A hybrid ensemble machine learning model for discharge coefficient prediction of side orifices with different shapes

Side orifices are widely applied for flow control and regulation in channel systems. Accurate estimation of the discharge coefficient of the side orifice is significant for water management. The main objective of current research is to accurately predict the discharge coefficients of circular and rectangular side orifices. Considering that traditional empirical regressions are hard to estimate the discharge coefficient precisely due to the complex nonlinear relationship between the discharge coefficient and relevant parameters, a new hybrid boosting ensemble machine learning model, BO-XGBoost, is developed, which combines the advantages of the boosting ensemble model (XGBoost) and Bayesian Optimization. To further evaluate the proposed hybrid model, it is also compared with other tree-based machine learning models, including standalone XGBoost, Random Forest (RF) and Decision Tree (DT). Literature experimental data of the flow and geometric parameters relevant to the discharge coefficients of circular and rectangular side orifices are collected and applied to develop the models. Four dimensionless parameters of the relative channel width (B/L), the relative bottom height (W/L), the relative upstream depth (Y/L) and the upstream Froude number (Fr) are taken into consideration for the prediction of discharge coefficient (C_d). Furthermore, four different input combinations are designed and then compared to determine the best one on the basis of RMSE. By using the optimal input combination, our results demonstrate that BO-XGBoost provides the best comprehensive performance among all the involved machine learning models in the discharge coefficient prediction for both types of side orifices. Besides, the uncertainty analysis also reveals that BO-XGBoost shows the narrowest uncertainty bandwidth and gives the highest prediction reliability.     

基于时空局部学习的集成自适应软测量方法

集成软测量方法已被广泛应用于流程工业关键质量参数实时估计。但是,常规集成建模方法在基模型构建过程中往往局限于挖掘样本之间的空间关系,忽略了样本间的时序关系,从而导致过程局部状态挖掘不充分、基模型间多样性不足等问题。其次,传统软测量方法由于缺乏自适应机制而无法有效处理过程时变特征,从而导致模型性能发生退化。为此,提出一种基于时空局部学习(STLL)的集成自适应软测量方法。该方法首先通过移动窗口、即时学习技术分别挖掘样本间的时序关系和空间关系,并采用统计假设检验实现冗余状态剔除,进而构建多样性的时空局部高斯混合回归(GMR)模型。然后,基于在线选择性集成策略实现局部预测结果的自适应融合。此外,引入双重自适应机制以缓解模型性能退化问题。实验结果显示,相较于非自适应全局GMR模型、时间局部学习集成GMR模型、空间局部学习集成GMR模型,所提方法在金霉素发酵过程中的预测精度分别提升了70.3%,14.9%,27.8%;在脱丁烷塔过程中,分别提升了31.9%,21.2%,19.3%。     

Adoption of machine learning in estimating compressibility factor for natural gas mixtures under high temperature and pressure applications

One of the essential properties of natural gas is its compressibility factor (z-factor), which is required for the efficient design of natural gas pipelines, storage facilities, gas well testing, gas reserve estimation, etc. Its importance has led to the development of several approaches involving new laboratory methods, equations of state (EOS), empirical correlations, and artificial intelligence for estimating gas compressibility factors. Most of the developed Z factor models have a limited range of applicability. They are unsuitable for predicting Z factors of highly pressurized gas reservoirs and natural gas systems with pseudo-reduced temperatures less than 1. Where such models exist, they are scarce and less accurate. In this study, three machine learning models, including the Gradient Boosted Decision Tree (GBDT), Support Vector Regression (SVR), and Radial Basis Function-Neural Network (RBF-NN), were developed for predicting the z-factor of natural gas mixtures with a range of P_pr and T_pr of 0–30 and 0.92–3.0, respectively. The results showed that the Gradient Boosted Decision Tree (GBDT) model outperformed other selected machine learning algorithms and published correlations. The proposed model gave a superior coefficient of determination (R² score), and root mean square (RMSE) of 0.99962 and 0.01033, respectively. Also, the variation of the Z factors from the GBDT model with pseudo-reduced pressures at different pseudo-reduced temperatures using the isotherm plot was found to be adequate. Hence, the GBDT model in this study is a reliable method for predicting Z factors of natural gas mixtures with P_pr and T_pr of 0–30 and 0.92–3.0, respectively. The plot revealed that the GBDT model performed extremely well in predicting compressibility factor with an MAPE of about 1%. The findings of this study shows that the proposed intelligent model can be utilized in predicting the gas Z-factor.     

