PREDICTIVE MAINTENANCE

predicitve-maintenance

The promise of predictive maintenance is to avoid unexpected system failures and to enable convenient maintenance planning. Predictive maintenance makes predictions about periods of time in which a machine, plant or other system is likely to fail. It is also possible to describe the component that will cause the failure. This is useful for the maintenance staff to bring the right equipment and spare parts. Other advantages include longer machine/plant life, increased plant safety and fewer accidents. Cost savings are generated by:

  • Reduction of maintenance frequency
  • Reduction of lost production time in the event of failures
  • Reduction of lost production time due to unnecessary planned maintenance
  • Reducing the cost of spare parts storage and deliveries
  • Very high ROI (investment returns) of predictive maintenance projects when many systems (machines, plants) of similar type are handled.

Predictive maintenance uses existing as well as newly installed sensors to measure physical quantities that are digitized and collected locally. They are then sent to a local (edge) computer or to a datacenter where predictive maintenance algorithms are executed.

A key element in this process is the Internet of Things (IoT). IoT enables different assets and systems to connect, collaborate, share, analyze and trade data. The transmission of data to data centers is usually carried out wirelessly, e.g. with Narrow-Band IoT (NB-IoT).

Predictive maintenance does not have to be carried out in real time, because the prediction of a failure, which would take place seconds or a few minutes after the corresponding sensor measurements, usually has no great practical value except to turn off the system in order to avoid major damage.

Examples of the use of predictive maintenance are:

  • Elevators and industrial printers
  • Wind turbines where downtime can be almost completely avoided
  • Manufacturing plants
  • Vehicles where extensive data collection of many sensors helps prevent expensive repairs or failures.

DETECTION OF DEFECTIVE PARTS

qualitaetspruefung

AI models, such as Convolutional Neural Networks (CNNs), can detect patterns, in some cases better (higher detection rate) than humans.  AI can identify objects that are different from others where the latter have the same or very similar characteristics. In this case, the model (a classification model) has the task of detecting defective parts.

If such a trained model reduces the proportion of unrecognized defective objects, from, for example, 3%  to 1% (AI), there are clear advantages for maintenance, logistics and production process, which lead to cost reduction and increase customer satisfaction.

MINIMIZATION OF ENERGY CONSUMPTION WHEN COOLING BUILDINGS

We will offer solutions that reduce the consumption of electrical energy for cooling buildings. We will start from mathematical foundations of DeepMind’s solutions, which led to a reduction in energy consumption of up to 40% in Google data centers.

The starting point is the following:

  • Rooms to be cooled according to a predetermined temperature distribution
  • Cooling systems are available, the power of which can be controlled via “setting parameters”
  • Sensors are installed that measure different physical quantities in and outside of the building, e.g. outside temperature and wind direction.

 

​Modell and training

Picture top left. A model, in this case a DNN, is created and trained in such a way that its predictions E are good estimates for the amount of electrical energy Et, (t stands for target) that is actually consumed for cooling the rooms during a time interval, e.g. five minutes. The input data of the model are Ez and the sensor values (Is) and cooling equipment parameter values ,(Ie).

Inference (operation)

Picture at the top right. The outputs E1, E2,…. EM of the trained model are each one estimate of the energy that would actually be consumed under the assumptions:

  • Actual values of the sensor measures (Is) and
  • M different sets of the cooling equipment parameters (Ie,1, Ie,2,…. Ie,M).

Essentially, the model is running trials in order to find the cooling equipment parameter values (Ie,R) which lead to the lowest energy estimate. Formally, Ie,R, equivalently R, is chosen such that

ER < En    for all  n ≠ R.

The calculation of the predictions En can be done sequentially on the same hardware or paralleled.

MINIMIZATION OF ENERGY CONSUMPTION WHEN COOLING BUILDINGS

We will offer solutions that reduce the consumption of electrical energy for cooling buildings. We will start from mathematical foundations of DeepMind’s solutions, which led to a reduction in energy consumption of up to 40% in Google data centers.

The starting point is the following:

  • Rooms to be cooled according to a predetermined temperature distribution
  • Cooling systems are available, the power of which can be controlled via “setting parameters”
  • Sensors are installed that measure different physical quantities in and outside of the building, e.g. outside temperature and wind direction.

 

​Modell and training

Picture top left. A model, in this case a DNN, is created and trained in such a way that its predictions E are good estimates for the amount of electrical energy Et, (t stands for target) that is actually consumed for cooling the rooms during a time interval, e.g. five minutes. The input data of the model are Ez and the sensor values (Is) and cooling equipment parameter values ,(Ie).

Inference (operation)

Picture at the top right. The outputs E1, E2,…. EM of the trained model are each one estimate of the energy that would actually be consumed under the assumptions:

  • Actual values of the sensor measures (Is) and
  • M different sets of the cooling equipment parameters (Ie,1, Ie,2,…. Ie,M).

Essentially, the model is running trials in order to find the cooling equipment parameter values (Ie,R) which lead to the lowest energy estimate. Formally, Ie,R, equivalently R, is chosen such that

ER < En    for all  n ≠ R.

The calculation of the predictions En can be done sequentially on the same hardware or paralleled.

Training

In the three applications described, the models must be trained before use.

Large amounts of data are required for training in these cases. It is a great advantage when historical data is available, the more the better. If such data does not exist, collecting data by measuring machines with suitable sensors could take a very long time