Performance of Improved Gaussian Extreme Learning Machine for Water Pipeline Leak Recognition

At present, many machine- or deep-learning algorithms have been applied to water distribution network (WDN) leakage recognition. The higher the algorithm classification accuracy, the higher the leakage accident recognition accuracy for WDNs. This is of great significance for the green water-saving systems construction. As a popular machine-learning algorithm, the extreme learning machine (ELM) is also applied in WDN leakage recognition due to its fast learning speed and high accuracy. However, the traditional ELM obtains the classification model output weights based on the mapping calculation of the randomly initialized input weights. Due to the random initialization action, the traditional ELM input weights do not consider the statistical characteristics of the leak samples data, which can easily lead to low accuracy or classification error phenomena. In this article, we will use the Gaussian mixture model (GMM) framework to evaluate the leak sample data, and the new input weights will be generated by GMM evaluation. The new input weights can prompt the ELM to avoid classification errors and low accuracy drawbacks. We have named this improved ELM algorithm as Gaussian ELM (GELM) algorithm. We also compared GELM with the traditional ELM and other machine-learning algorithms, and the classification accuracy of the GELM was significantly higher.