Predictive Maintenance (PM) Optimal Servicing by Machine Learning Prediction

A technical installation is then reliably available if it is repaired in a timely manner, i.e. maintenance and repair are done directly. This will only succeed if the company can access appropriate resources. In short: the appropriate skilled workers, spare parts and logistics etc. are present.

If the company continues to use an asset with an unrecognized component maintenance requirement and does not repair it promptly, troubleshooting often becomes even more expensive. It may then also other parts of the system affected and the maintenance takes longer accordingly.

In order to be able to react as quickly and correctly as possible, a reliable detection and prediction of damage events would be ideal. In this way, the necessary means for economically optimal maintenance can be kept available as needed. The aim is therefore a maintenance strategy that recognizes possible errors before they happen and thus enables the planning of optimal maintenance.

Reliable prediction of future events is an integral part of any Predictive Maintenance (PM) system. An important key lies in the patterns of past events. The recognition of complex patterns as well as their own momentum and trends are the playing field of the PM system developed by us.


DIN 3105 differentiates between three types of maintenance:

  1. trouble-dependent
  2. time-dependent
  3. state-dependent

Predictive Maintenance is an extension of condition-based maintenance that can predict when the equipment will fail. The main goal is to maximize the use of the components while eliminating any errors or failures before they happen.

Models and Techniques

Depending on the objective, the models then make different statements. Here are some examples:

  • With what probability does a specific event occur in the next operating period of a certain time interval, in a certain operating mode of the technical system?
  • Which are decisive influencing factors for certain event types? How can they be ranked?
  • In which state is the system currently located? What are the transition probabilities from given state in a certain period of time to a certain other state?
  • How can multivariate anomalies be detected and interpreted?

We use machine learning techniques such as:

  • Cluster Analysis: Subspace Clustering, Data Streaming Clustering
  • Feature Selection/Reduction of Dimensions: PCA, ICA, Auto Encoders, Mutual Information, First order utility
  • Event Analysis: Markov Chains, CTMC, Non-homogeneous Poisson processes, Cox Regression, Bayesian Model Averaging, MCMC
Scheme Predictive Maintenance
© Fraunhofer ITWM

Scheme Predictive Maintenance

Benefits of Predictive Maintenance

Predictive maintenance can be introduced quickly, especially where a large amount of data has already been collected. However, the subsequent upgrade of monitoring systems may be worthwhile. In general, PM has the following advantages:

Immediate benefits:

  • the chance to predict and avoid faults, errors and failures before they occur
  • less costly plant shutdowns, optimal exploitation of the potential of the plants
  • maintenance is planned more cost-efficiently and in advance

Indirect benefits that result:

  • higher utilization of the profit potential
  • increase of productivity
  • lowering storage costs
  • improvement of product quality
  • preservation of resources
  • increased competitiveness
  • Good maintenance also increases customer satisfaction and
  • offers the possibility to develop new service concepts.