CAJ | 학술논문

Kalman 神经网络以其良好的自适应非线性逼近能力，被广泛用于复杂非线性动态工业过程建模。传统噪声估计方法难以得到观测噪声不确定动态工业过程的噪声估计值，因而常将观测噪声估计值置零以进行 Kalman 神经网络建模，影响Kalman神经网络的建模效果，限制了Kalman神经网络在观测噪声不确定动态工业过程建模中的应用。有效利用观测输入输出数据，提出样本有效噪声估计(Gamma test, GT)改进的Kalman神经网络建模方法。采用衰减记忆的GT对输入输出数据进行实时估计，得到准确的观测噪声估计值，再利用Kalman神经网络实现精确建模。验证结果表明，该方法对EKF神经网络模型和 UKF 神经网络模型均有很好的改善作用，有效解决观测噪声不确定引起的 Kalman 神经网络模型发散问题，为采用Kalman神经网络建立噪声不确定动态工业过程的精确模型提供了一条有效途径。
Kalman 신경망락이기량호적자괄응비선성핍근능력，피엄범용우복잡비선성동태공업과정건모。전통조성고계방법난이득도관측조성불학정동태공업과정적조성고계치，인이상장관측조성고계치치령이진행 Kalman 신경망락건모，영향Kalman신경망락적건모효과，한제료Kalman신경망락재관측조성불학정동태공업과정건모중적응용。유효이용관측수입수출수거，제출양본유효조성고계(Gamma test, GT)개진적Kalman신경망락건모방법。채용쇠감기억적GT대수입수출수거진행실시고계，득도준학적관측조성고계치，재이용Kalman신경망락실현정학건모。험증결과표명，해방법대EKF신경망락모형화 UKF 신경망락모형균유흔호적개선작용，유효해결관측조성불학정인기적 Kalman 신경망락모형발산문제，위채용Kalman신경망락건립조성불학정동태공업과정적정학모형제공료일조유효도경。
Kalman filter neural network(KFNN) have been widely used in modeling for complex industrial process, because they have abilities to adaptive approximate the nonlinear and dynamic properties of the process. However, the performances of KFNN will diverge because it can’t get accurate statistics of unmeasurable noise by traditional noise estimation methods. A new KFNN with gamma test(GT) is proposed for industrial process modeling with the unmeasurable noise. The moving window idea is introduced to GT algorithm, and the improved GT is used to track the changes of the observable noise covariance in real time because it can get the accurate statistics of the unmeasurable noise only use the input-output data. Then the covariance in the traditional KFNN is replaced by real-time estimation from the improved GT algorithm. In this way, the KFNN is enhanced by the GT algorithm. In order to verify, the proposed KFNN is used to model the industrial process. The efficiency of the new KFNN is verified by complex hydrocyanic acid(HCN) industrial process. Verification show that the performance of the proposed KFNN model superior to those of the traditional KFNN, e.g. the extended Kalman filter artificial neural network(EKFNN) and the unscented Kalman filter artificial neural network(UKFNN). Therefore, the proposed method provides a new solution to get the accurate model of the industrial process with unmeasurable noise.