In modern industry, the health monitoring of unit equipment is crucial. Due to the complexity of unit equipment, there exist intricate relationships among various parameters sensed by sensors, which influence the operational state of the unit. To enhance the accuracy of short-term unit state prediction, this paper proposes a joint framework based on CNN, LSTM, and attention mechanism. In this study, we introduce a parallel expanded framework where CNN captures complex dependencies among variables instead of conventional matrix transformations to obtain the required Q in the attention mechanism, LSTM captures the temporal correlations among data instead of K, while the sequence itself serves as the value vector V. The attention mechanism balances the importance of features for prediction. Through multiple comparative experiments, the proposed model demonstrates higher prediction accuracy compared to other methods, showing promising reliability in predicting unit states based on sensor parameters.