| In large industrial processes,it is necessary to monitor the key parameters of the system in order to accurately optimize and control the industrial process.Some key parameters are difficult to be measured by hardware devices due to environment,cost or technology reasons,but soft sensor technology can realize the monitoring of these key parameters.So soft sensor has been widely used.Because the spatial distribution of industrial equipment makes the measurement sequence of key variables inconsistent with the sampling sequence of process variables,it shows obvious time-delay caused by signal,material transmission or installation location,and such time-delay has time-varying characteristics.In order to monitor the state of the system more reliably,it is necessary to consider the time-varying time-delay characteristics between the monitoring variables of the system when building the soft sensor model.In view of the above problems,this paper mainly focuses on the soft-sensing modeling problem with delay characteristics.The main work of this paper is as follows:(1)In the process of industrial data collection,sensor or network abnormality and other reasons will cause data missing or data distortion.Aiming at the problem of missing data,this paper uses linear interpolation,KNN interpolation and multiple linear regression interpolation methods to repair missing data,and compares and analyzes the effects of different methods.To solve the problem of data distortion,this paper proposes an outlier detection algorithm based on leave-one cross validation.The method does not need to assume that the original data obey a certain distribution,and analyzes whether the data is distorted from the perspective of the model estimation error.The results show that the method is effective in outlier detection.(2)Aiming at the problem that there are many kinds of variables collected in industrial process and there is redundancy among variables,this paper introduces a two-stage auxiliary variable selection method.In this method,irrelevant features were deleted by the maximum information coefficient method,and redundant features were eliminated by the approximate markov blanket method.Experimental results based on real industrial data show that the proposed method can significantly reduce the dimension of auxiliary variables and improve the modeling efficiency while ensuring the modeling accuracy.(3)Aiming at the key variable estimation problem with multi-stage time-delay characteristics in industrial processes,this paper proposes a key variable estimation method based on the online identification of time-delay parameters.In this method,the time-delay parameter data set was obtained from the original data by the maximum information coefficient method based on the sliding window,and the online time-delay parameter estimation model was constructed according to the time-delay parameter data set.The original data were reconstructed in time sequence by the time-delay parameter data set,and then the soft sensor model was established according to the reconstructed data.Experimental results based on industrial data with multi-stage time-delay characteristics show that the proposed method has a better estimation effect than the traditional method.(4)Aiming at the estimation of key variables with unclear time-delay characteristics in industrial processes,this paper proposes a weighted ridge regression method based on local information.By referring to the structure of time-delay neural network,this method takes the historical information into consideration to increase the information of the model,uses ridge regression to replace the neuron nodes to improve the stability of the model,searches the optimal parameters of the model through particle swarm optimization algorithm,and realizes the accurate estimation of the key variables.Compared with time-delay neural network and ridge regression,the proposed method is more effective in the estimation of industrial process variables with uncertain time-delay characteristics. |
PDF Full Text Request