False Alarm-Improved Detection Capabilities of Multi-sensor-Based Monitoring of Vibrating Systems

Monitoring the State-of-Health of vibrating mechanical systems is useful but complex. Besides challenges associated with dynamical behaviors of the systems monitored, supervision tasks are complex with respect data acquisition, feature extraction, and/or statistical modeling for feature classification. Data acquisition strategy addresses sensor types, quantities, and locations. Feature extraction task details the selection and processing of features sensitive to a change/fault present and if required the development of statistical models for change/fault classification. In this contribution, the above-mentioned challenges associated with supervision are explained and detailed using an elastic mechanical structure applying the Probability of Detection method. Previously solved problems relating to simultaneously accessing all mentioned challenges are briefly repeated for understanding. This serves as a prelude to the newly developed data driven noise analysis and improved detection procedure. An experimental example using different real sensor types in combination with mechanical modifications of an elastic beam is presented. The adapted Probability of Detection method helps to determine a suitable feature, sensor type, and position for least change/fault detection. In this article a new data driven noise analysis approach is introduced to ensure optimal sensor-specific flaw size detection. Optimality in this context is related to the selection of the appropriate feature and threshold values for desired false alarms. The noise analysis permits the selection of a decision threshold (threshold beyond which change/fault is considered present in a signal) with the corresponding detectable flaw size and related false alarm rate. Selecting different sensors implies changing the signal distribution character and the decision threshold. This change results in different values and hence can be exploited to decide the optimal sensor. The implemented noise analysis allows a trade-off between flaw size detection and probability of false characterization of faults with 90% detection at 95% reliability level. The novel approach provides a graphical representation that illustrates the diagnostic capabilities of a sensor as its decision threshold is varied.