Denial Constraint (DC) discovery aims to effectively and efficiently identify data quality issues. Nevertheless, balancing efficiency with task-specific applicability poses a substantial challenge. Unfortunately, most existing methods heavily rely on extensive enumeration and fail to adequately consider the requirements of downstream data cleaning tasks. To address these issues, we propose DCRer, a novel DC discovery framework based on deep reinforcement learning. DCRer models the DC discovery process as a Markov Decision Process (MDP) and utilizes a DC tree structure combined with a masking mechanism to guide predicate node selection. To improve training, we design a reward function and a Deep Q-Network (DQN). The reward function jointly considers the coverage, diversity, and completeness. Furthermore, the DQN integrates both Dueling DQN and Double DQN to improve model stability and mitigate estimation bias. Extensive experiments demonstrate the advantages of DCRer in terms of effectiveness, efficiency and scalability. Additionally, we integrate DCRer with existing data cleaning frameworks, validating its effectiveness in practical tasks.