Anomaly detection and other practical considerations for estimating acoustical metrics from time series data

Abstract

The scope of environmental acoustic monitoring and rate of data collection are growing rapidly. These increases in the quantity of information have elevated the necessity of detecting anomalous data and the difficulty of doing so. Analysis of contaminated data leads to incorrect results, including biased parameter estimation and flawed model selection. Censoring of data requires strong justification, as the loss of information can lead to these same problems. For acoustics, this issue is compounded by the widespread use of time-average sound levels (Leq), which are especially sensitive to anomalous measurements. This talk will discuss applications of anomaly detection to acoustic pressure time series data for the calculation of long-term metrics. There is no single universally applicable or generic approach to data cleaning and preparation, which often consumes a disproportionate amount of time and effort relative to interpretation of results. Examples drawn from terrestrial and aquatic acoustical monitoring data sets of varying resolution and associated metadata will provide context for issues that compromise data quality, anomaly detection methods, and mitigation efforts.