decision-making basis for further customer satisfaction measures

  • Very good differentiation between positive and negative calls

Highlights

  • Consideration of high data protection regulations since the content of the calls must not be recorded

  • Analysis of hundreds of thousands of calls

  • Validation of the framework by means of past, manually collected satisfaction analyses

More information
Read More
Fabrik Maschinen- und Anlagebau
Data Science, Use Case, Use Case - EN

Artificial intelligence increases adherence to delivery dates of manufacturing companies

Challenge

Machine and plant manufacturers are familiar with this problem: The very complex planning of purchased parts and in-house production of a multi-variant production is already a difficult challenge without inaccurate planning values.  Also, the data often does not match reality; production planning is then calculated with 10 instead of 20 days, for example. The result of this is very inaccurate delivery dates and frequent deadline postponements. The manufacturing industry has the advantage that a lot of data is already available in several systems. The disadvantage is that the data collected is usually far removed from realistic values. The consequence is very imprecise planning, which makes it difficult to give customers precise delivery dates.

Solution

The use of machine learning helps to achieve strong improvements, for example, to determine replenishment times more reliably and precisely. Inaccurate data can be eliminated through data cleansing using machine learning and master data quality can be increased in the long term. Historical SAP data such as supplier, materials or purchase orders help to improve replenishment lead times by up to 42 percent.

The precise planning values can then be integrated, among other things, into mobile, dynamic store floor management, in which key productivity figures can be viewed at every machine via a dashboard directly on the shop floor. In doing this, production can also be made digital and paperless.

Machine learning improves planning values in mechanical and plant engineering.

Added value

  • 98% Operational Excellence

  • 42% better data

  • 90% more precise planning

  • 95% faster response time

Highlights

  • 159 monitored sensors

  • 6 years training data

More information
Read More
Hände eines Roboters und eines Menschen, die sich auf dem Hintergrund einer großen Datennetzverbindung berühren
Data Science, Use Case, Use Case - EN

Precise planning data for mechanical engineering thanks to machine learning

Challenge

A German mid-volume manufacturer of tooling systems with around 400 employees faced the challenge that planning data for production as well as upstream and downstream areas were inaccurate and core data was incorrectly maintained. As a result, planning stability was lacking and deadlines could often not be met.

Solution

A regression/machine learning algorithm has been developed to accurately predict operation durations based on operation and PDC information.

Added value

  • An average of 30% more accurate estimation of operation durations.

  • Rapid identification of data quality and process problems in production data collection.

Highlights

  • Algorithm learns from over 1,000,000 PDC bookings and hundreds of thousands of completed operations.

  • In addition, it takes into account dozens of relevant input variables such as current resources, material or core data.

More information
Read More