Building collapses during recent earthquakes have brought up the need for research on factors pertaining to collapse and the safety of structures.This requires response replication of structures that account for uncer...Building collapses during recent earthquakes have brought up the need for research on factors pertaining to collapse and the safety of structures.This requires response replication of structures that account for uncertainties from ground motions and structural properties.Structural collapse often implies that the structural system is no longer capable of maintaining its gravity load-carrying capacity,which often points to factors involving strength and stiffness degradation.In this study,the polynomial chaos nonlinear autoregressive with exogenous input form(PC-NARX)model is explored for dynamic response replication of a nonlinear single-degree-of-freedom(SDOF)structure.The generalized hysteretic Bouc-Wen model is applied to emulate stiffness and strength degradation for an SDOF structure close to collapse.A stochastic ground motion model is used to represent the uncertainties in seismic excitation.The PC-NARX model is employed and further evaluated for response replication of an SDOF system with inherent uncertain structural properties.A generic algorithm(GA)is used to select the terms for structural dynamics,and polynomial chaos expansion(PCE)is used to incorporate uncertain parameters into NARX model coefficients.It is demonstrated that the PC-NARX model provides good accuracy to account for both ground motion and structural uncertainties into response replication of SDOF structures with significant strength and stiffness degradation.The PC-NARX model thus presents a promising technique for collapse safety analysis of structures.展开更多
随着新能源大规模接入电网,发电型燃气轮机常需频繁切换工作状态,导致故障风险上升,因此,异常检测对燃气轮机安全运行更加重要。针对燃气轮机异常检测问题,提出了一种基于NARX-Catboost算法的基线建模方法。采用NARX建立燃气轮机声压特...随着新能源大规模接入电网,发电型燃气轮机常需频繁切换工作状态,导致故障风险上升,因此,异常检测对燃气轮机安全运行更加重要。针对燃气轮机异常检测问题,提出了一种基于NARX-Catboost算法的基线建模方法。采用NARX建立燃气轮机声压特征信号的基线模型,引入CatBoost算法以增强NARX拟合能力,并运用贝叶斯优化对模型超参数进行寻优,最终通过实验数据验证了该融合方法在异常检测方面的有效性。另外,将所提NARX-Catboost与基于向前回归正交最小二乘法的NARX模型(NARX-FROLS)和集成深度随机向量函数链接网络(Ensemble Deep Random Vector Functional Link network, edRVFL)方法及性能进行对比。结果表明:NARX-CatBoost方法对正常声压均方根值的拟合均方根误差(RMSE)值为0.008 50,拟合准确度明显优于NARX-FROLS与edRVFL方法;NARX-CatBoost方法对异常声压均方根的异常检测准确率为96.94%,表明通过正常声压特征数据建立基线模型进行异常检测的可行性与准确性。展开更多