Machine learning methods for wastewater hydraulics

Wastewater hydraulics problems are frequently addressed by investigation on physical models. Dimensional analysis is a powerful tool that allows discovering essential information about the investigated phenomenon, but in some cases it is affected by significant limitations. In such cases, many issues can be addressed by means of machine learning algorithms, resulting from the theories on pattern recognition and computational learning. In order to show the potential of such an approach, in this study Regression Tree M5P model, Bagging algorithm and Random Forest algorithm were applied to the solution of some complex problems of wastewater engineering: the prediction of energy loss, the pool depth, the air entrainment in a drop manhole, and the forecasting of the lateral outflow in a low crested side weir. The algorithms were trained and tested on data obtained from experimental tests that were carried out at the Water Engineering Laboratory of the University of Cassino and Southern Lazio. In most of the considered cases, regression trees and ensemble methods were able to provide very accurate predictions.     

基于多特征融合多核学习支持向量机的液压泵故障识别方法

提出基于多特征融合多核学习支持向量机的液压泵故障识别方法。该方法首先对原始信号进行集总经验模态分解,然后分别用AR模型和奇异值分解两种特征提取方法提取故障特征,最后将不同类型的特征分别用相应的核函数进行映射,用多核学习支持向量机来识别液压泵的工作状态和故障类型。实验结果表明该方法显著地提高了故障诊断的准确性。     

基于集成自适应欠采样的铝管表面缺陷检测方法研究

铝管作为一种常见的传输零件,对其表面缺陷进行检测是保证其生产质量、运行安全的必要措施。基于机器视觉的铝管表面缺陷检测方法因其检测精度高、速度快等优点,已取代人工检测,成为主流检测方法之一。但由于缺陷样本与背景样本之间分布不平衡,导致分类器决策边界偏移、检测精度下降,限制了其应用范围。针对这一问题,提出一种基于集成自适应欠采样的铝管表面缺陷检测方法,首先利用支持向量描述方法对数据分布间的重叠区域进行识别,其次通过构建样本局部密度关系自适应确定欠采样对象及数量,最终利用随机空间生成技术同时对数据样本空间和特征空间进行优化。试验结果表明,所提方法在铝管表面缺陷数据集上识别精确率达到98.52%,优于其他先进检测方法。     

直升机自动倾斜器竖向振动的神经网络识别

工程中的许多大型结构中的一些关键部件的振动无法直接测量，因而探索一种间接测量或识别的方法就显得尤为重要。由于直升机飞行状态的复杂性，不但测试困难，而且经典方法处理不理想。文中提出基于BP神经网络方法对直升机自动倾斜器竖向振动进行识别及评估。介绍了网络结构、训练学习过程、试验测试及数据处理过程和识别与估计方法。估计评价的准则主要考虑以估计第一阶谐和频率（6．4Hz)所对应的结果进行比较。由于神经网络方法考虑了不确定性因素，从而估计结果与真实结果相符得较好。最后分析了人工神经网络在用于动力学系统的识别与估计中的多方面问题。     

Ensemble semi-supervised Fisher discriminant analysis model for fault classification in industrial processes

