In microseismic monitoring systems (MMS), efficiently and accurately recognizing fracture waveforms and picking their arrival times of P-wave and S-wave are two crucial signal processing tasks for timely rockburst warnings. Previous studies have overlooked the potential relationship between these two tasks and have processed them independently, leading to redundant computations and performance bottlenecks. An intuitive solution is to integrate the two tasks into a multitask learning (MTL) framework. However, due to their competing optimization objectives, where recognition primarily focuses on global shape features and picking relies on local detail features, conventional parallel MTL structures often suffer from conflicting gradients. This conflict leads to a severe seesaw phenomenon that results in significant performance degeneration in picking task. To address this issue, we proposeMSNet, a cascade MTL framework forMicroseismicSignal processing. In this framework, a self-supervised fracture-perception module and a dual-scale fusion module are proposed to model the hierarchical dependence between recognition and picking tasks. Specifically, these modules enable a forward cascaded feature refinement from recognition to picking, as well as an information supplement from picking to recognition. Thus, MSNet transforms the two tasks from parallel competing learning into a mutually cooperative process that effectively mitigates the seesaw phenomenon. Finally, MSNet is validated on two datasets collected from different tunnel boring machine excavation projects. The results demonstrate its superiority over state-of-the-art methods in terms of accuracy that achieves an approximate 2% improvement under laboratory conditions. Furthermore, the practical applicability of MSNet is further validated through its deployment in a real-world tunnelling project, where it successfully detects rockburst events and provides real-time warnings.