Abstract: To address the problem that time-varying statistical characteristics of normal operating conditions in nonstationary industrial processes are easily confused with abnormal features, resulting in elevated false alarm rates and degraded abnormal monitoring performance, we propose a joint-optimized-based abnormal monitoring method for nonstationary process. Within a unified framework, shallow statistical features and deep representation features are adaptively fused to characterize the global time-varying behavior of nonstationary processes. On this basis, we construct an abnormal monitoring model based on a Gaussian mixture model, and introduce an estimation network to enable efficient parameter updating, thereby avoiding the computational burden associated with traditional iterative algorithms. Meanwhile, we formulate a joint optimization objective incorporating reconstruction error and sample likelihood is formulated to achieve collaborative optimization of feature learning and abnormal modeling. We then design an abnormal monitoring strategy based on the negative log-likelihood to effectively identify abnormal states. Experimental results obtained from the Tennessee Eastman benchmark and real wastewater treatment process demonstrate that the proposed method can significantly reduce false alarms under nonstationary normal operating conditions while maintaining high abnormal monitoring rates, indicating its strong potential for practical industrial applications.