AI-based predictive maintenance for manufacturing plants

Prof. Dr. habil. Alfred Schöttl
Department of Electrical Engineering and Information Technology

Background
The fourth industrial revolution aims to digitalize production and make it smarter. One of the core components of Industry 4.0 are predictive maintenance systems.

By continuously monitoring characteristic values or plant components during operation, it is possible to draw conclusions about their condition. Various internal and external sensors can be used to record data from the production plant. Important measured variables are, for example, currents, temperatures, accelerations or occurring forces.

A consortium consisting of the Fraunhofer Institute for Microsystems and Solid State Technologies (EMFT), Procon IT GmbH, Mühlbauer GmbH and Munich University of Applied Sciences is working on this project, funded by the VDI/VDE-IT. The project officially started in October 2021 and is scheduled to run for three years.

Goal
The goal of the project is to use a limited number of sensors to create a reliable predictive maintenance system that can detect anomalous behavior of a production plant. In addition to the current operating status, it is also possible to draw conclusions about the remaining lifetime of components. The sensors to be used can include accelerometers, thermal elements as well as imaging sensors.

Methods
Existing methods are tailored to specific applications. Within the scope of this project, different methods will be used to solve the problem. Besides classical methods like clustering approaches or statistical models, supervised and unsupervised training algorithms based on deep neural networks will be used. Special attention will be paid to the last mentioned, since these networks have a high generalization capability and thus seem to be well suited for the task. Especially generative models are hardly used in this area yet, but their results for simplified problems are promising.

In order to find the optimal solution, different approaches have to be implemented and evaluated before the predictive maintenance system is applied to production lines.

A prediction component based on a stochastic reliability model will help determine optimal maintenance times.

Challenges
In real-world scenarios, there are a number of challenges in automatically detecting anomalies. Firstly, it cannot be guaranteed that processes within a production line run independently of each other. In addition, data comes from a wide variety of sources with many different value domains. They are multimodal. Another difficulty is the processing of this data due to the strongly varying sampling rates. Depending on the sensor type, sample rates can range from a few hertz to several kilohertz.

Anomalies can occur in different forms: Events occurring punctually in time series may be due to a failure of a component. If sensor values are observed over a longer period of time, a changing average value of a signal may indicate degradation. In addition, unknown, i.e., never before observed, misbehaviors may occur and must be detected as anomalies.

A general problem with anomaly detection datasets is the strong imbalance of the data, making conventional classification algorithms inappropriate. It is not uncommon for public datasets to consist of far more than 90% data describing the nominal state of a component or an entire production facility, and only a few examples are present that reveal anomalous behavior.

Recording a dataset in a production environment is time consuming and expensive. For this reason, a procedure model for model and measurement selection is also being designed as part of the project.

The prediction of the remaining lifetime is another challenge. For this purpose, existing model-based methods will be used and adapted to the application.

Outlook
A novel, generative model for anomaly detection in time series has already been designed and validated on the basis of publicly available reference data sets from industrial environments.

The next steps suggest an adjustment of the network to real data of the production plant.

Running duration:
01.10.2021 - 30.09.2024

Funded by:
Bavarian Ministry of Economic Affairs, Regional Development and Energy

Projectpartner:

• Mühlbauer GmbH & Co. KG
• Fraunhofer Gesellschaft e. V.
• Procon IT GmbH