Complex industrial process data modeling method based on dynamic convolutional neural network

Abstract

The invention discloses a complex industrial process data modeling method based on a dynamic convolutional neural network. The method comprises the following steps: process variables with strong correlation with industrial process objective variables are selected, and through sampling, a time sequence for each process variable is obtained; an equal depth box plot is used to carry out abnormal point detection and elimination on the time sequences, and a linear interpolation method is then used for filling; the time sequence for each process variable in a former process time delay range at the sampling moment of the objective variable is extracted, a two-dimensional matrix containing dynamic characteristics of the process is formed, and picture samples are formed; and the dynamic convolutional neural network is built to analyze the dynamic characteristics of the industrial process data, the time and space relation of each sensitive variable is recognized automatically, and a prediction model for the objective variable is built. A large amount of historical data accumulated in the actual production process field is used, a data model of predicting an unpredictable objective variable by using a predictable process variable is built accurately, and an important role is played in online production process evaluation, dynamic adjustment and even energy conservation and emission reduction.

Description

technical field [0001] The invention relates to the technical field of data-driven soft sensor modeling, in particular to a dynamic convolutional neural network data prediction modeling method suitable for complex industrial processes. Background technique [0002] The performance indicators of large-scale and complex industrial processes often rely on the "fixed-point and timing" offline assay method for detection. This detection method has serious hysteresis and is very unfavorable to the online evaluation and adjustment of the overall operating status of the process; and, considering operability, For issues such as safety, many indicators cannot or are difficult to be measured by sensors in the form of online assays. Therefore, it is particularly important to model complex industrial processes and predict indicators in real time. Predictive modeling is mainly divided into two types: mechanism-based and data-driven. Due to the complexity of the industrial process mechanis...

AI Technical Summary

Problems solved by technology

Linear regression is a commonly used data prediction modeling method in the early days, but this method is not suitable for nonlinear and complex industrial processes

SVR, KNN and other methods mainly use appropriate distance functions such as Euclidean distance to establish regression prediction models. This type of method can effectively solve the regression problem of high-dimensional features, but it is difficult to choose a suitable kernel function and distance function.

Algorithms such as CART, RF, and GBDT are predictive modeling machine learning methods that use decision tree-based predictors representing the mapping relationship between object attributes and objects. This type of method is robust to noise and can handle it well. Large-scale, high-dimensional data, but if the decision tree is too complex, it will lead to overfitting

Algorithms such as BPNN and ELM establish a hierarchical prediction model including nodes and node relationships based on the principle of graph theory. This type of algorithm has a strong theoretical foundation, rigorous derivation process, strong versatility, and can theoretically simulate any complex nonlinear process. However, the algorithm is easy to fall into local optimum, and the phenomenon of "gradient explosion" or "gradient dispersion" will appear, and there is a lack of unified and complete theoretical guidance when designing the structure

The above-mentioned complex industrial process data predictive modeling algorithms also have a common problem that they cannot effectively analyze the dynamic characteristics of the process and solve the inconsistency of data time series correlation

Images