Self-Organizing Map Anomaly Detection Using Temporal Node Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting process anomalies in technical installations using self-organizing maps often fail to recognize deviations early enough or accurately determine their cause, especially in dynamic processes, due to similarities between abnormal and normal states, and lack temporal behavior detection.
Innovation Solution
A method that trains a self-organizing map with historical data to determine chronological node paths and tolerance, sets threshold values for Euclidean distance, and tracks node sequences to identify deviations and determine symptom vectors, enabling early and reliable anomaly detection and cause analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If self-organizing maps are used to classify system statuses, then fault statuses can be recognized, but deviations are not recognized early enough or are missed entirely when abnormal states resemble normal states
Solution Approach 1:
The patent transitions from static status classification to dynamic temporal sequence analysis. By examining the chronological sequences of node activations and comparing them against learned normal patterns, the system detects deviations in the temporal dimension rather than relying solely on spatial classification of individual states. This allows early detection of anomalies even when individual states appear normal.
Solution Approach 2:
The system pre-learns and stores chronological sequences of node hits for normal operating conditions during a training phase. These reference sequences are established beforehand and used to compare against actual runtime sequences, enabling early anomaly detection by identifying deviations from the pre-established normal patterns before critical failures occur.
2Reliability
If traditional self-organizing maps are used for monitoring, then system states can be classified, but reliable cause determination is not possible due to incorrect symptom vector determination
Solution Approach 1:
The system uses the chronological sequence of node activations as feedback to identify the root cause. By tracking the temporal order in which nodes are hit and comparing this sequence against learned normal patterns, the system can determine which nodes deviate from expected behavior and use this information to accurately construct symptom vectors that point to the actual cause of anomalies.
3Adaptability or versatility
If static self-organizing maps are used for state classification, then deviations can be identified, but temporal behavior deviations cannot be detected
Solution Approach 1:
The patent transforms the static self-organizing map into a dynamic monitoring system by tracking the temporal sequences of node activations. Instead of merely classifying individual states, the system monitors how nodes are activated over time and detects anomalies in the temporal patterns. This dynamic approach adds temporal behavior detection capability while building upon the existing self-organizing map structure.
Data Source
Figure 1
Figure 2
AI summary
The invention essentially relates to a method and a diagnostic system for improved detection of a process anomaly in a technical plant, in which a self-organizing map is first trained using the historical process data as good states of the plant, wherein the good states are used to determine the temporal sequence ora path of the hit nodes and the tolerances of the neuron hits are used, and thresholds for the Euclidean distance for the good states are determined and stored, and the current process data of the plant in the form of a state vector are evaluated using the trained self-organizing map, the Euclidean distance of the current state vector to the hit neuron is checked for exceeding the threshold, a neuron that should be hit is determined using the determined path, provided the threshold has not already been exceeded during the check with the hit neuron, and a symptom vector is determined from the current state vector and either the hit neuron or the neuron that should be hit.If the number of hits for a node is optionally stored along with the paths, deviations from the normal state over time can be determined. The advantages lie primarily in the early and reliable detection of deviations from the norm, the additional possibility of accurately determining a symptom vector for better identifying the root cause of the error.