Monitoring of Complex Industrial Processes based on Self-Organizing Maps and Watershed Transformations Christian W. Frey 2012
Monitoring of Complex Industrial Processes based on Self-Organizing Maps and Watershed Transformations 1. Motivation 2. Monitoring of Industrial Processes 3. Self-Organizing Maps 4. Demonstrator System at IOSB 5. Industrial Application PROCMON 6. Summary 2012 2
1. Motivation Monitoring of complex Industrial Processes: Production processes in process industry are characterized by high complexity and system order Typically the production plants are equipped with a high number of sensors and actuators Distributed automation systems require a continuous and integral monitoring functionality to increase operational availability and cost effectiveness Online monitoring and diagnosis of the asset functionality Detection of process anomalies and component malfunctions Application-oriented visualization of the diagnosis information 3
2. Monitoring of Industrial Processes Model-Based Monitoring: Diagnosis of technical processes is performed by measuring the residual between the process model and the measured system states Analytical process models require a mathematical description of the process relations: Y (t) Comprehensive physical system knowledge is required and can be very expensive due to experimental investigations X (t) Process Model Y m(t) Highly complex processes e.g. chemical plants are difficult to describe by an analytical process model due to their high system order E (t) 4
2. Monitoring of Industrial Processes Model-Based Monitoring: Data-driven process modeling concepts can be applied to obtain model of the unknown system behavior: No mathematical description of the physical behavior Recorded input / output samples are used to induce an data-driven process model X (t) Y (t) Novelity Detection N (t) Machine learning algorithms are capable of modeling the input / output behavior based on training data For anomaly detection an multiple-input multiple-output model (MIMO) of the process is not required The measured system states only need to be classified with respect to their novelty 5
3. Self-Organizing Maps Topology Preserving Mapping: High dimensional input data is mapped topological ordered into a lower dimensional output space f : A n B, A R, B R m f f Similar input vectors are mapped topological close together and dissimilar apart Topology preserving concepts: Self Organizing Maps (SOM) Growing Grid / Neural Gas (GGRID) Generative Topographic Mapping (GTM) 6
3. Self-Organizing Maps Self-Organizing Maps (SOM): f f Self Organizing Maps developed by Teuvo Kohonen in 1989 The self-organizing map performs a topological mapping from a higher dimensional input space to a lower dimensional map space The map consists of neurons arranged in a defined topology (ring, toroid) No data connection between neurons the neurons are characterized by their position in the topology Each Neuron holds an prototype vector the dimensionality corresponds to the number of input signals 7
3. Self-Organizing Maps Self Organizing Maps (SOM): High dimensional input data is topological ordered by so called neighborhood function SOM forms a semantic map where similar input vectors are mapped topological close together and dissimilar apart Approximation quality of the map is calculated by the distance (e.g. Euclidian) between input vector and neuron vector (quantification error) d Input Vector Prototype Vector d l x, m x, m i j i m m,..., j 1 m jd x x,..., i 1 x id il jl 2 8
3. Self-Organizing Maps Unified-Distance-Matrix (UMatrix): UMatrix transformation adds a third dimension to the 2D map of the SOM which corresponds to the distance between the neighboring neurons Valleys (blue) in the UMatrix are clusters in the map where the stored prototype vectors are similar (specific process phases) Mountain Ridges (red) reflect the cluster boarders in the data set (transient phases) By analyzing the UMatrix: Detection of process phases and transient process phases BMU trajectory can be interpreted as process phase sequence 9
3. Self-Organizing Maps Unified-Distance-Matrix (UMatrix): UMatrix transformation adds a third dimension to the 2D map of the SOM which corresponds to the distance between the neighboring neurons Valleys (blue) in the UMatrix are clusters in the map where the stored prototype vectors are similar (specific process phases) Mountain Ridges (red) reflect the cluster boarders in the data set (transient phases) By analyzing the UMatrix: Detection of process phases and transient process phases BMU trajectory can be interpreted as process phase sequence 10
3. Self-Organizing Maps Unified-Distance-Matrix (UMatrix): UMatrix transformation adds a third dimension to the 2D map of the SOM which corresponds to the distance between the neighboring neurons Valleys (blue) in the UMatrix are clusters in the map where the stored prototype vectors are similar (specific process phases) Mountain Ridges (red) reflect the cluster boarders in the data set (transient phases) By analyzing the UMatrix: Detection of process phases and transient process phases BMU trajectory can be interpreted as process phase sequence 11
3. Self-Organizing Maps Watershed-Transformation : UMatrix representation holds information about the number of clusters and the cluster borders The Watershed-Transformation is capable of extracting the number of clusters and the cluster areas in the UMatrix The basic idea is to flood the UMatrix valleys the watershed lines corresponds to the cluster borders in the training data set By applying the Watershed Transformation: Process phases can be determined Process phase sequence can be monitored 12
4. Demonstrator System at IOSB 13
4. Demonstrator System at IOSB 14
4. Demonstrator System at IOSB 15
4. Demonstrator System at IOSB Prozessdaten Fehler 16
5. Industrial Application PROCMON Process and Condition Monitoring: Embedding of the monitoring functionalities in a software engine (PROCMON) Implementation in C++ with interfaces for: ANSI C++ Microsoft.NET3.5 /.NET4.0 32/64 BIT (C++/Cli, C#, VB) Matlab Implementation of standardized graphical user interface (GUI) for Microsoft operating systems (WIN32/64) 17
5. Industrial Application PROCMON Bayer Technology Services PuMon: Development of monitoring application for process industry in cooperation with Bayer Technology Services (BTS) Process-Unit-Monitoring (PUMon) functionality is optimized for monitoring of chemical and pharmaceutical production plants 18
5. Industrial Application PROCMON Monitoring of drinking water quality: 19
5. Industrial Application PROCMON Monitoring of wind power plants: P [MW] V [m/s] f1 [HZ] Error fn [HZ] 20
6. Summary Monitoring concept based on machine learning methods Appliance of self organizing maps for generating an data driven model of the physical process behavior Watershed algorithm for detecting automatically the typical process phases Verification of the concept by appliance to an chemical test plant Appliance of the diagnosis concept to a wide range of processes in process industry 21