Sparse Neural Networks for Multi-Sensor Time-Series Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing anomaly detection methods for multi-dimensional time series are ineffective due to the need for domain-driven sensor selection, loss of sensor-dependency information, and impracticality of building separate models for each sensor, especially in complex industrial systems.
Innovation Solution
A processor-implemented method using a sparse neural network that combines dimensionality reduction via LASSO-constrained feedforward layers with a recurrent neural network encoder-decoder model to learn normal behavior and detect anomalies by computing error vectors and generating anomaly scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If domain-driven sensor selection is used for anomaly detection, then the detection accuracy for known anomalies is improved, but other types of anomalous signatures in excluded sensors are missed
Solution Approach 1:
The patent applies universality by designing a single RNN-based anomaly detection model that can process all sensor inputs simultaneously without requiring domain-driven sensor selection. The model is trained to learn from the temporal patterns and dependencies across all sensors, enabling it to detect various types of anomalies including those that may be present in previously excluded sensors, thus achieving both accuracy and comprehensive coverage
2Adaptability or versatility
If separate RNN models are built for each sensor to handle all anomaly types, then detection coverage is improved, but computational complexity and practical deployment become infeasible
Solution Approach 1:
The patent merges multiple sensor processing functions into a single unified RNN model. Instead of building separate models for each sensor, the invention combines all sensor inputs into one model that processes them jointly, learning the temporal dependencies and interactions between sensors. This significantly reduces computational complexity and deployment difficulty while maintaining comprehensive anomaly detection coverage
3Adaptability or versatility
If separate RNN models are built for each sensor, then all anomaly types can be detected, but sensor-dependency information is lost
Solution Approach 1:
The patent uses the RNN's internal state representations as copies of sensor information that preserve dependency relationships. The model creates hidden state representations at each time step that encode the temporal dependencies between sensors, allowing the system to maintain and utilize sensor-dependency information without requiring separate models for each sensor
Data Source
AI summary
Anomaly detection from time series is one of the key components in automated monitoring of one or more entities. Domain-driven sensor selection for anomaly detection is restricted by knowledge of important sensors to capture only a certain set of anomalies from the entire set of possible anomalies. Hence, existing anomaly detection approaches are not very effective for multi-dimensional time series. Embodiments of the present disclosure depict sparse neural network for anomaly detection in multi-dimensional time series (MDTS) corresponding to a plurality of parameters of entities. A reduced-dimensional time series is obtained from the MDTS via an at least one feedforward layer by using a dimensionality reduction model. The dimensionality reduction model and recurrent neural network (RNN) encoder-decoder model are simultaneously learned to obtain a multi-layered sparse neural network. A plurality of error vectors corresponding to at least one time instance of the MDTS is computed to obtain an anomaly score.