Maintenance costs are normally a major portion of the overall operating costs in most production facilities. Those costs are constituted from labor and material costs. Therefore the maintenance optimizing is a very important task that needs to be addressed in every production system.
Unlike in Run-to-Failure Management or Preventive Maintenance strategies, in Predictive Maintenance we continuously utilize automated analysis (Analytics) of large sensor datasets (Big Data) and are thereby able to identify operational problems before they actually occur.
1. Avoid unscheduled downtime
The studies show that the downtime caused by equipment or system failures can be reduced 40% to 60% within the first two years and up to 90% within five years.
2. Boost manpower utilization
By identifying the precise repair task, as well as the parts, tools and support needed to rectify the problem, predictive maintenance can dramatically increase the manpower utilization.
3. Reduce labor and material costs
Effective use of predictive maintenance will eliminate 35% to 60% of maintenance expenditures (labor and material).
4. Increase useful life
Making minor adjustments or repairs and thereby not permitting a minor deficiency from becoming a serious problem can extent the useful operating life of system assets by 33% to 60%.
Showcase: Predicting rainfall by applying Analytics on climate data
In this showcase we demonstrate how we can solve the task “recognition of correlations between atmospheric events” by applying our Analytics technology – which is normally used for the prediction of incidents and failures of systems and equipment. This is possible because the nature of the two tasks is basically identical and because the analytics methodology does not consider the semantics of the data being analyzed.
Although the main purpose of our climate sensor WolkSensor is the sensing and transferring of the climate data of closed spaces and refrigeration equipment, the observation that the air pressure in closed spaces is absolutely identical to the outdoor air pressure tempted us to use his recorded office air pressure data together with the publicly available rainfall data for a small experiment.
In particular, we used this experiment to put our hypothesis “Relevant drop of the air pressure (at least 1.0 hPa/hour) is most probably followed by the rainfall.” to the test. In the processed data we used the timestamp as the id attribute and the rainfall occurrence as the labelled attribute. To train our prediction model we used the data starting from the 1st till 28th of November 2015. After its training the prediction model has been applied to the two last days of November:
In the diagram above we presented with the green line the values of a rainfall probability multiplied with 10. As you can see on the right side of the diagram, our model predicted the rainfall with high probability (values range from 0.8 to 1.0) on both days, which corresponds to the historical rainfall data (presented with the red line at 12PM of each day) and affirms our hypothesis.
To make sure that our results are not accidentally correct, we applied the trained model to the first 20 days of December 2015 and in most cases confirmed the previous results:
One case, where our model failed, is 6th December. On that day there was no rain, but the prediction says the opposite. The reason for this is the fact that our model does not take all other relevant climate parameters (air temperature, humidity, wind direction and intensity) into account.