非线性有源自回归(Nonlinear Auto Regressive eXogenous,NARX)模型在辨识过程中要求以随机激励作为输入信号,由于转子系统无法产生该信号,因此传统辨识方法不适用于此类系统。针对该问题,提出一种频域模型结构及对应的扫频辨识方法。首...非线性有源自回归(Nonlinear Auto Regressive eXogenous,NARX)模型在辨识过程中要求以随机激励作为输入信号,由于转子系统无法产生该信号,因此传统辨识方法不适用于此类系统。针对该问题,提出一种频域模型结构及对应的扫频辨识方法。首先,通过快速傅里叶变换(Fast Fourier Transform,FFT)得到系统扫频过程中全部时域响应的频谱,提取关键频率成分对应的频域数据,并组建频域模型结构下的候选模型项库。其次,引入基于预测残差平方和的正交向前回归算法(the Orthogonal Forward Regression algorithm,based on the Predictive Residual Sum of Squares,PRESS-based OFR)完成模型项选取。最后,通过简单转子系统数学模型得到的仿真数据及转子实验台得到的真实数据验证所提出建模方法的有效性和适用性,同时,也通过上述两种数据探讨了扫频过程中引入不同的频率成分信息对预测精度的影响。结果表明,采用频域数据得到的模型结构和对应的系数能准确预测系统时域响应。在仿真数据上,模型预测的归一化均方误差(Normalized Mean Square Error,NMSE)不超过0.3%,而实验数据的NMSE进一步降低至0.2%。此外,通过对比不同模型结构,包含二倍频对应数据的模型(模型二)在预测系统二倍频处对应的幅值时的精度比仅使用基频数据建立的模型(模型一)更准确,NMSE在转速为339.3 rad/s工况下提升了95.58%;因此,丰富的频域数据信息有利于得到更准确的模型,提出的方法为谐波激励系统的辨识提供了新思路。展开更多
基于台架采集数据,采用外部输入非线性自回归(nonlinear autoregressive model with exogenous input,NARX)神经网络建立了具备瞬态特性的柴油机排气温度计算模型作为虚拟传感器,并采用并发式训练方法对模型进行训练。将结果与前馈神经...基于台架采集数据,采用外部输入非线性自回归(nonlinear autoregressive model with exogenous input,NARX)神经网络建立了具备瞬态特性的柴油机排气温度计算模型作为虚拟传感器,并采用并发式训练方法对模型进行训练。将结果与前馈神经网络、长短期记忆网络(long short term memory,LSTM)神经网络及量产发动机的排温传感器采集结果进行对比。经验证,稳态工况下,两种神经网络均能达到较高精度;欧洲瞬态循环(European transient cycle,ETC)工况下,NARX神经网络计算温度的最大偏差为6.6℃,量产发动机排温传感器测得温度最大偏差为45.9℃。NARX神经网络所需的计算时间约为现有电控单元排温模型的2.5倍。展开更多
针对非线性自回归模型(Nonlinear Auto-Regressive with extrainput,NARX)系统辨识问题,利用非正交的方法来构造较为稀疏的逼近NARX模型的径向基函数模型。与已有的径向基或其他的核模型只采用同一固定的尺度不同,采用多个尺度,通过最...针对非线性自回归模型(Nonlinear Auto-Regressive with extrainput,NARX)系统辨识问题,利用非正交的方法来构造较为稀疏的逼近NARX模型的径向基函数模型。与已有的径向基或其他的核模型只采用同一固定的尺度不同,采用多个尺度,通过最小化当前训练误差,选择最佳的核中心和尺度参数。在学习过程中,采用非正交核函数的方法进行模型逐步回归。对样本数据利用k均值聚类算法得到核函数中心参数备选项,同时设置多个备选尺度,并通过最小二乘法求得相应核函数的权值,利用前向选择方法从中找出使模型误差最小的最优核函数。仿真实验验证了方法在泛化性能和稀疏性方面的可行性。展开更多
智能电能表在复杂电网环境下的计量精度易受非线性误差影响。为提高其准确性,提出一种融合非线性自回归外生输入(nonlinear auto-regressive with exogenous inputs,NARX)模型与畸变功率的误差校正方法。利用NARX模型构建电能表的非线...智能电能表在复杂电网环境下的计量精度易受非线性误差影响。为提高其准确性,提出一种融合非线性自回归外生输入(nonlinear auto-regressive with exogenous inputs,NARX)模型与畸变功率的误差校正方法。利用NARX模型构建电能表的非线性误差模型,以捕捉其动态特性;从测量数据中分离基波与谐波电能,并计算谐波电能比差以量化谐波影响;引入畸变功率概念,构建以谐波电能比差和畸变功率为输入的误差校正模型,对非线性误差进行补偿。实验结果表明:经所提方法校正后,在不同谐波含量(5%、10%、20%)条件下,智能电能表的最大计量误差由校正前的2.4%、3.1%、10.3%均降至0.5%左右,同时非线性误差预测结果的拟合度得到了提升,有效提高了谐波环境下的计量精度。展开更多
Many physical processes have nonlinear behavior which can be well represented by a polynomial NARX or NARMAX model. The identification of such models has been widely explored in literature. The majority of these appro...Many physical processes have nonlinear behavior which can be well represented by a polynomial NARX or NARMAX model. The identification of such models has been widely explored in literature. The majority of these approaches are for the open-loop identification. However, for reasons such as safety and production restrictions, open-loop identification cannot always be done. In such cases, closed-loop