近红外光谱分析在工业过程故障检测方面具有独特的优势, 是一种准确且高效的方法。 结合互信息熵和传统的主成分分析, 对近红外光谱特征信息进行提取, 通过构建过程的模式来刻画工业过程的运行状态。 利用近红外光谱数据, 从有机分子含氢基团振动信息中获取工业系统的过程模式, 从微观分子层面探索提高工业过程故障检测准确率的有效方法, 结合贝叶斯统计学习技术, 提出了基于近红外光谱数据的工业过程故障检测技术。 针对近红外光谱信息量丰富, 谱带较宽, 特征性不强的特点, 首先对工业过程不同运行状态下的近红外光谱吸光度数据进行一阶导数预处理, 采用主成分分析法(principal component analysis, PCA)压缩光谱数据量, 扩大不同运行状态下光谱特征信息的差异性, 提取光谱的内部特征信息。 然后采用互信息熵(mutual information entropy, MIE)作为光谱特征信息相关性度量函数, 基于最小冗余最大相关算法进一步减少光谱特征信息间的冗余并最大化光谱特征信息与类别的相关性, 弥补了PCA无监督特征波长选择的不足, 提出一种基于PCA-MIE的过程模式构建方法, 获得的过程模式子集更紧凑更具类别表现力。 再利用贝叶斯统计学习算法, 根据后验概率对构建的模式子集进行决策, 判别生产过程的正常状态和故障状态。 由于过程模式子集结合了PCA浓聚方差的优势和互信息熵相关性测度的特征信息选择方法, 蕴含了更多的近红外光谱的本质信息与内在规律, 从而更能刻画工业过程的运行状态。 接着, 设置测试准确率TA作为评估标准, 用以评价故障检测方法的性能效果。 最后利用某化工厂提供的原油脱盐脱水过程近红外光谱数据对所提方法进行验证, 并与传统近红外光谱特征信息提取方法PCA和MIE方法性能进行对比分析, 结果表明基于PCA-MIE的过程模式故障检测方法几乎在所有维数子集上性能都优于其他两种方法, 在特征维数为18维时获得最高的准确率94. 6%, 证明了方法的优越性。

The technology of near infrared spectroscopy that has unique advantage in fault detection in industrial processes is an accurateand effective method. Combining the mutual information entropy and the traditional principal component analysis, a new method for extracting the near infrared spectral feature information was first developed. The operating states of the industrial process was described by the constructed process pattern.Near infrared spectroscopy data were used to obtain the process pattern of industrial systems from the vibration information of hydrogen groups in organic molecules in this paper. An effective method to improve accuracy of fault detection in industrial processes was explored from the microscopic molecular level. Combined with Bayesian statistical learning method, an industrial processes fault detection technique based on near infrared spectroscopy data was proposed. Firstly, for the characteristics of rich information, wide spectrum band and weak characteristic, first-order derivative preprocessing of near infrared spectroscopic absorbance data under different operating states of industrial process was applied. Principal component analysis(PCA) was used to compress the amount of spectral data, expand the differences in spectral feature information under different operating states, and extract the internal feature information of the spectrum. Then, mutual information entropy(MIE) was used as correlation measure function of spectral feature information, and the minimum redundancy maximum relevance algorithm was used to further reduce the redundancy between the spectral feature information and maximize the relevance between the spectral and class.It made up for the deficiency of unsupervised feature wavelength selection of PCA. Therefore, a process pattern construction method based on PCA-MIE was proposed. The obtained process pattern subset was more compact and more expressive. Furthermore, Bayesian statistical learning method was applied to make decisions based on posterior probability of the constructed process pattern subset to identify the normal and accident state of the production process. Because the process pattern subset combines the advantages of PCA in density variance reduction and the feature information selection method of mutual information entropy correlation measure, it contains more essential information and inherent laws of near infrared spectroscopy, which can better describe the operating states of the industrial process. Next, The test accuracy (TA) was set as the evaluation criteria to evaluate the performance of the fault detection method. Finally, the data of crude oil desalination and dehydration process provided by the chemical plant was used to verify the effectiveness of the proposed method. Compared with the performance of traditional near infrared spectral feature information selection methods PCA and MIE, the results showed that the process pattern fault detection based on PCA-MIE outperforms the other two methods on almost all dimensions subsets. The highest accuracy rate is 94.6% when the feature dimensions is 18, which proves the superiority of the proposed method.