To address the issues of long average flow time, high work-in-process (WIP) inventory and frequent equipment changeovers in semiconductor packaging workshop, an integrated scheduling of production and maintenance in semiconductor packaging workshop based on state feature fusion and multi-agent cooperation is studied under the consideration of limited buffer capacity and equipment maintenance constraint. Firstly, a multi-objective mathematical optimization model is established to minimize the makespan, total tardiness and total setup time. Secondly, according to the production characteristics and buffer capacity limits of different stations in the semiconductor packaging workshop, three heterogeneous production agents are designed by employing a twin delayed deep deterministic policy gradient (TD3) algorithm. Finally, to overcome the problems such as a large number of features and the difficulty of neural network learning in multi-agent TD3 (MATD3), a convolutional neural network (CNN) is employed to compress and fuse the state features of different agents, and an improved MATD3 algorithm based on CNN state feature fusion is constructed, named as CNN-MATD3. A large number of experimental results show that: (1) state feature fusion plays a significant role in CNN-MATD3 algorithm, and the average contributions to inverted generational distance (IGD) and hypervolume (HV) indicators are 216.93% and 26.52%, respectively; (2) CNN-MATD3 algorithm is significantly superior to two classical meta-heuristic algorithms, five single-agent algorithms and two multi-agent algorithms, the relative percentage deviations (RPDs) of each type of algorithm in IGD indicator are separately up to 16652.49%、7009.01% and 16788.91%, and their RPDs in HV indicator are separately no less than -95.46%、-56.81% and -86.65%; and (3) buffer capacity and product urgency have significant influences on the integrated scheduling results that are found by sensitivity analysis.