To realize the quick and accurate measurement of the dissolved oxygen concentration (DO) in wastewater treatment process, a novel chaos-fireworks algorithm (FWA) based hybrid optimization algorithm was proposed and a neural network on-line soft-sensor model was built based on the improved algorithm. According to the property of the data collected from wastewater treatment process, a new measure of similarity degrees between samples was defined to extract more responsive modeling data. In the novel algorithm, a modified two-level sinusoidal chaotic mapping was defined and the initial members of FWA were well selected by utilizing the ergodicity of chaos. As a result, the quality of the initial population in standard FWA was improved. Next, the search mechanism of FWA was modified by introducing chaos optimization algorithm. The optimization procedure was divided into two phases and the population was divided into two subpopulations according to the predefined criterion. Test results confirmed that the improved FWA had higher convergence speed and convergence accuracy. The novel soft-sensor modeling method and the sample data extraction method was used to build a soft sensor model for real-time measuring DO in wastewater treatment process. Application results indicated the root mean square error and the root mean square error of this model were 0.0175 and 0.0118 respectively, it had good generalization ability.