identification is necessary. This paper presents a two-step approach to closed-loop identification of the polynomial NARX/NARMAX systems with variable structure control （VSC）. First, a genetic algorithm （GA） is used to maximize the similarity of VSC signal to white noise by tuning the switching function parameters. Second, the system is simulated again and its parameters are estimated by an algorithm of the least square （LS） family. Finally, simulation examples are given to show the validity of the proposed approach.展开更多
基金National Natural Science Foundation of China under Grant No.51878390。
文摘Building collapses during recent earthquakes have brought up the need for research on factors pertaining to collapse and the safety of structures.This requires response replication of structures that account for uncertainties from ground motions and structural properties.Structural collapse often implies that the structural system is no longer capable of maintaining its gravity load-carrying capacity,which often points to factors involving strength and stiffness degradation.In this study,the polynomial chaos nonlinear autoregressive with exogenous input form(PC-NARX)model is explored for dynamic response replication of a nonlinear single-degree-of-freedom(SDOF)structure.The generalized hysteretic Bouc-Wen model is applied to emulate stiffness and strength degradation for an SDOF structure close to collapse.A stochastic ground motion model is used to represent the uncertainties in seismic excitation.The PC-NARX model is employed and further evaluated for response replication of an SDOF system with inherent uncertain structural properties.A generic algorithm(GA)is used to select the terms for structural dynamics,and polynomial chaos expansion(PCE)is used to incorporate uncertain parameters into NARX model coefficients.It is demonstrated that the PC-NARX model provides good accuracy to account for both ground motion and structural uncertainties into response replication of SDOF structures with significant strength and stiffness degradation.The PC-NARX model thus presents a promising technique for collapse safety analysis of structures.
文摘随着新能源大规模接入电网,发电型燃气轮机常需频繁切换工作状态,导致故障风险上升,因此,异常检测对燃气轮机安全运行更加重要。针对燃气轮机异常检测问题,提出了一种基于NARX-Catboost算法的基线建模方法。采用NARX建立燃气轮机声压特征信号的基线模型,引入CatBoost算法以增强NARX拟合能力,并运用贝叶斯优化对模型超参数进行寻优,最终通过实验数据验证了该融合方法在异常检测方面的有效性。另外,将所提NARX-Catboost与基于向前回归正交最小二乘法的NARX模型(NARX-FROLS)和集成深度随机向量函数链接网络(Ensemble Deep Random Vector Functional Link network, edRVFL)方法及性能进行对比。结果表明:NARX-CatBoost方法对正常声压均方根值的拟合均方根误差(RMSE)值为0.008 50,拟合准确度明显优于NARX-FROLS与edRVFL方法;NARX-CatBoost方法对异常声压均方根的异常检测准确率为96.94%,表明通过正常声压特征数据建立基线模型进行异常检测的可行性与准确性。
