Sensor Data Processing for Real-Time Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for analyzing dynamical systems using time-series data from sensors, such as accelerometers, require extensive offline dataset collection and heavy computational resources for training artificial neural networks, making them inefficient for real-time anomaly detection and resource-constrained environments.
Innovation Solution
A method that processes time-series data using a pipeline involving domain transformation, concept drift detection, normalization, and auto-encoder neural network processing with randomly initialized weights, allowing for online learning of normal operating conditions and reduced computational requirements, enabling efficient anomaly detection in dynamical systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ANN processing with extensive offline dataset collection is used, then anomaly detection accuracy is improved, but computational resources and time required for training increase significantly
Solution Approach 1:
The patent applies preliminary action by performing domain transformation and feature extraction offline before the actual anomaly detection runtime. The concept drift detector and normalization processes are prepared in advance, allowing the online anomaly detection to use pre-processed data without requiring heavy computational resources during real-time operation.
Solution Approach 2:
The patent segments the data processing pipeline into distinct stages: domain transformation, concept drift detection, normalization, and anomaly detection. This segmentation allows each component to be optimized independently and enables the system to use reduced computational power by processing data through simplified representations at each stage.
2Quantity of substance
If extensive offline dataset collection is performed, then training data sufficiency is improved, but data collection time and operational complexity increase
Solution Approach 1:
The patent implements self-service by enabling the system to automatically learn normal operating conditions from data streams without requiring manual intervention or pre-collected datasets. The concept drift detector automatically identifies when to update the normal condition model based on detected drifts, eliminating the need for time-consuming offline data collection procedures.
Solution Approach 2:
The patent applies dynamics by transitioning from static offline dataset collection to dynamic online learning. The system continuously updates its understanding of normal operating conditions as data streams in, adapting to changing system behavior without requiring predetermined training datasets. This dynamic approach eliminates fixed data collection time requirements.
3Measurement precision
If heavy computational power is allocated to ANN training, then processing accuracy is improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent replaces the heavy mechanical computation of traditional ANN training with a more efficient computational approach using domain transformation and simplified anomaly detection algorithms. By substituting complex neural network training with mathematical transformations and statistical methods, the system achieves comparable accuracy with significantly reduced computational power requirements.
Solution Approach 2:
The patent applies parameter changes by transforming the data representation through domain transformation and normalization processes that convert complex sensor data into simplified feature spaces. This transformation allows the anomaly detection to operate with reduced computational complexity while maintaining accuracy through the preserved semantic meaning of the transformed parameters.
Data Source
AI summary
An embodiment method comprises applying domain transformation processing to a time-series of signal samples, received from a sensor coupled to a dynamical system, to produce a dataset of transformed signal samples therefrom, buffering the transformed signal samples, obtaining a data buffer having transformed signal samples as entries, computing statistical parameters of the data buffer, producing a drift signal indicative of the evolution of the dynamical system as a function of the computed statistical parameters, selecting transformed signal samples buffered in the data buffer as a function of the drift signal, applying normalization processing to the buffered transformed signal samples, applying auto-encoder artificial neural network processing to a dataset of resealed signal samples, and producing a dataset of reconstructed signal samples and calculating an error of reconstruction. The error of reconstruction reaching or failing to reach a threshold value is indicative of the evolution of dynamical system over time.