Abstract: Dam seepage safety monitoring suffers a drawback of low detection efficiency and low accuracy by previous abnormal detection methods, and a problem that most of them fail to consider comprehensively the randomness and fuzziness of monitoring data in formulating the abnormal threshold values. This paper presents an improved deep sparse autoencoder (DSAE) method for the anomaly detection of dam seepage safety monitoring data. This method optimizes the hyperparameters of DSAE by using the improved Aquila optimization (IAO) algorithm based on chaotic initialization and nonlinear flight rate, and extracting residual component of monitoring data by the singular spectrum analysis method. The new IAO-DSAE method so obtained can realize a high-precision reconstruction of the residual components of monitoring data. To improve the reliability of anomaly detection, we modify the process of anomaly threshold formulation by adopting the expectation and entropy of the reverse cloud algorithm in replacement of the mean and standard deviation used in the traditional 3σ method, so as to achieve a comprehensive consideration of the influence of randomness and fuzziness of monitoring data on the formulation. An engineering case study shows that compared with the anomaly detection methods based on statistical model method and 3σ method, the prediction based on the seepage safety monitoring data that are processed by our new method can raise the accuracy by 5.56% and 6.99%, respectively, verifying the applicability and effectiveness of the new method.

Key words: seepage safety, anomaly detection, deep sparse autoencoder (DSAE), reverse cloud, improved Aquila optimization (IAO) algorithm, singular spectrum analysis