Research On Water Quality Anomaly Event Detection Methods For Urban Water Supply System Based On Empirical Mode Decomposition

The safety of water quality is such a vital issue that is beneficial to national economy and people’s livelihood and has called so much concern and attention to countries all over the world. Water contaminants event often leads to serious consequences, so to establish mechanisms to prevent water contamination and to build water pollution early warning monitoring system is quite essential. Water quality anomaly detection algorithm is the core of early warning monitoring system, including some common used anomaly detection methods based on statistical analysis, clustering and artificial intelligence. However, water quality indicator time series are non-stationary itself and the timescale of water quality event can be uncertain too, thus the performance of some regular algorithms to multi-scale event is poor, with a low detection rate and a high false alarm rate, especially for the events that are relatively long and smooth.In order to further explore the fluctuation characteristics of components in water quality indicators time series that come from different sources, to combine multi-scale and time-frequency variation characteristics of water event and to improve the detection capability and utilization of information for water quality anomaly detection algorithm, this paper proposed several Empirical Mode Decomposition (EMD) based water quality anomaly detection methods. The theory study of EMD is briefly introduced as well as its application fields, principal defects and solutions. At last, the effectiveness of the proposed methods for water quality anomaly detection is illustrated by contrast with some regular algorithms through simulation analysis.The main work and innovation of this paper are as follows:1. For issues of components in water quality indicators time-series coming from different sources and the time-scale of water quality event is uncertain, this paper proposed a water quality anomaly detection algorithm based on EMD and Generalized Zero Crossing (GZC) theory, which involves local time-scale match. First, utilize EMD to "sift" raw water quality time-series after pretreat, thus to obtain a series of Intrinsic Mode Functions (IMF), of which the vibration cycle is gradually increased. Then use 3δ threshold detection method to perform anomaly detection for each IMF, respectively. Then calculate the local time scale of every time step in each IMF, thus to attain the average time scale of each mode. For every observation point, calculate its local dynamic time, and decide its degree of belongings according to the definition of time-scale belonging. In this way, the most sensitive modes to specific water quality event are selected out and are fused to implement anomaly detection. The proposed method can be performed for event detection of different time-scale, has certain time-scale adaptability, and has certain detection ability for water quality events that are relatively long and smooth.2. For issues of mode mixing phenomenon in EMD and the ignorance of inner connection of different IMFs in former method because of the fusion strategy, proposed a method based on Ensemble Empirical Mode Decomposition (EEMD) and multi-dimensional anomaly detection. Utilize EEMD to decompose water quality indicator time series to get a series of IMFs, and after selecting all the sensitive scale, utilize the values at every time point of every sensitive mode to build target feature vector, then use training set data to train the model. For every observation point, calculate the Mahalanobis distance between its target feature vector and the training center, and decide whether it’s abnormal or not by setting different confidence coefficient and threshold. The proposed method can restrain the mode-mixing problem caused by non-stationary and intermittent characteristics of water quality indicator time series, and take full advantage of instinct links between sensitive IMFs and have a better performance in event detection. 3. For issue of distortion or radiation in both ends of the signal, known as edge effect, in the online application of EEMD, proposed a symmetric extreme points extension based EEMD method for online event detection. Through symmetric extension, the distortion and radiation can be effectively restrained. For each new observation, a slide window, which has a FIFO (first in first out) mechanism, is introduced to use improved EEMD to decompose the signal, thus to get a series of IMFs. The values of each IMF at the new observation in turn are used to build the target feature vector. And the state of current point is determined by its Mahalanobis distance between its target feature vector and training center and seeing if the distance is beyond threshold under different confidence. This method can effectively inhibit edge effect of traditional EEMD in real-time application and promote the detection rate of water anomaly event detection.