LSTM Sensor Drift Identification via Autocorrelation Thresholds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current abnormal event detection (AED) processes in the petrochemical industry, relying on statistical criteria and linear models, struggle to accurately distinguish between sensor drift and true abnormal events, leading to false alarms and potential equipment failures.
Innovation Solution
The implementation of a long-short term memory (LSTM) model combined with an accumulated slow drift error (ASDE) model, utilizing deep neural network machine learning and graphics processing unit (GPU) processing, to identify sensor drift by learning the natural rhythm of process variables and setting autocorrelation thresholds based on historical data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If statistical criteria and linear models (PCA, PLS) are used for abnormal event detection, then the detection process is simple and computationally efficient, but the system produces false alarms when sensors drift because it cannot distinguish drift from actual abnormal events
Solution Approach 1:
The patent transforms the detection approach by changing from linear statistical parameters to nonlinear machine learning parameters. The LSTM model learns temporal patterns and autocorrelation characteristics of sensor data, using parameters like autocorrelation thresholds and drift detection metrics that adapt to nonlinear sensor behavior, enabling distinction between drift and actual abnormalities
Solution Approach 2:
The patent replaces traditional statistical mechanical models (PCA, PLS) with intelligent computational systems (LSTM neural networks, GPU processing). This substitution enables the system to handle nonlinear relationships and temporal dependencies that linear models cannot capture, significantly reducing false alarms from sensor drift
2Measurement precision
If the AED process uses multiple variables and linear models to predict sensor readings, then the system can detect abnormalities through prediction mismatch, but it fails when the relationship between variables is nonlinear
Solution Approach 1:
The patent introduces dynamic adaptability through the LSTM model that continuously learns from historical sensor data. The system dynamically adjusts its understanding of normal sensor behavior patterns, allowing it to adapt to nonlinear relationships between process variables without requiring explicit mathematical models of those relationships
Solution Approach 2:
The system changes from using fixed linear relationship parameters to dynamic learned parameters through machine training. The LSTM model captures complex nonlinear parameter relationships through its internal state mechanisms, enabling accurate prediction even when variable relationships change over time or are inherently nonlinear
3Reliability
If sensor drift is not distinguished from abnormal events, then all deviations trigger alarms, but this leads to false alarms and unnecessary interventions
Solution Approach 1:
The patent segments the detection space by introducing separate detection pathways: one for detecting sensor drift (using autocorrelation analysis of prediction errors) and another for detecting actual abnormal events. This segmentation allows the system to classify different types of deviations appropriately, reducing false alarms while maintaining sensitivity to real problems
Solution Approach 2:
The patent introduces an intermediary drift detection layer between raw sensor data and final abnormal event classification. The LSTM model acts as an intermediary that first predicts normal sensor behavior, then analyzes prediction errors through autocorrelation to determine if drift is occurring before triggering abnormal event alarms, thereby filtering out false positives
Data Source
AI summary
A method for identifying sensor drift can include: setting an autocorrelation threshold for a sensor in a long-short term memory (LSTM) model developed based on historical process measurements from an analogous sensor to a sensor; collecting measured data from the sensor; applying the LSTM model to the measured data from the sensor, wherein applying the LSTM model comprises: applying the LSTM model to the measured data from the sensor to yield LSTM predicted data; calculating key performance indicators (KPIs) of the LSTM data based on an accumulated slow drift error (ASDE) model, wherein the KPIs comprise an error, an accumulated prediction error, an accumulated slow-drift error, and an estimated autocorrelation; and identifying sensor drift when the estimated autocorrelation violates the autocorrelation threshold.