Siamese Auto-Encoder Anomaly Detection for Industrial Time-Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex industrial systems face challenges in accurately detecting anomalies due to the complexity of interdependencies between components, leading to difficulties in distinguishing between regular and anomalous operating states, resulting in either missed anomalies or excessive false alarms, which can overwhelm operators and reduce system availability.
Innovation Solution
A semi-supervised deep anomaly detection method using digital twin simulation data, where a Siamese auto-encoder neural network is trained with triples of data samples from both digital twin simulation and sensor data to differentiate between regular and anomalous operating states, achieving high detection performance with minimal human intervention and resource effort.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If rule-based monitoring systems are used with conservative rules, then false alarms are reduced, but detection sensitivity decreases and anomalies may remain undetected
Solution Approach 1:
The patent replaces rule-based monitoring systems with a machine learning model (autoencoder neural network) that automatically learns patterns from data. The model substitutes manual rule formulation with automated anomaly detection based on learned representations, achieving both high sensitivity and low false positives through data-driven pattern recognition rather than conservative threshold rules
Solution Approach 2:
The patent transforms the monitoring approach by changing from fixed rule-based parameters to dynamic learned parameters. The autoencoder learns optimal detection thresholds and patterns during training, allowing the system to adapt parameters based on actual system behavior rather than relying on pre-defined conservative rules, thereby improving both detection sensitivity and reliability
2Measurement precision
If rule-based monitoring systems implement comprehensive checks, then detection sensitivity improves, but false alarms increase and operators are overwhelmed
Solution Approach 1:
The patent replaces manual rule formulation with automated machine learning-based anomaly detection. The autoencoder learns to distinguish normal from anomalous patterns through training, automatically optimizing the balance between detection sensitivity and false alarm rate without requiring operators to manually tune numerous rules
Solution Approach 2:
The system performs self-learning and self-optimization through the autoencoder training process. The model automatically adjusts its internal parameters and detection thresholds based on training data, eliminating the need for continuous manual rule tuning and reducing operator workload while maintaining high detection accuracy
3Measurement precision
If expert knowledge is used to define feasible ranges, then detection accuracy improves for simple systems, but system complexity increases for complex technical systems with many interrelated components
Solution Approach 1:
The patent replaces complex manual rule formulation with automated machine learning. The autoencoder automatically learns to represent complex interdependencies in the system data, substituting the need for experts to manually define rules for each component interaction with a unified data-driven model that handles complexity internally
Solution Approach 2:
The patent segments the complex monitoring task into distinct phases: data collection, autoencoder training on normal operation data, and anomaly detection. This segmentation allows the system to handle complexity by breaking down the overall task into manageable steps, with the model learning complex patterns during training and applying them during inference
4Ease of manufacture
If digital twin simulation data is used for training, then the need for labeled anomaly data is reduced, but the complexity of integrating simulation and real data increases
Solution Approach 1:
The patent uses digital twin simulation to create copies of normal operating conditions. The autoencoder is trained on these simulated copies of normal data, eliminating the need to collect extensive real-world anomaly data. The simulation copies provide sufficient training examples for the model to learn normal patterns without requiring actual anomaly instances
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for detecting an anomalous operating status of a technical system with a training phase and a monitoring phase. The training phase obtains a first set of time-series values generated by a digital twin simulation of the technical system for a regular operating status of the technical system and a second set of time-series values measured by a plurality of sensors. The sensors are configured to monitor a set of operational parameters of the technical system and collect the second set of time-series values in an anomalous operating status of the technical system. The method then executes a training step by adjusting parameters of a machine learning model for detecting the regular operating status of the technical system and for discriminating data samples of the regular operating status from data samples of the anomalous operating status by processing triples of data samples, the each of the triples of data samples comprising a first data sample and a second data sample each from the first set of time-series values, and a third data sample from the second set of time-series values to generate a trained machine learning model. In a monitoring phase, the method obtains a set of multivariate time-series values measured by the plurality of sensors, calculates an anomaly score value adapted for determining whether the technical system is in an anomalous operating status based on the obtained set of multi-variate time-series values and the trained machine learning model, and then generates and outputs a signal including information on the determined anomalous operating status of the technical system.