An unsupervised water quality anomaly detection method based on a combination of time-frequency analysis and clustering.

Anomalies in water quality, which frequently arise due to pollution, constitute a substantial menace to human health. The preservation of public welfare critically entails the timely recognition of abnormal water quality. Conventional techniques for detecting water quality anomalies face obstacles such as the necessity of expert knowledge, limited accuracy in detection, and delays in identification. In this paper, we proposed an original unsupervised technique for identifying water quality anomalies combined with time-frequency analysis and clustering (TCAD). We chose time-frequency analysis because it effectively evaluates water quality changes, generating distinct multi-band signals that reflect different aspects of water quality dynamics. We also proposed a clustering technique which can identify water quality markers and amalgamate data from multi-band signals for accurate anomaly detection. We seek to clarify the reasoning behind our methodology by portraying how time-frequency analysis and clustering address the deficiencies of conventional methods. Our experiments evaluated various indicators of water quality, and the effectiveness of our proposed approach was supported by comparative analyses with commonly used models for detecting anomalies in water quality.

An improved graph convolutional network with feature and temporal attention for multivariate water quality prediction.

The analysis and prediction of water quality are of great significance to water quality management and pollution control. In general, current water quality prediction methods are often aimed at single indicator, while the prediction effect is not ideal for multivariate water quality data. At the same time, there may be some correlations between multiple indicators which the conventional prediction models cannot capture. To resolve these problems, this paper proposes a deep learning model: Graph Convolutional Network with Feature and Temporal Attention (FTGCN), realizing the prediction for multivariable water quality data. Firstly, a feature attention mechanism based on multi-head self-attention is designed to capture the potential correlations between water indicators. Then, a temporal prediction module including temporal convolution and bidirectional GRU with a temporal attention mechanism is designed to deal with temporal dependencies of time series. Moreover, an adaptive graph learning mechanism is introduced to extract hidden associations between water quality indicators. An auto-regression module is also added to solve the disadvantage of non-linear nature of neural networks. Finally, an evolutionary algorithm is adopted to optimize the parameters of the proposed model. Our model is applied on four real-world water quality datasets, compared with other models for multivariate time series forecasting. Experimental results demonstrate that the proposed model has a better performance in water quality prediction than others by two indices.