TRANSACTIONS OF THE INSTITUTE OF MEASUREMENT AND CONTROL 2018, VOL. XX(X) XXXX 1 Comparison of delayless digital filtering algorithms and their application to multi-sensor signal processing Anna Swider*, Eilif Pedersen Abstract In the phase of industry digitalisation, data are collected from many sensors and signal processing techniques play a crucial role. Data preprocessing is a fundamental step in the analysis of measurements, and a first step before applying machine learning. To reduce the influence of distortions from signals, selective digital filtering is applied to minimise or remove unwanted components. Standard software and hardware digital filtering algorithms introduce a delay, which has to be compensated for in order to avoid destroying signal associations. The delay from filtering becomes more crucial while analysis measurement from multiple sensors, therefore in this paper we provide an overview and comparison of existing digital filtering methods with application based on real-life marine examples. Additionally, design of special purpose filters is a complex process and for preprocessing data from many sources, application of digital filtering in the time domain can have a high numerical cost. For this reason we describe Discrete Fourier Transformation digital filtering as a tool for efficient sensor data preprocessing, which does not introduce a time delay and has low numerical cost. The Discrete Fourier Transformation digital filtering has a simpler implementation and does not require expert-level filter design knowledge, what is beneficial for practitioners from various disciplines. Finally, we exemplify and show the application of the methods on real signals from marine systems. Index Terms digital filtering, DFT, preprocessing, delay, delayless preprocessing, synchronisation, big data, IoT, sensors. I. I NTRODUCTION D ATA analysis has a vital role in many different industries, e.g.: medicine (gens analysis), economics (stock exchange) and in marine engineering (the ship industry enters the Shipping 4.0 phase Rødseth et al. (2016)). There is a need for collecting and processing huge quantities of measurements as time series data (signals) which comes from many sources, often referred to as big data DNVGL (2017). The marine industry is now entering the challenging phase of smarter shipping, including on-board monitoring systems, and advisory tools. Modern vessels will be equipped with various on-line data collection and advanced monitoring systems. The on-board measurements from sensors of many installations play a crucial role, and their availability expands the functionality of marine products. The aim of data analysis in the marine application is developing on-shore and on-board advisory tools using prediction of propulsion power or ship performance monitoring, as well as enhancing knowledge about specific systems and components, and the relationship between systems Swider & Pedersen (2017). Machine learning algorithms and statistical modelling become widely used tools in equipment monitoring and advisory systems. However they are sensitive to data quality, and in particular to relationships between subsystems retained as correlations in the data. A fundamental step in the analysis of measurements, and before applying machine learning is data preprocessing Garc´ ıa et al. (2016), Taleb et al. (2015). Unfortunately, in the literature from different industries e.g. Kuhn & Johnson (2013), the importance of the quality of the time series is limited. Because measurements play an important role in marine applications, proper data preprocessing and improvement of their quality is critical to ensure correct interpretation on board the vessel or during off-line analysis. A major source of the disturbances and distortions in measurements is the Data Acquisition System (DAS). The role of the DAS is the collection of measurements of the desired variables, transmission and conversion of the recorded signals to digitized form Bendat & Piersol (2010). Among the most common disturbances and distortions are Vaseghi (2009), Bendat & Piersol (2010): white noise, poor calibration and digitization effects (ADC quantization or aliasing) Randall (2011). Data cleaning is a necessary stage, where distortions and disturbances are eliminated as far as possible. It plays an important role during data analysis Frnay & Verleysen (2014). A typical data analysis scheme is depicted in Figure 1. It contains some major steps like data collection, data cleaning and feature extraction before the machine learning or statistical modelling can be applied. In this paper we focus on the data cleaning stage. Processing of data from several marine systems creates many challenges. One of them is lack of data synchronisation, which can be introduced by specific systems setups, different time intervals or preprocessing. In the literature from data analytics, data Anna Swider*, correspondence author: anna.swider@rolls-royce.com, is with Rolls-Royce Marine AS and Norwegian University of Science and Technology (NTNU), Department of Marine Technology Eilif Pedersen is with Norwegian University of Science and Technology (NTNU), Department of Marine Technology Manuscript received February 14, 2018. Manuscript revised June 29, 2018.