Abstract: Thermal power units are a multivariable, time-varying, and strongly coupled object, with related state parameters interconnected and influencing each other. During operation, there are a large number of delays and inertia phenomena. Traditional alarm strategies based on fixed threshold discrimination should not adjust the alarm threshold spontaneously in response to changes in operating conditions, on-site disturbances, or changes in system parameters associated with monitoring points, Make the alarm criteria more scientific and reasonable, increase the effectiveness of parameter warning, and reduce the false alarm rate. This article proposes a neural network-based anomaly monitoring method for multivariable systems. By combining neural network algorithms with historical operating data of the unit, monitoring parameter models under different operating conditions are trained. When conducting early warning, input the relevant parameters of the monitoring parameters into the trained model to obtain the evolution trend of the parameters to be warned under the current operating conditions. Subsequently, calculate the confidence interval of the parameters to be warned that no abnormalities have occurred, and then evaluate and judge whether there are abnormalities. Compared to traditional methods, anomalies can be detected faster. It has important reference significance for improving the operational reliability of similar air compressors in thermal power plants.