Parallel reservoir computing based signal outlier detection and recovery method for structural health monitoring

Abstract

The presence of outliers in signals collected by structural health monitoring systems, caused by sensor failure, equipment malfunction, or transmission interruption, can lead to misjudgments of a structure's working status and damage degree. This study proposes a novel, fast, accurate, and automatic method which are capable of reconstructing signals according to adjacent channels, detecting outliers by amplifying and sorting reconstructing errors, and recovering normal values to the corresponding locations. A parallel reservoir computing-based reconstructor with a decomposition module which purifies frequency components of input for each sub-network is utilized for improved precision. In addition, the adopted local outlier factor algorithm simplifies outlier detection work as simplex threshold comparison. The proposed method is analyzed for its effectiveness in detecting various types of outliers, such as spikes, abnormal segments, external trends, shifting, and baseline drift, using an acceleration dataset from the Shanghai Tower.

Keywords

• Structural health monitoring
• Anomaly detection
• Outlier recovery
• Recurrent neuron network
• Reservoir computing
• Continuous wavelet transformation