文摘非线性有源自回归(Nonlinear Auto Regressive eXogenous,NARX)模型在辨识过程中要求以随机激励作为输入信号,由于转子系统无法产生该信号,因此传统辨识方法不适用于此类系统。针对该问题,提出一种频域模型结构及对应的扫频辨识方法。首先,通过快速傅里叶变换(Fast Fourier Transform,FFT)得到系统扫频过程中全部时域响应的频谱,提取关键频率成分对应的频域数据,并组建频域模型结构下的候选模型项库。其次,引入基于预测残差平方和的正交向前回归算法(the Orthogonal Forward Regression algorithm,based on the Predictive Residual Sum of Squares,PRESS-based OFR)完成模型项选取。最后,通过简单转子系统数学模型得到的仿真数据及转子实验台得到的真实数据验证所提出建模方法的有效性和适用性,同时,也通过上述两种数据探讨了扫频过程中引入不同的频率成分信息对预测精度的影响。结果表明,采用频域数据得到的模型结构和对应的系数能准确预测系统时域响应。在仿真数据上,模型预测的归一化均方误差(Normalized Mean Square Error,NMSE)不超过0.3%,而实验数据的NMSE进一步降低至0.2%。此外,通过对比不同模型结构,包含二倍频对应数据的模型(模型二)在预测系统二倍频处对应的幅值时的精度比仅使用基频数据建立的模型(模型一)更准确,NMSE在转速为339.3 rad/s工况下提升了95.58%;因此,丰富的频域数据信息有利于得到更准确的模型,提出的方法为谐波激励系统的辨识提供了新思路。
文摘基于台架采集数据,采用外部输入非线性自回归(nonlinear autoregressive model with exogenous input,NARX)神经网络建立了具备瞬态特性的柴油机排气温度计算模型作为虚拟传感器,并采用并发式训练方法对模型进行训练。将结果与前馈神经网络、长短期记忆网络(long short term memory,LSTM)神经网络及量产发动机的排温传感器采集结果进行对比。经验证,稳态工况下,两种神经网络均能达到较高精度;欧洲瞬态循环(European transient cycle,ETC)工况下,NARX神经网络计算温度的最大偏差为6.6℃,量产发动机排温传感器测得温度最大偏差为45.9℃。NARX神经网络所需的计算时间约为现有电控单元排温模型的2.5倍。
文摘针对非线性自回归模型(Nonlinear Auto-Regressive with extrainput,NARX)系统辨识问题,利用非正交的方法来构造较为稀疏的逼近NARX模型的径向基函数模型。与已有的径向基或其他的核模型只采用同一固定的尺度不同,采用多个尺度,通过最小化当前训练误差,选择最佳的核中心和尺度参数。在学习过程中,采用非正交核函数的方法进行模型逐步回归。对样本数据利用k均值聚类算法得到核函数中心参数备选项,同时设置多个备选尺度,并通过最小二乘法求得相应核函数的权值,利用前向选择方法从中找出使模型误差最小的最优核函数。仿真实验验证了方法在泛化性能和稀疏性方面的可行性。
文摘智能电能表在复杂电网环境下的计量精度易受非线性误差影响。为提高其准确性,提出一种融合非线性自回归外生输入(nonlinear auto-regressive with exogenous inputs,NARX)模型与畸变功率的误差校正方法。利用NARX模型构建电能表的非线性误差模型,以捕捉其动态特性;从测量数据中分离基波与谐波电能,并计算谐波电能比差以量化谐波影响;引入畸变功率概念,构建以谐波电能比差和畸变功率为输入的误差校正模型,对非线性误差进行补偿。实验结果表明:经所提方法校正后,在不同谐波含量(5%、10%、20%)条件下,智能电能表的最大计量误差由校正前的2.4%、3.1%、10.3%均降至0.5%左右,同时非线性误差预测结果的拟合度得到了提升,有效提高了谐波环境下的计量精度。
文摘Many physical processes have nonlinear behavior which can be well represented by a polynomial NARX or NARMAX model. The identification of such models has been widely explored in literature. The majority of these approaches are for the open-loop identification. However, for reasons such as safety and production restrictions, open-loop identification cannot always be done. In such cases, closed-loop identification is necessary. This paper presents a two-step approach to closed-loop identification of the polynomial NARX/NARMAX systems with variable structure control （VSC）. First, a genetic algorithm （GA） is used to maximize the similarity of VSC signal to white noise by tuning the switching function parameters. Second, the system is simulated again and its parameters are estimated by an algorithm of the least square （LS） family. Finally, simulation examples are given to show the validity of the proposed approach.
