EXTRACTING METHOD OF CHANNEL-FREQUENCY FEATURES IN DAS SENSORS

Abstract

A method for calculating channel-frequency feature for detecting multi channel (propagated spatially) activities in DAS sensor data is provided. This method can be generalized for linearly spaced sensor arrays. In the method, spectrograms of different channels are generated and frequency features are calculated for different time windows. In channel frequency image, frequency features are concatenated in spatial domain, so that horizontal axis represents the spatial dimension.

Claims

1. A method for extracting a channel-frequency feature for linearly spaced sensors for detecting multi channel activities, comprising steps of: calculating spectrograms of linearly spaced channels for different time windows to generate three dimensional data as a channel, a time, a frequency, concatenating frequency bins in a spatial dimension to extract a channel-frequency image.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0009] FIGS. 1A and 1B show spectrogram extracted from the seismic signals of a train in different channels.

[0010] FIG. 2 shows flow chart of channel-frequency feature extraction.

[0011] FIG. 3 shows graph of channel-frequency feature of a train.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] In railway applications, DAS sensors are becoming popular due to wide distance of sensing. Continuous monitoring of railway and train conditions and detecting anomalies improves the safety in railway transportation. Estimation of train location, speed, number of wagons and detection of flat wheels, broken rails constitute main application areas of DAS sensor in railways. For these tasks, DAS sensors offer great potential with high performance.

[0013] In order to exploit the advantage of wide range coverage of DAS sensors, sophisticated algorithms and fast computation techniques are required. Traditionally, signal processing techniques are applied on seismic signals but recently machine learning approaches outperforms the other techniques. Machine learning approaches also allow better feature extraction in spatial dimension.

[0014] In audio applications, spectrogram is a widely used representation of time series data where one dimensional time domain signals are converted to 2-dimensional time-frequency features. In a spectrogram, FFT magnitudes of short time windows are concatenated horizontally by sliding the window. In FIGS. 1A and 1B, we see spectrograms extracted from the seismic signals of a train. In DAS sensors instead of single channels, we have time series data from multiple channels.

[0015] Spectrograms represent data in time frequency domain. In DAS sensors, there exists one more dimension, spatial dimension, due to multiple channels. Calculating the spectrograms of different channels generates three dimensional data (Channel, Time, Frequency).

[0016] In this work, a channel frequency feature extraction method for estimating multi-channel activities in DAS sensor data is proposed. This method utilizes the linearly spaced channel features of DAS sensor and the method is not applicable for other sensor geometries. This interpretation of data enables the adaptation of commonly used advanced image processing techniques. Processing channel-frequency features in different time frames is similar to video processing.

[0017] In spectrogram images, frequency features are calculated for different time windows. In channel-frequency image, frequency features are concatenated in spatial domain, so that horizontal axis represents the spatial dimension.

[0018] In proposed algorithm, FFT of each channel is calculated for T samples. If magnitudes of frequency content are calculated, F=T/2+1 frequencies extracted for each channel. Then frequency bins are concatenated in spatial dimension so that C×F channel-frequency image is extracted.

[0019] If there is an activity that spans multiple channels, it has a specific signature in channel-frequency image of said activity. In FIG. 3, channel-frequency feature of a train which has a length of 200 m can be seen. The proposed channel-frequency feature has a better representation of raw data and allows implementation of advanced image processing techniques. It can be applicable for linearly spaced sensor geometries and provides a rich feature space due to the fusion of data obtained by multiple channels.