Koji Komatsu

Classification-based behavior model for detection of abnormal states in systems

Abstracf - In !his paper, a special model in the form of lookup Table is proposed for detecting different abnormal states in mechanical and other industrial systems. The model is generated by repetitive use of a classification procedure that uses a SelfOrganized Map and Ndural Gas algorithm for competitive unsupervised learning. This classification procedure makes a hind of mapping from the measured “Parameter-Space” to the preliminary defined “Operating Mode-Space”. The computed Modes Recognition Vectors show the relative weight of each operating mode in the entire set of operating modes. They represent the first part of the system behavior model, while the sewad part consists of the vectors of the Deviated (abnormal) Parameters. These are preliminary generated in a systematical, combinatorial way. Thus produced Classification-based Behavior Model (CBB Model) is further used for solving the general problem of the fault diagnosis. Here the most plausible deviated parameters vector has to be found which produces as close as possible the computed Modes Recognition Vector. Simulations with real dbta from a hydraulic excavator are used in order to prove the applicability and the merits of the proposed method for abnormality detection and fault diagnosis.

Learning of RBF network models for prediction of unmeasured parameters by use of rules extraction algorithm

The paper presents three different methods for learning of normalized RBF network models that are similar in structure to the Takagi-Sugeno fuzzy models. These methods use different groups of parameters for optimization and incorporate a rules extraction algorithm for numerical evaluation of the connection weights, as a part of the optimization. Combinations of the methods give different learning strategies, which are analyzed in the paper through two simulated and one real example.

Identification of Dynamic Cause-Effect Relations for Systems Performance Evaluation

Incremental cause-effect relations are used in this paper in order to create respective CER dynamic models that are further used to reveal the existing relationships between the pairs of parameters in the observed set of system parameters.

Solving the inverse dynamics problem by self-organizing maps

In this paper a recursive computation procedure for recovering the inputs of a dynamic process based on a preliminary assumed number of measured discrete outputs is proposed and analyzed. A specially constructed self-organizing map is first trained in off-line mode and is further used in real-time as a tool for classification and revealing the existing input and the measured outputs. The proposed computation procedure gives a discrete solution of the inverse problem constrained within the preliminary assumed discrete levels of the input. The number of these levels is directly connected to the final computation accuracy. The consistency of the proposed computation scheme for solving the inverse problem is extensively analyzed on a test dynamic process. Simulation results show its applicability for solving different backward tracking problems, including fault diagnosis problems that heavily rely on a robust and plausible solution of the inverse problem.

Detection and Analysis of Deterioration Trends in Construction Machines Operation

The paper proposes a computational strategy for discovering trends of changes and gradual performance deterioration of construction machines and other complex systems. If a significant change in the machine performance is detected, it further activates a fault diagnosis procedure in order to find the possible faulty condition. In order to send the information from various sensors of the construction machine to the maintenance center in an efficient form, a special information compression method is proposed in the paper. It uses original unsupervised learning algorithm to locate the given number of neurons from the parameter space in the densest data areas. These neurons are considered as information granules, which are later on sent in a wireless way to the maintenance center to represent the current machine operation. Two information recovery procedures are also proposed in the paper for analyzing the compressed information from the neurons. The first one is a specialized version of the moving window method, while the other is fuzzy inference-based approach for discovering possible machine deterioration. Results based on real experimental data from a hydraulic excavator are used to explain the proposed computational strategy and its merits.