Condition Based Maintenance (CBM) holds the promise of predicting machinery maintenance by using sensing technology, signal processing and software techniques. Condition-based maintenance allows to take planned corrective measures following sensing of asset performance degradation.
The first step in implementing CBM is selecting the suitable sensors. The next step is acquisition of data indicating the condition of the equipment. Various sensors like, ultrasonic sensors, vibration sensors, and acoustic sensors, have been designed to generate different data.
Generally, there are two steps to deal with the signals from sensors. The one is signal processing, which enhances the signal characteristics and quality. The techniques in signal processing include filtering, amplification, data compression, data validation, and de-noising. The other is extracts features from processed signals that are characteristic of an incipient failure or fault.
The last step is decision-making on taking maintenance actions. The models for decision-support could be divided into four categories: (1) physical model, (2) statistic model, (3) data-driven model, and (4) hybrid model. Because CBM mostly depends on signals and data reflecting the condition of equipment, data-driven model will be in a dominant place.
Sensors can report data on condition, but, ultimately, it is the system to which they are attached that provides the intelligence to interpret the data and take action. Based on a model of detoriation the asset can be repaired preventive, during the next scheduled maintenance intervention, as opposed to making repairs after an unanticipated breakdown.
If the historical data can be obtained easily, the data-driven model is used to compare and evaluate the fault and the condition. When none or only a small part of historical can be obtained, the hybrid techniques combine the data- driven techniques and model-based techniques to evaluate the condition. A semi-supervised learning method can be used to create a physical model based on new data. Techniques for maintenance decision-making can be divided into two main classes: diagnostics and prognostics. Diagnostics focuses on detection, isolation, and identification of faults when they occur. Prognostics attempts to predict faults or failures before they occur.
However, in practice, it is not easy to apply prognostic techniques due to the lack of training data and specific knowledge, which are required to create the models. New, more promising methods to create these models include artificial neural networks, fuzzy logic systems, fuzzy-neural networks, neural-fuzzy systems, evolutionary algorithms, and swarm intelligence. Compared to diagnostics, the application of prognostics is much smaller.