The paper investigates a methodology on knowledge compilation and refinement after its acquisition through a process of car diagnostics on the case of abnormal noise and vibration. The knowledge compilation and refinement involve three basic steps; making knowledge hierarchic, finding useful features of knowledge, and designing a simpler decision tree. The better the useful features of knowledge are found, the better the decision tree is designed. This led to effective use of the domain knowledge obtained in cooperation with experts. In order to facilitate the actual fault diagnosis, the decision tree was designed by solving a combinatorial problem to minimize the number of total branches on each primary subset of classification. As a result of the knowledge compilation and refinement, the number of queries on the fault diagnosis was reduced by 83 to 88 percent and the error rate of diagnosis was 10 percent in comparison with the judgement of experts.
For practical pattern recognition, it is required not only to recognize geometrically similar patterns but also to detect the difference of translation, rotation and size from their templates. This paper proposes a method to recognize the similar patterns by a multilayer net and then detect the difference on the common basis of well-known geometrical characteristics (center of gravity, angle of principal axis, and variance). It is found from experimental results that, with the proposed method, a small net can classify the similar patterns at a high recognition rate and detect their rotated angles and scale ratios with a high accuracy.
This paper describes a method of generating the real-time fault detection model for manufacturing systems controlled by PC (Programmable Controller) automatically using Complementary-places Petri nets. To generate a fault detection model, we use two kinds of information : logging data of limit switches and ladder programs. By using the generated model, we can detect failures in the PC based manufacturing systems. Furthermore, we evaluate the method proposed in this paper by applying it to fault detection of a virtual production system.
This paper deals with a scheduling problem in a real electric wire production process. The process is of a flow-shop with two processes. In each process plural equipments are arranged in parallel. The scheduling problem is large-scale and has complicated constraints. A genetic algorithm (GA) is applied to the problem. If GA is designed so as to search all solutions in the fundamental space, it is inefficient because the number of such solutions is enormous. On the other hand some dispatching rules can be applied to scheduling problems if they are effective. However it is not easy to find an effective rule for complicated production processes. In this paper, GA is designed in such a way that appropriate dispatching rules are combined with the conventional GA operator. Consequently, a part of decision variables are determined by appropriate rules and the others are determined by the operation of GA. The effectiveness of the proposed method is confirmed by computational results.
We consider a polarimetric SAR data classification method which uses scattering models. The proposed method is a neural network classifier composed of two classification procedures. First, SAR data is pre-classified into three scattering classes (odd, even, and diffuse & unable to classify) by individually computing the Mueller matrix and Stokes vector based on the van Zyl approach. Second, a neural network which suits to each scattering class is constructed and the data is identified as a final land cover type. Either the competitive or back-propagation neural network can be employed as a classifier. It is possible to identify more detail category in order to analyze scattering classes. As a result of the procedure using SIR-C C band data, 11 land cover types by combination of five categories and three scattering classes are identified. We assume the effectiveness of the proposed method in comparison to classification by checking the ground truth data.
This paper proposes a technique to derive candidates of counter actions applying both functional modeling and qualitative reasoning methods with the purpose of supporting operator actions in an anomalous situation of engineering plants. A plant model is constructed by adding necessary knowledge to derive counter actions to a functional model of a target system by the MFM (Multilevel Flow Modelling). Algorithms are given to infer the affection of an anomaly, to determine the priority of plant states and/or goals to be recovered, and to derive candidates of counter actions. The applicability of the proposed technique is discussed using an anomalous case in a fluidized-bed refuse incineration plant.
For the purpose of implementing the coordinated traffic signal control, urban road networks in Japan are divided into several sub-areas. Each sub-area consists of a series of signal intersections. The traffic signal timing is optimized within each sub-area and reflects the local traffic situation. Since there is no well-established method of decomposing a traffic network into sub-areas, the decomposition is determined by skillful engineers based on their experience. This paper proposes a new method of decomposition, which is based on an effectiveness index reflecting the traffic delay and the topological structure of the sub-area network. To evaluate the performance of our approach, simulation experiments are carried out for an actual road network. The results show that the decomposition obtained by our method outperforms the conventional decomposition in terms of the employed performance index.