In this paper, an ensemble form of the semi-supervised Fisher Discriminant Analysis (FDA) model is developed for fault classification in industrial processes. This method uses the K Nearest Neighbor (KNN) algorithm to merge the metric level outputs, which are obtained by the sub-classifiers in the ensemble model, to get the final classification result. An adaptive form is further proposed to improve the classification performance by putting forward to a new method of weight adjustment. While semi-supervised learning can generate a better model by exploiting additional information contained in unlabeled data, ensemble learning achieves the promotion of algorithm robustness by integrating a series of weak learners. In addition, the property of diversity in ensemble learning can be boosted by incorporating different unlabeled data to different weak learners. Therefore, the combination of those two methods can achieve great generalization for the fault classification model. The performances of two proposed methods are evaluated through an industrial benchmark process.     

Analog circuit fault diagnosis based on Quantum Clustering based Multi-valued Quantum Fuzzification Decision Tree (QC-MQFDT)

Analog circuit fault diagnosis is challenging due to the parametric deviation and the difficulty in signal quantizing. There still lacks effective approaches to provide reliable fault detection and classification results for a comprehensive diagnosis. In this paper, we propose a fault diagnosis methodology based on a new classification model called Quantum Clustering based Multi-valued Quantum Fuzzification Decision Tree (QC-MQFDT). QC-MQFDT incorporates the adaptive fuzzification method to discretize continuous-valued data. The fuzzification mechanism is devised by incorporating quantum clustering (QC) as well as the quantum membership function (QMF), where the former has the ability to sense the internal dependencies of data, and the latter uses the number of energy levels to approximate the optimal shape for fuzzy membership functions. The QF-C4.5 algorithm is developed as the decision tree learning algorithm, which employs quantum fuzzy entropy (QFE) to evaluate the information in the target variable space. The proposed method is validated using both simulated data and the real time data for the application studies of two benchmark analog circuits. The classification performances are discussed and the diagnostic capability of the model is verified through the application studies.     

基于Hankel矩阵联合近似对角化的结构模态识别

针对基于二阶盲辨识(second order blind identification,简称SOBI)的模态参数识别方法存在的不足,提出了一种基于Hankel矩阵联合近似对角化(Hankel matrix joint approximate diagonalization,简称HJAD)技术的结构运行模态分析(operational modal analysis,简称OMA)的新方法。该方法通过对随机子空间类模态识别方法常用的Hankel矩阵进行联合近似对角化,以分离各阶模态响应,进行模态识别。与基于SOBI的模态识别方法相比,在具体实施过程中,仅需要在分析数据中添加与实测振动响应对应的时间延迟的数据,实现难度较小。数值算例和物理模型试验的分析结果表明,所提出的基于HJAD技术的结构运行模态分析方法,不仅具有鲁棒性强和计算效率高的优点,还可以克服传统的基于SOBI的模态识别方法的模态识别能力受测点数目限制的问题。     

简支梁桥损伤程度识别的模拟和试验

在结合钢桁桥损伤程度识别方法的基础上,提出了适用于简支梁结构的两种损伤程度识别方法:整体振型的相关系数法和保证准则法,将其应用到实验室简支梁结构上分别进行数值模拟和试验。脉冲激励下的结果表明,两种方法能较准确地识别损伤单元的等效损伤程度,具有很强的抗噪能力。最后,探讨了激励对提出方法的影响,为工程应用奠定了基础。     

Research on soft sensing modeling method of gas turbine’s difficult-to-measure parameters

During the operation of a gas turbine, there are many key parameters that are difficult to directly measure or to ensure measurement accuracy, which can only be measured by offline analysis methods. However, the data obtained by offline analysis has a large time lag, and it is difficult to realize real-time monitoring, control and optimization of gas turbines. In recent years, with the widespread application of data-driven methods, data-driven soft sensing technology has become a breakthrough method for online prediction of difficult-to-measure variables. Due to the time-varying nature of the gas turbine operation process, the predictive performance of the offline modeling method will inevitably degrade over time. Therefore, an adaptive soft-sensing multi-level modeling method based on the combination of the just in time learning and the ensemble learning is proposed in this paper. Taking compressor inlet air flow and turbine inlet temperature as examples, the research is carried out and verified by actual operating data. The results verify the effectiveness of the method.

