Developing Condition-Based Maintenance

Scientists combine equipment health monitoring, detection, and forecasting to keep systems operating.

Like any manufacturing equipment, semiconductor fabrication systems have a finite lifetime. Technicians normally perform maintenance on these hardware systems according to preset schedules and regardless of actual need, which results in unnecessary equipment downtime and needless costs incurred as a result of lost production time and additional maintenance labor. AFRL scientists teamed with researchers from the University of New Mexico (UNM) to examine the feasibility of establishing prognostics for such expensive and valuable machinery and to devise a mechanism for scheduling equipment maintenance based on needs rather than calendar cycles. This so-called condition-based maintenance has the potential to increase equipment availability, improve productivity, enhance safety, and reduce expenses. The ultimate objective of the AFRL/UNM collaboration is to develop a data-driven prognostic system that provides advanced warning of failures, faults, and other error events that occur in complex systems.

Depiction of the relationship between an intelligent, system-level prognostic technique and the individual subsystem diagnostic tools which it communicates to correlate information

A functioning system, be it a machine tool or an aircraft, creates a dataset in n-dimensional space. This dataset contains certain recognizable characteristics, or signatures, with each signature denoting a particular system condition or event. Conversely, a malfunctioning system generates different datasets and creates different signatures; complex partitions in the n-dimensional dataset separate these signatures. Because a complex system can have a substantial number of unique conditions or events, the volume of generated information applicable to forecasting maintenance needs is well beyond the capacity of the engineer or machine operator to process. Scientists can detect the distinct signatures associated with the state of a system using methods such as bounds checking, statistical analysis, neural networks, fuzzy logic, data mining, classical expert systems, and hierarchical and hybrid systems. Each of these signature detection techniques is a good candidate for use in prognostic systems, and all are described in more detail in the following paragraphs. The AFRL/UNM scientific team chose to integrate various soft computing techniques with these established methods to discern the subtle clues contained in this massive network of information.

Operators routinely employ the original diagnostic technique—bounds checking—as an effective prognostic tool. In this approach, instruments monitor operating parameters, such as temperature, pressure, voltage, current, and load, to ensure they fall within predetermined operating limits established by experience and prior statistical studies. Additionally, operators can apply several well-known statistical analysis techniques to effectively compare the system's current state to distinct feature vectors. Statistical analysis can also provide prognostic information when software routines employ it in conjunction with prior operational knowledge.

Neural networks model how neurons work, and researchers can use these networks to create linear, plain, or volume pattern recognition systems. The many types of neural networks include multilayer, polynomial, radial, self-organizing, and probabilistic. In addition, a number of neural networks are considered "gated experts," because they employ an algorithm that assigns the experts based on their best performance record. Fuzzy logic, which emulates the way humans make decisions, is yet another model that researchers can leverage for nontraditional pattern recognition and classification.

Along with their use of neural networks and fuzzy logic, researchers are developing data mining algorithms to predict system failures before they occur. Data mining algorithms can identify the signs of impending failure, eroding conditions, and reduced dimensionality required to make prognostic predictions. Intelligent mining tools extend these algorithms by including the capability to self-learn new circumstances associated with the health of repetitive systems and equipment. The addition of intelligent mining techniques therefore promises to increase the accuracy of current analysis and probability-based programs. Maintainers currently use both pattern recognition and data mining algorithms to identify existing equipment and component faults; capturing and analyzing the data necessary for predicting the onset of abnormalities and faults prior to equipment/component failure marks the next logical step.

Once other techniques detect a problem, maintenance personnel employ classical expert systems, which usually present corrective actions as text-based instructions such as, "if this happens, then do this." The AFRL/UNM team developed three different electronic automated experts based on auto-associative neural networks, Kohonen self-organizing maps, and radial-based clustering algorithms, respectively. Hierarchical and hybrid systems make the best possible use of available information (see figure on previous page). Combining these tools—from collection and detection to prediction— can provide advance warning of failures, faults, and other error events. Armed with this information, operators can plan for condition-based maintenance in real time. The AFRL/UNM team is currently building a library of advanced pattern recognition and data mining techniques to create a prognostic system for AFRL's chemical oxygeniodine laser component of the Airborne Laser program. The team is also in the process of applying these prognostic techniques to other lasers, aircraft structures, biomedical devices, and industrial fabrication equipment.

Mr. Victor Stone and Dr. Mo Jamshidi, of the Air Force Research Laboratory's Directed Energy Directorate, wrote this article. For more information, contact TECH CONNECT at (800) 203-6451 or place a request at http://www.afrl.af.mil/techconn_index.asp . Reference document DE-H-05-05.