Bayesian Multivariate Time-Series Prediction for Missing Manufacturing Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manufacturing systems face challenges in effectively predicting and adjusting parameters using multivariate time series data, particularly due to missing values and asynchronous data acquisition, which affects the overall performance and maintenance of equipment.
Innovation Solution
A method utilizing a Bayesian model to interpolate missing values in multivariate time series data, store prediction data with uncertainty, and optimize model parameters for predicting future values, allowing for adjustments to be made in manufacturing system parameters based on the predicted data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional time series prediction methods are used, then the system can handle complete data, but it fails to effectively predict parameters when missing values and asynchronous data acquisition occur
Solution Approach 1:
The patent introduces a Bayesian imputation model as an intermediary component that processes missing and asynchronous data before it reaches the prediction system. This mediator fills gaps in the time series data using probabilistic methods, allowing the downstream prediction algorithms to work with complete, synchronized data sequences, thereby resolving the contradiction between prediction reliability and adaptability to incomplete data
Solution Approach 2:
The system dynamically adjusts the temporal resolution and synchronization parameters of the time series data based on the observed data quality and acquisition patterns. By changing parameters such as sampling intervals, aggregation windows, and imputation confidence thresholds, the system adapts to varying data completeness levels while maintaining prediction accuracy
2Measurement precision
If multivariate time series data is used to improve prediction accuracy, then the system can capture complex dependencies, but the complexity of data processing and model optimization increases
Solution Approach 1:
The patent segments the multivariate time series prediction problem into distinct modular components: data acquisition module, synchronization module, Bayesian imputation module, feature extraction module, and prediction module. Each module handles a specific aspect of the data processing pipeline, allowing complex multivariate dependencies to be managed through a series of simpler, specialized operations rather than a monolithic complex system
Solution Approach 2:
The system performs preliminary actions by pre-processing the multivariate time series data through synchronization, imputation, and feature extraction before the actual prediction occurs. This preliminary processing transforms raw complex data into structured, ready-to-predict formats, reducing the computational complexity during the critical prediction phase while maintaining the ability to capture complex dependencies
3Reliability
If frequent data collection is performed to improve prediction reliability, then more data points are available, but the data storage and processing requirements increase
Solution Approach 1:
The system implements periodic data collection and processing cycles rather than continuous high-frequency sampling. By collecting data at strategically determined intervals based on process stability, prediction confidence requirements, and resource availability, the system accumulates sufficient data for reliable predictions while avoiding the excessive storage and processing burdens of truly continuous high-frequency data acquisition
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure describes a method of controlling a manufacturing system using multivariate time series, the method comprising: recording data from one or more devices in the manufacturing system; storing the recorded data in a data storage as a plurality of time series, wherein each time series has a first recorded value corresponding to a first time and a final recorded value corresponding to an end of the time series; interpolating, within a first time window, missing values in the plurality of time series using a Bayesian model, wherein the missing values fall between the first and end time of the respective time series; storing the interpolated values as prediction data in a prediction storage, wherein the interpolated values include the uncertainty of each interpolated value; loading the recorded data that fall within a second time window from the data storage; loading prediction data from the prediction storage that fall within the second time window and for which no recorded data are available; optimizing the parameters of the Bayesian model using the loaded recorded data and the prediction data; predicting, using the Bayesian model, values for each of the time series for which loaded recorded and prediction data are not available; storing the predicted values as prediction data in the prediction storage, wherein the prediction values include the uncertainty of each prediction value; and adjusting one or more of the devices that generate the recorded data based on the prediction data within the second time window.