Neural Anomaly Detection From Aggregate Sensor Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In electronic device manufacturing, existing monitoring systems rely heavily on human operators to interpret data from multiple sensors, leading to increased costs, subjective decision-making, and potential oversight of anomalies, especially in complex processes like chemical vapor deposition and physical vapor deposition, where simultaneous changes in sensor readings may indicate deteriorating conditions.
Innovation Solution
The implementation of a method using neural networks to process and analyze aggregated sensor statistics, reducing data dimensionality, applying anomaly detection models, and generating anomaly scores to automatically detect manufacturing process deviations, thereby reducing human intervention and improving efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If human operators interpret data from multiple sensors, then subjective decision-making and potential oversight of anomalies occur, but extensive human oversight is required
Solution Approach 1:
The system enables self-service by implementing automated anomaly detection through neural networks that process sensor statistics independently. The detector neural network automatically generates anomaly scores from aggregated sensor data without requiring human operators to interpret the data, allowing the system to monitor itself and detect manufacturing anomalies autonomously.
Solution Approach 2:
The patent replaces the mechanical human decision-making process with an automated neural network system. Instead of human operators subjectively interpreting sensor data, the detector neural network processes aggregated sensor statistics and generates objective anomaly scores, substituting human cognitive functions with automated computational mechanisms.
2Measurement precision
If data from multiple sensors is aggregated and analyzed using neural networks, then anomaly detection accuracy improves, but system complexity increases
Solution Approach 1:
The system segments the complex anomaly detection task into distinct functional components: an aggregator neural network that processes individual sensor statistics separately, and a detector neural network that integrates these processed statistics. This segmentation allows each component to specialize in specific aspects of data processing, improving overall detection precision while managing system complexity through modular architecture.
Solution Approach 2:
The patent introduces an intermediary reduced representation that bridges the gap between raw sensor statistics and final anomaly detection. The aggregator neural network transforms complex multi-sensor data into a simplified reduced representation, which then serves as input to the detector neural network. This intermediary layer reduces data dimensionality and complexity while preserving essential anomaly information.
3Productivity
If reduced representation of sensor statistics is used, then processing efficiency improves, but information loss may occur
Solution Approach 1:
The system applies parameter changes by transforming the original sensor statistics through the aggregator neural network into a different parameter space - the reduced representation. This transformation changes the form and dimensionality of the data parameters, enabling more efficient processing by the detector neural network while the trained model preserves the essential information needed for accurate anomaly detection.
Data Source
AI summary
Implementations disclosed describe a method and a system to perform the method of obtaining a reduced representation of a plurality of sensor statistics representative of data collected by a plurality of sensors associated with a device manufacturing system performing a manufacturing operation. The method further includes generating, using a plurality of outlier detection models, a plurality of outlier scores, each of the plurality of outlier scores generated based on the reduced representation of the plurality of sensor statistics using a respective one of the plurality of outlier detection models. The method further includes processing the plurality of outlier scores using a detector neural network to generate an anomaly score indicative of a likelihood of an anomaly associated with the manufacturing operation.