Abstract: Fault diagnosis is a critical technology for ensuring production safety and maintaining product quality within the chemical industry. However, many contemporary fault diagnosis methods that leverage deep neural networks are typically trained using samples from specific operating conditions, often neglecting the domain generalization capabilities of these models. This oversight results in a marked decline in diagnostic performance when the models are applied to the diverse and complex conditions of real industrial scenarios. This paper proposes a weighted-based domain-specific feature removal network (WBDSFRN) to address this challenge. WBDSFRN incorporates a weight-based domain-specific feature removal module, which effectively discriminates between domain-invariant and domain-specific features during the training phase. In the testing phase, the model largely eliminates domain-specific features of the target domain, thereby alleviating domain shift and improving the model’s generalization ability. The effectiveness of the proposed method is rigorously validated via comparative experiments and ablation studies using the Tennessee Eastman Process (TEP). The results demonstrate that WBDSFRN significantly outperforms existing approaches and exhibits robust diagnostic performance even under complex and variable operating conditions.