Anomaly Detection for Seasonal Sensor Drift in Equipment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting anomalous states in industrial plants are limited in handling slow progressive changes, high oscillations, unknown anomalous states, and sensor data with long seasonal trends, leading to inconsistent and unstable performance in predicting equipment behavior.
Innovation Solution
A computer-implemented method using machine learning models, specifically LSTM recurrent neural networks and Gaussian mixture models, to analyze sensor data, preprocess for oscillations and trends, and determine anomalous states by calculating divergence and variation from predicted behavior, with confidence scoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing detection methods are used for equipment with distinct static operating states, then detection is straightforward, but they fail to handle slow progressive changes, seasonal trends, high oscillations, and unknown anomalous states
Solution Approach 1:
The patent transforms the detection approach by changing parameters from static threshold-based detection to dynamic model-based prediction. It uses LSTM networks to learn temporal patterns and Gaussian mixture models to capture distribution characteristics, enabling reliable detection across diverse conditions including slow changes, seasonal trends, and oscillations that static methods cannot handle
Solution Approach 2:
The patent introduces dynamic elements by using recurrent neural networks that adapt to changing operating conditions over time. The system continuously learns from historical data and adjusts its detection thresholds and models dynamically, rather than relying on fixed static operating state definitions, thereby achieving both adaptability and reliability
2Measurement precision
If complex data driven algorithms are used to predict component behavior, then detection capability improves, but computational complexity and data requirements increase
Solution Approach 1:
The patent segments the complex detection task into two distinct components: an LSTM network for capturing temporal dependencies and predicting normal behavior, and a Gaussian mixture model for modeling the distribution of prediction errors. This segmentation allows each component to specialize in one aspect, improving overall precision while making the system more manageable and interpretable
Solution Approach 2:
The patent introduces an intermediary layer (the prediction model) that bridges raw sensor data and anomaly detection. Instead of directly detecting anomalies from raw data, the system first predicts normal behavior and then compares actual behavior against predictions, using the divergence as an intermediary signal for anomaly detection. This approach improves precision while maintaining reasonable complexity
Data Source
AI summary
Embodiments are directed to a computer-based tool that can identify an anomalous state of a component in a real-world environment, even if the component experiences gradual and/or seasonal trends. The tool receives data from sensors monitoring a component. The tool uses a trained machine learning model to calculate a predicted behavior of the monitored component. Actual behavior of the component, captured by current sensor readings, is compared to the predicted behavior of the component, calculated by the machine learning model, to compute a divergence. The computed divergence is used by a statistical learning method to determine if the component in the real-world environment is in an anomalous state.