The ultimate maintenance goal, of PM is the development of a dynamical maintenance schedule based on algorithms fed with the actual condition. Such a PM policy uses prognostics to analyze data to assess its current condition and to detect and diagnose faults in the machine. Predictive maintenance (PrM) is based on educated estimates when the next failure is likely to occur. When a maintenance department is capable to predicting failures, maintenance can be planned in advance. It also has positive consequences for spare part inventory, can reduce downtime, and will at the end increase operational efficiency.
A PrM policy requires a method to assess the actual condition of the asset and detect incipient faults in a timely manner. It therefore requires use of both sensor technology and knowledge of the asset. Sensor are placed on the sources of faults. These sources must be observed to translate symptoms of incipient failure into analyses. Sources can be the core components of assets. The number, type and location of sensors, and their reliability and redundancies all affect costs and extensive processing of data. Data as temperature, pressure, voltage, noise, or vibration measurements are collected using dedicated sensors. Sensors derives metrics from predictable changes when the asset degrades. Algorithm perform fault detection and diagnosis by comparing new data against the established markers of faulty conditions.
Data preprocessing includes simple techniques such as outlier and missing value removal and mean value of the data over time. More advanced signal processing requires techniques such as Fourier transformations. Complex signal analysis is used to measure the frequency of the peak magnitude in a signal spectrum, or describing changes in the spectrum over time.
Designing an algorithm begins with data from multiple sensors and multiple machines running at different times and under different operating conditions. Extended knowledge of the asset help determines what preprocessing methods to use. Typical data sources are: normal operation, operating data under faulty conditions, system failures (run-to-failure). In many cases, failure data from machines are not available, or only a limited number of failure datasets exist because of regular maintenance being performed and the relative rarity of such incidents.
The existing (tacit and explicit) knowledge must be converted into mathematical models of the system and its possible faults, effects and causes from the insights of domain experts. Understanding system dynamics involves detailed knowledge of relationships among various known signals from the asset machinery, the asset operating conditions, and the maintenance history.
Modeling requires identifying appropriate condition indicators and training a model to interpret them. Some models analyze residual computation, state estimation, and parameter estimation. More complex model is based on fault-diagnosis approach, a classifier is trained to compare the current value of one or more condition indicators to values associated with fault states, and returns the likelihood that one or another fault state is present.
Prognostics algorithm is forecasting when a failure will happen based on the current and past state of the machine. A prognostics algorithm typically estimates the machine’s remaining useful life (RUL) or time-to-failure. Prognostics can use modeling, machine learning, or a combination of both.