Research On Soft Sensor Model And Interpretability Based On Deep Learning

In industrial processes,many process variables such as temperature,pressure,flow rate,etc.are very important because they can affect the quality and output of products.However,due to the harsh industrial production environment and considering factors such as cost and economic benefits,many key variables are difficult to directly measure with hardware sensors.The emergence of soft measurement technology provides a new approach to solving this problem.Soft measurement technology is based on establishing mathematical models to infer and estimate variables that are difficult to measure directly,as well as predicting and optimizing relevant parameters in industrial production processes.Soft sensing technology provides an efficient detection and control method for existing sensors,making it a research focus in the field of complex industrial process control.At present,research on soft sensing mainly focuses on modeling methods.With the rapid development of high-performance computing technology,data-driven models,especially those based on deep learning,are increasingly being used for soft sensing modeling.However,the black box nature of data-driven models leads to a lack of interpretability,which limits their application in modeling schemes,as models lacking interpretive results may not gain trust.In this paper,it is hoped that the interpretability of the model can be improved by designing and improving the model structure,thereby enhancing the credibility of the model results.The main research content of this paper includes:(1)The air preheater of a 600MW power plant boiler is taken as the research object in this paper.First,collect relevant industrial data,use the method to eliminate the outlier in the process data,and then use the Spearman correlation coefficient method to select the variables that have a greater impact on the rotor thermal deformation as the input variables of the model.Finally,use the min-max method to standardize the data.(2)Two interpretable soft measurement methods based on deep learning are proposed in this paper to address the issue of unreliable measurement results due to the lack of interpretability of deep learning soft measurement models based on complex industrial data,and applies them to predict the thermal deformation of air preheater rotors.Firstly,Scheme 1 proposes an LSTM network model based on attention mechanism,which extracts temporal features from the dataset using LSTM network.By adding attention mechanism to LSTM network,the deep network focuses on more important information,improving the prediction accuracy of the model.And obtained attention weights provide time-dependent and interpretable results.(3)Based on the model proposed in Scheme 1,Scheme 2 proposes a deep learning model called GCN-ATTENTION-LSTM.The network structure consists of a temporal feature extraction module composed of the model in Scheme 1,and a stacked graph convolutional network with added attention mechanism forms the spatial feature(relationship between features)extraction module of the model,further improving prediction performance.In order to make the data suitable for the input form of graph convolutional networks,this model treats each data sample as a node and uses a neural network to learn the relationship between two nodes in the data to obtain the graph structure.Then,a graph convolutional network with an attention mechanism is used to extract features from the graph structure.When the model training is completed,the learned optimal graph structure can intuitively reflect the relationship between variables in the dataset,together with the attention matrix,the interpretability of the model in both temporal and spatial dimensions is demonstrated.In order to verify its performance,the model was applied to predict the thermal deformation of the air preheater rotor.Compared with other commonly used soft measurement models and existing more advanced soft measurement models,the deep learning based soft measurement model proposed in this paper not only improves the prediction accuracy,but also has better interpretability due to the addition of interpretable components.(4)Due to the fact that the interpretable results on the spatial dimension obtained by the model proposed in Scheme 2 can only show which features have an impact on the prediction of rotor shape variables,but it is not clear how these features affect the prediction results.Therefore,in order to further understand how the model system makes decisions,the SHAP method is adopted for post interpretability of the model,further improving its interpretability.