Predictive Maintenance (PrM-Predictive Maintenance) can best be described as a series of activities that are undertaken as a result of changes in the physical condition of an asset to prevent the risk of failure.
Predictive maintenance itself can be divided into methods (sensors) to detect starting failure (e.g. wear) and algorithms based on statistical data (sensors) for predicting failure.
The use of the sensors as an indicator for determining the condition and therefore the right moment of intervention (CBM-Condition Based Maintenance) can be supported continuously or periodically than for maintenance decisions. The condition of components is then automatically assessed via algorithms and patterns. Failure can then be recognized early and corrective measures can then be planned.
Unplanned outages (as the promise is) can be avoided and both staff and resources can be deployed more effectively. This is the first step in the implementation of a PrM-data acquisition process. Sensors generate information (vibrations, temperature, capacity) about the condition of the components. Different sensors, ultrasonic sensors, vibration sensors and acoustic emission sensors, are designed to collect different data.
Two steps are needed to handle the signals from sensors. One is signal processing, which improves the signal characteristics and quality. The techniques in signal processing include filtering, amplification, data compression, validation and noise reduction. The next step is to extract those processed signals precisely those data that are characteristic and appropriate for a starting failure.
The next step is to make decisions about maintenance actions based on that date. The decisions can be supported with a: physical model, statistical model, data or a combination (hybrid) model. These models are used to evaluate the current condition. The most commonly used method is the comparison with historical data (data driven) to identify and evaluate the failure. When historical data is missing, model-based techniques are used to evaluate the condition. These techniques can only be used after an analysis and knowledge of the asset. After all, a model of existing reality (current condition and condition progress) must first be made.
Techniques for supporting maintenance decisions can be divided into: diagnosis and prognosis. Diagnosis focuses on the detection and identification of failures and the moment when they occur. The prognosis focuses on predicting failures before they occur. In practice, however, it is not easy to apply prognosis and diagnosis due to the lack of procedures, data and knowledge required to create the models. Usually only experimental data is used to prepare a model or algorithm. What is still missing are techniques such as artificial intelligence, neural networks, fuzzy logic systems. Compared to diagnostics, the number of prognostic possibilities is much smaller.