Industrial Process Anomaly Detection Using Dynamic Data Clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Industrial machines generate vast amounts of data, but existing methods require predefined notions of 'normal' behavior to detect anomalies, which can be cumbersome and inefficient, especially when dealing with complex multi-dimensional data sets.
Innovation Solution
A method that uses clustering to identify anomalies by determining the distance of new data points to existing clusters, allowing for automatic detection without predefining normal behavior, using user-supplied variables for cluster formation and outlier identification, and enabling notification and processing of anomalies for improved supervision and control.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional anomaly detection methods are used that require predefined normal behavior, then detection accuracy may be maintained, but system complexity and setup time increase significantly
Solution Approach 1:
The system performs self-learning by automatically analyzing historical process data to establish normal behavior patterns through clustering algorithms. The anomaly detection mechanism serves itself by autonomously identifying clusters and determining thresholds without external intervention, eliminating the need for manual configuration of normal behavior definitions while maintaining high detection accuracy
Solution Approach 2:
The system performs preliminary learning during an initial phase where historical data is collected and analyzed to pre-establish clusters and thresholds before actual anomaly detection begins. This preliminary action prepares the system in advance, so that when operational mode is activated, anomaly detection can proceed immediately without requiring manual setup of normal behavior parameters
2Reliability
If continuous monitoring of all process data is performed, then complete anomaly detection coverage is achieved, but processing time and computational resources increase
Solution Approach 1:
The continuous data stream is segmented into discrete clusters based on similarity metrics. Instead of analyzing every data point continuously, the system groups similar acquisitions into clusters and only performs detailed anomaly analysis when data points fall outside established cluster boundaries, significantly reducing processing time while maintaining comprehensive coverage
Solution Approach 2:
The system implements periodic cluster updates rather than continuous re-analysis of all data. Clusters are established during an initial learning phase and then updated periodically with new data, allowing the system to maintain reliable anomaly detection coverage while avoiding the computational burden of continuous full-data processing
3Measurement precision
If manual definition of normal behavior is required, then detection rules can be precise, but adaptability to new process conditions decreases
Solution Approach 1:
The system transitions from static manual definitions to dynamic automatic learning. Clusters and thresholds are not fixed but are continuously adapted based on incoming process data. The system can dynamically adjust to new process conditions by automatically updating cluster structures and threshold values, maintaining both precision and adaptability simultaneously
Solution Approach 2:
The system automatically adjusts detection parameters such as cluster thresholds, distance metrics, and similarity criteria based on the characteristics of the process data being analyzed. These parameter changes enable the system to adapt to different process conditions and variations while maintaining precise anomaly detection through data-driven optimization rather than fixed manual settings
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Present invention concerns a method and a system for monitoring data of an industrial process. The method comprises steps of acquiring a parameter related to the industrial process; determining the distance of said acquisition to a cluster; and determining that said acquisition represents an anomaly if its distance to the cluster exceeds a predetermined threshold. In this, the cluster comprises at least a predetermined number of acquisitions that are closer to each other than a predetermined distance and the distance between the acquisition and the cluster is the shortest distance between said acquisition and any acquisition of the cluster.