Machine Fault Detection via Virtual Sensors for Unobserved Wear
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for error detection in mechanical systems, such as hydraulically driven machines, are inefficient in detecting anomalies in unobserved quantities that cannot be directly measured, leading to potential failures due to lack of visibility into system components like friction and leakage, which are not accessible during regular operation.
Innovation Solution
An iterative estimation method combining physical models with sensors and virtual sensors to estimate unobserved quantities, allowing for early detection of anomalies by analyzing deviations in wear parameters using distance measures like Mahalanobis distance, implemented in local computing units like PLCs, reducing computational effort and enabling continuous error detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning-based models are used for error detection, then detection capability is improved, but computational effort and complexity increase significantly
Solution Approach 1:
The patent introduces a physical model as an intermediary between sensor measurements and error detection. This physical model serves as a simplified mediator that translates complex system behavior into manageable relationships, avoiding the need for computationally intensive machine learning models while maintaining effective error detection capability.
Solution Approach 2:
The patent replaces complex computational systems (machine learning models) with a simpler physical model-based approach. By substituting the mechanical/computational complexity of neural networks with a more straightforward physical model that processes sensor data, the system achieves error detection with minimal computational effort.
2Loss of information
If sensors are installed to measure all system parameters, then measurement completeness is improved, but system cost and complexity increase
Solution Approach 1:
The patent creates virtual copies of physical sensors through mathematical models. Instead of installing physical sensors to measure unobserved variables like friction and leakage, the system computes virtual sensor readings from the physical model and available measurements, providing complete system state visibility without additional physical hardware.
Solution Approach 2:
The physical model acts as an intermediary that bridges the gap between limited sensor measurements and complete system state knowledge. It mediates between the available sensor data and the unmeasured system parameters, deriving information about friction, leakage, and wear without requiring direct physical measurement of these quantities.
3Device complexity
If physical models are used for estimation, then computational effort is reduced, but ability to detect anomalies in unobserved variables is limited
Solution Approach 1:
The patent creates virtual sensor copies that specifically target unobserved variables like friction and leakage. These virtual sensors are mathematical constructs derived from the physical model that provide precise estimates of parameters that would be difficult or impossible to measure directly, enabling anomaly detection in previously inaccessible system aspects.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method for fault detection in a machine system, comprising acquiring (110) measured values for one or more measured variables of the system; performing (120) an iterative estimation method (42, 46, 50) using the measured values to determine an estimate for a system state that includes at least one unobserved variable (22, 24, 26, 28, 44, 48, 52) of the system, wherein the iterative estimation method is based on a physical model of the system, and wherein the model includes at least one interaction of the at least one unobserved variable with the one or more measured variables; and applying (140) anomaly detection, wherein a fault of the system is detected (160) when an anomaly of the at least one unobserved variable is detected.