Regularized Echo State Network Based on Human Learning Optimization Algorithm and Its Application in Fault Diagnosis

The input weights and reservoir node activations in Echo State Networks (ESN) are typically initialized randomly, which often leads to unstable output performance. Moreover, the output connection weights obtained through traditional pseudo-inverse methods may yield ill-conditioned solutions, adversely affecting the overall network performance. To address these issues, this study proposes a Regularized Echo State Network based on Human Learning Optimization (RESNHLO).The proposed method employs the Human Learning Optimization algorithm to dynamically search and update the ESN's input weight,while utilizing binary encoding to selectively activate reservoir nodes. This approach not only significantly improves the optimization efficiency of connection weights but also enables adaptive adjustment of the reservoir structure. Additionally, an L2 regularization strategy is incorporated to optimize the output connection weights, further enhancing the network's generalization capability. Comparative simulation results on fault diagnosis in an oceanic shipboard desalination system demonstrate that the RESNHLO method achieves superior diagnostic performance.