Abstract: In recent years, Chinese iron and steel enterprises have mainly adopted the “sampling after the event” method to inspect the product quality before it leaves the factory. Due to the inability to achieve quality inspection for all products, customers often claim and return defective products, leading to major economic losses in steel enterprises. To improve the stability and reliability of product quality, the use of machine learning methods to realize the online monitoring, optimization, and preset of product quality is the key technology to be solved in iron and steel enterprises. Therefore, the online identification and diagnosis of abnormal product quality based on the soft hypersphere, online optimization of the process parameters based on manifold learning and process specification formulation based on the multivariate statistical process control were proposed. In this study, integrated methods of online monitoring, diagnosis, and optimization of product quality were proposed in which the abnormal point of the product quality by the soft hypersphere method, based on the support vector data description, was identified online, and the process parameters were diagnosed through the contribution chart. Optimizing in real time, abnormal process parameters via a local projective transformation of neighbor points was then achieved. The process parameter setting model based on manifold learning by multiclass neighborhoods to extract the manifold of process parameters was established. Meanwhile, the process specification model, based on the maximum inner rectangle of the soft hypersphere, was established to obtain an effective control interval of the process parameters. Through system integration with the proposed methods and using industrial internet technology and big data analysis methods, the system of intelligent online monitoring of product quality has been successfully developed. At present, the system has been applied to more than ten production lines in iron and steel enterprises. The accuracy rate of online quality determination is 99.2%, and the online detection time is less than 0.1 s.