Autoencoder Process Monitoring for Batch Deviation Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting deviations in batch processes, such as principal component analysis (PCA), struggle to identify individual drift parameters and small deviations in process parameters, leading to potential quality issues and inefficiencies.
Innovation Solution
A computer-implemented method using an autoencoder process model to detect deviations in batch processes by generating and training a model with reference process data, assigning model nodes to process parameters, and determining reconstruction errors to identify individual contributions outside predetermined value ranges.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If principal component analysis is used to detect deviations, then overall process deviation can be detected, but individual drift parameters cannot be identified
Solution Approach 1:
The patent segments the overall deviation detection into individual parameter contributions by using an autoencoder model where each neuron's weight contribution can be individually analyzed. This allows the system to identify which specific process parameters are causing deviations rather than just detecting that a deviation exists.
Solution Approach 2:
The patent transitions from traditional PCA's 2D projection space to a neural network weight space where individual parameter contributions can be measured. By analyzing the weight matrix W between input and hidden layers, the system adds a dimensional aspect that enables identification of individual parameter drifts.
2Device complexity
If principal component analysis is used, then complex statistical measurements can be performed, but single deviating process parameters cannot be detected
Solution Approach 1:
The patent introduces an autoencoder neural network as an intermediary between the raw process data and the deviation detection. This intermediary model, trained on reference data, enables the system to identify single deviating parameters by analyzing weight contributions while maintaining the statistical analysis capabilities of the original approach.
3Speed
If principal component analysis is calculated per time point, then real-time monitoring is achieved, but batch maturity differences cause detection issues
Solution Approach 1:
The patent performs preliminary training of the autoencoder model using reference batch data that captures the normal process variations across different batch maturities. This preliminary action enables the model to distinguish between normal maturity-related variations and actual deviations, improving reliability while maintaining real-time monitoring.
Data Source
AI summary
A computer-implemented method for detecting deviations in a production process with process parameters includes providing reference process data of a reference production process which comprises reference parameters of the production process, generating and training a process model with model nodes and corresponding model weights based on an autoencoder using the reference process data, at least partly assigning model nodes of the process model to process parameters of the production process, providing current process data from a current production process that comprises current parameters of the production process, ascertaining process deviations of the current process data using the process model by determining a reconstruction error and outputting the model weights of the model nodes, estimating the future curve of the reconstruction error, and checking whether the estimated future curve of the reconstruction error lies within a specified value range for the specified duration and if so continuing, otherwise outputting an alarm.