Abstract: The debutanizer column is a key unit in the naphtha fractionation system, responsible for separating light hydrocarbons such as C3~C5, and plays a crucial role in determining gasoline quality. In particular, the bottom C4 content is a critical quality indicator but is difficult to measure online due to the time delay and high cost of traditional laboratory analyses, limiting the efficiency of real-time control. To overcome this challenge, this work proposes a high-accuracy soft sensor model that integrates a deep belief network (DBN) and an improved whale optimization algorithm (IWOA) into an interval type-2 Takagi-Sugeno-Kang fuzzy logic system (IT2 TSK FLS) with an A2-C1 structure. The DBN is employed for deep feature extraction, enhancing data representation and reducing noise in the input process variables. Subsequently, the IWOA enhanced with cosine adjustment and step-size correction mechanisms is applied to optimize both the antecedent membership functions and the consequent parameters of the fuzzy logic system. This joint approach improves the models prediction accuracy and robustness. To evaluate the effectiveness of the proposed model, a comprehensive set of comparison experiments is conducted. The benchmark methods include support vector machines (SVM), long short-term memory (LSTM) networks, gated recurrent units (GRU), and IT2 TSK fuzzy logic systems optimized using backpropagation (BP), particle swarm optimization (PSO), grey wolf optimization (GWO), whale optimization algorithm (WOA), improved WOA (IWOA), and the proposed DBN-IWOA approach.The model achieves lower RMSE and MAE values, along with higher R2 scores, thereby validating its effectiveness and robustness in practical applications. In conclusion, the proposed approach shows strong potential for advancing soft sensor modeling in complex industrial processes, enabling more accurate and efficient real-time quality prediction and process control.

Key words: soft sensor modeling, debutanizer column, interval type-2 fuzzy logic system, deep belief network, premature convergence detection mechanism

摘要： 针对化工过程中存在的强非线性和复杂性问题，本工作提出了一种基于深度信念网络(DBN)与改进鲸鱼优化算法(IWOA)优化的区间二型TSK模糊逻辑系统(DBN-IWOA-IT2 TSK FLS)方法，以提升软测量建模的精度和稳定性。首先，DBN通过深度特征提取能力对输入数据进行处理，以减少噪声干扰并提取关键信息。随后，结合区间二型TSK模糊逻辑系统(IT2 TSK FLS)的建模优势，采用IWOA算法对前件参数和后件参数进行优化，以进一步增强模型的预测能力。IWOA通过引入早熟收敛检测机制，提高了全局搜索能力，加快了收敛速度，并降低了陷入局部最优的风险。最后，将所提出的方法应用于脱丁烷塔软测量建模，选取了支持向量机(SVM)、长短期记忆网络(LSTM)、门控循环单元网络(GRU)，以及分别基于反向传播算法(BP)、粒子群优化算法(PSO)、灰狼优化算法(GWO)、鲸鱼优化算法(WOA)、改进鲸鱼优化算法(IWOA)和DBN-IWOA优化算法的区间二型TSK模糊逻辑系统作为对比模型进行实验评估。结果显示，DBN-IWOA-IT2 TSK FLS在预测准确性、收敛速度均优于现有方法，验证了其有效性和工程应用价值。

关键词: 软测量建模, 脱丁烷塔, 区间二型模糊逻辑系统, 深度置信网络, 早熟收敛检测机制