     

[设备健康管理]基于深度学习多样性特征提取与信息融合的行星齿轮箱故障诊断方法

针对行星齿轮箱振动信号噪声干扰大、单一分类器泛化能力不强的问题,提出了一种基于深度学习多样性特征提取与信息融合的行星齿轮箱故障诊断方法。利用多目标优化算法优化多个堆栈去噪自动编码器(SDAE)以获得多个性能优异的SDAE,并提取多样性的故障特征;采用多响应线性回归模型集成多样性故障特征实现信息融合,得到多目标集成堆栈去噪自动编码器(MO-ESDAE),最后将其应用于行星齿轮箱故障诊断。实验结果表明：该方法能有效提高故障诊断精度与稳定性,具有较强的泛化能力。     

基于流行-IMF奇异值熵的转子故障特征提取方法

针对转子振动信号的非平稳性以及微弱故障特征难以提取的问题，提出一种基于集合经验模式分解(ensemble empirical mode decomposition，简称EEMD)的奇异值熵和流形学习算法相结合的故障特征提取方法。首先，对原始振动信号进行EEMD分解，得到若干本征模态函数(intrinsic mode function，简称IMF)分量，根据峭度 欧式距离评价指标选取故障信息丰富的敏感分量，组成初始特征向量，求其奇异值熵；其次，利用近邻概率距离拉普拉斯特征映射算法（nearby probability distance Laplacian eigenmap,简称NPDLE）对奇异值熵组成的特征矩阵进行降维处理；最后，将得到的低维特征子集输入到K-近邻(K-nearest neighbor，简称KNN)中进行模式辨识。用一个双跨度转子实验台数据集和Iris仿真数据集对所提方法进行了验证，结果表明，IMF奇异值熵和NPDLE相结合的方法可以有效地实现转子故障特征提取，提高了故障辨识的准确性。     

Mass flow rate measurement of gas-liquid two-phase flow using acoustic-optical-Venturi mutisensors

The measurement of multiphase flow parameters is essential for the online monitoring of industrial production and energy metering. In this paper, a multi-sensor experimental measurement device is designed based on NIR, acoustic emission sensors, and throated Venturi. The measurement information is decomposed using modal decomposition, and the characteristic variables of the gas volume fraction are extracted by flow noise decoupling and light attenuation analysis. A new gas volume fraction model is proposed based on Gradient Boosting Decision Tree (GBDT) through feature-level fusion, and the Mean Absolute Percentage Error (MAPE) of the gas volume fraction prediction models is within 4% for the three flow patterns. A new flow rate model is established based on the Homogeneous and Collins models. Laboratory results indicate that the MAPE of the flow rate model is 1.56%, and 98.61% relative deviations are within ±20% error band. The study provides a new method for online measurement of multiphase fluid motion and a theoretical basis for sensing mechanism and measurement of multiphase flow.     

Biphase randomization wavelet bicoherence for mechanical fault diagnosis

Wavelet bicoherence is one of the most useful tools for quadratic nonlinear behavior identification of stochastic system, which has been used in many fields. However, current wavelet bicoherence algorithm can neither eliminate the spurious peaks coming from components with long coherence time, nor distinguish the quadratic phase coupling and non quadratic phase coupling signals, which may constraint the application of wavelet bicoherence. In this article, biphase randomization wavelet bicoherence technique is proposed to solve this problem. In this method, an ensemble average biphase randomization algorithm is established, in which the biphase randomization is employed to damage the biphase dependence among bispectrum samples. The spurious bicoherence coming from long coherence time waves and non phase coupling waves is eliminated efficiently by using the proposed method. Based on that, two diagnosis features are established for mechanical fault diagnosis. Simulation and experiment results demonstrate that the performance of the proposed method is much better than that of current wavelet bicoherence method.     

