Industrial Process Time-Series Forecasting with CNN Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex industrial processes generate vast amounts of real-time data, making it challenging to predict future outcomes effectively due to the complexity and temporal nature of the data, with existing methods not fully leveraging historical data to forecast future behaviors or potential failures in processes like electricity production from wind turbines and pig iron production.
Innovation Solution
A method utilizing a convolutional neural network architecture, comprising a first convolutional layer, inception-module layers, a recurrent layer, and a dense layer, to generate and analyze time-series data subsets, allowing for the prediction of future outcomes by dividing data into observations with historical input variables and output variables recorded at sequential time points, enabling forecasting beyond the latest sequential time point.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If complex industrial process data is collected in real-time with multiple parameters, then the completeness and detail of process information is improved, but the complexity of data analysis and prediction increases
Solution Approach 1:
The patent segments complex industrial process data into structured time-series subsets with defined temporal relationships. Each subset contains input parameters, intermediate states, and output parameters organized by time sequences, making the complex data manageable for analysis while preserving complete process information.
Solution Approach 2:
The patent introduces an intermediary data processing layer that transforms raw complex process data into structured time-series subsets. This intermediary representation layer simplifies the data structure while maintaining information completeness, serving as a bridge between raw data and prediction algorithms.
2Measurement precision
If historical data is used to predict future process outcomes, then the accuracy of predictions is improved, but the time required for data processing and prediction increases
Solution Approach 1:
The patent performs preliminary organization of historical process data into structured time-series subsets before actual prediction is needed. By pre-structuring the data with defined temporal relationships and parameter associations, the system enables faster real-time predictions without sacrificing accuracy, as the data is already prepared in an analysis-ready format.
3Measurement precision
If data is divided into overlapping segments for analysis, then the detection of temporal patterns is improved, but the computational load increases
Solution Approach 1:
The patent divides process data into overlapping time-series segments where each segment contains input parameters, intermediate states, and output parameters. The overlapping structure allows temporal patterns to be detected across segment boundaries while maintaining computational efficiency through structured organization of parameters within each segment.
4Reliability
If machine learning models are trained with complex multi-dimensional time series data, then the predictive capability is improved, but the training time and resource requirements increase
Solution Approach 1:
The patent introduces structured time-series subsets as an intermediary representation that simplifies the input data for machine learning models. By organizing complex multi-dimensional time series data into standardized subsets with clear temporal and parameter relationships, the training process becomes more efficient while maintaining or improving predictive capability through better data structure.
Data Source
AI summary
Disclosed is a computer-implemented method of generating a time-series of data sets for predictive analysis from data comprising input variables and an output variable recorded at sequential time points, the method comprising: dividing the data into a collection of observations, each observation comprising: a subset of sequential time points; associated input variables; and an output variable recorded at a forecasting time point beyond the latest sequential time point of the subset; and using the collection of observations in a convolution neural network to predict the output at the forecasting time point of a streaming data set.