A multi-step progressive fault diagnosis method for rolling element bearing based on energy entropy theory and hybrid ensemble auto-encoder

It is meaningful to efficiently identify the health status of bearing and automatically learn the effective features from the original vibration signals. In this paper, a multi-step progressive method based on energy entropy (EE) theory and hybrid ensemble auto-encoder (HEAE), systematically blending the statistical analysis approach with the deep learning technology, is proposed for rolling element bearing (REB) fault diagnosis. Firstly, a preliminary detection about the REB health status is performed by the statistical analysis technique integrated with the EE theory. Secondly, if fault exists in REB, a new HEAE is constructed based on denoising auto-encoder and contractive auto-encoder to strengthen the feature learning ability and automatically extract the deep state features from the raw data. Subsequently, a modified t-distributed stochastic neighbor embedding (M-tSNE) algorithm is developed to achieve the features reduction to further improve the diagnosis efficiency. Finally, the low-dimensional representations after features reduction are as the inputs of softmax classifier to recognize the fault conditions. The proposed method is applied to the fault diagnosis of REB. The results confirm the effectiveness and superiority of the proposed method, and it is more suitable for the actual engineering applications compared with other existing methods.     

Neural Network Compact Ensemble and Its Applications

There has been many methods in constructing neural network (NN) ensembles, where the method of simultaneous training has succeed in generalization performance and efficiency. But just like regular methods of constructing NN ensembles, it follows the two steps, first training component networks, and then combining them. As the two steps being independent, an assumption is used to facilitate interactions among NNs during the training stage. This paper presents a compact ensemble method which integrates the two steps of ensemble construction into one step by attempting to train individual NNs in an ensemble and weigh the individual members adaptively according to their individual performance in the same learning process. This provides an opportunity for the individual NNs to interact with each other based on their real contributions to the ensemble. The classification performance of NN compact ensemble(NNCE) was validated through some benchmark problems in machine learning, including Australian credit card assessment, pima Indians diabetes, heart disease, breast cancer and glass. Compared with other ensembles, the classification error rate of NNCE can be decreased by 0.45% to 68%. In addition, the NNCE was applied to fault diagnosis for rolling element bearing. The 11 time-domain statistical features are extracted as the properties of data, and the NNCE is employed to classify the data. With the results of several experiments, the compact ensemble method is shown to give good generalization performance. The compact ensemble method can recognize the different fault types and various fault degrees of the same fault type.     

神经网络在温度控制系统中的应用

与常规控制方法相比较，神经网络控制系统有许多典型的先进特性，介绍了一些人工神经网络的基本概念及学习算法，最后，将神经网络PID控制器取代基本PID控制器用在浴室水箱温度控制中，仿真结果表明这种控制方法有很好的控制效果。     

Three-stage Method for Identifying the Dynamic Model Parameters of Stranded Wire Helical Springs

The dynamic behavior of the stranded wire helical spring is described by a modified Bouc-Wen model while the model parameters must be identified using an identification method and experimental data. Existing identification methods usually relies either solely nonlinear iterative algorithms or manually trial and error. Therefore, the identification process can be rather time consuming and effort taking. As a result, these methods are not ideal for engineering applications. To come up with a more practical method, a three-stage identification method is proposed. Periodic loading and identification simulations are carried out to verify the effectiveness of the proposed method. Noises are added to the simulated data to test the performance of the proposed method when dealing with noise contaminated data. The simulation results indicate that the proposed method is able to give satisfying results when the noise levels are set to be 0.01, 0.03, 0.05 and 0.07. In addition, the proposed method is also applied to experimental data and compared with an existing method. The experimental data is acquired through a periodic loading test. The experiment results suggest that the proposed method features better accuracy compared with the existing method. An effective approach is proposed for identifying the model parameters of the stranded wire helical spring.