Koichi Inoue’s research while affiliated with Osaka Sangyo University and other places

Conventional fault trees composed of AND and OR logic gates do not consider event sequence dependency. This paper presents a simple evaluation method for fault trees containing AND and OR gates plus priority AND gates, which restrict the occurrence order of basic events. Basic logical operations of priority AND gates with conventional logic gates are given. Due to these simple operations, not only the minimal cut set derivation procedure, but also the inclusion-exclusion method can be applied in the same way as coherent fault trees. An illustrative example of a standby system shows the merits and characteristics of the proposed method.

A minimal cut set (MCS) of FTA (fault tree analysis) shows a minimal combination of component failures leading to system failure, based on which system failure probability can be evaluated. However, the conventional inclusion-exclusion method based on MCSs does not necessarily give the system failure occurrence probability. This paper proposes an analytical probability evaluation method of system failure occurrence based on critical states for each basic event. System failure occurrence is caused by a basic event if it occurs at its critical state. Critical states for a basic event can be obtained using logical expression in terms of MCS's. Thus, the system failure occurrence probability can be obtained as the sum of system failure occurrence probabilities caused by basic events. The proposed method assumes any general distribution for component failure probability. An illustrative example shows the details and merits of the method

Severe accidents can occur even in advanced systems such as airplanes and nuclear power plants. In the accident analysis, most of the attention focuses on the active failure which initiates the occurrence of system failure, but the latent condition which allows the active failure to cause system failure must be considered. This paper proposes a simple method to deal with latent conditions in FTA (fault tree analysis) based on the detectability of basic events. In FTA analysis, the system failure occurs if all basic events in a minimal cut set occur, but all the event sequences leading to the occurrence of a minimal cut set do not always occur without their detection. Using the detectability index of a basic event, it is evaluated whether the critical state of a basic event is latent where its occurrence can cause the system failure. Based on the number/probability of critical states that can be detected by the detectability of a basic event, a new measure of its importance in contributing to the reduction of system failure is proposed. A simple illustrative example shows these properties

This paper introduces the basic concept of Probalilistic Risk Assessment (PRA) , which evaluates the safety of large-scale systems. Firstly, the use of probability concept in the design of technical system is briefly reviewed, and then the concept of risk is defined. Then, PRA procedure is reviewed which is composed of risk analysis and risk assessment. Risk analysis obtains accident sequences and estimates their probabilities and effects using Event Tree Analysis and Fault Tree Analysis. Risk assessment examines whether the risk analysys result satisfies the risk acceptance level. If the risk acceptance level is not achieved, some measure must be taken to reduce the risk. PRA continues until the risk acceptance is achieved. Human plays an important role not only in the safety assessment, but also in accident sequences leading to system failure. Finally, the human-related problems in PRA are discussed from three viewpoints: the definition and evaluation of probability, the risk acceptance, and human error analysis.

This paper proposes two efficient methods to design a robust feedback control system by use of neural networks. The first method is based on L2 gain, and two different neural networks are used. The controller is trained to be robust as a result of competition between neural networks. The second method is based on MiniMax optimization, and is useful to treat parametric uncertainties. In both methods, robustness of the neural network can be quantified. It is very easy to combine proposed methods so that effective methods for various problems can be derived.

The interaction between/among operators must be considered in research domains such as a task analysis or human modeling of a teamwork operation. This paper shows that the interaction between/among operators can be represented by three kinds of "Mental Images" and operator's cognitive activity can be analyzed by using these mental images. Furthermore, it also shows (1) how mental images are formed in an operator's mind and (2) what kind of action would be taken when he/she senses a discrepancy between mental images. Finally, an illustrative example of an airplane accident shows how to make use of mental images in human error analysis.

Failure diagnosis is generally composed of three activities: (1) failure detection, (2) cause identification, and (3) protective action. This paper proposes the use of bond graphs with a means-ends hierarchy structure for the support of the entire failure diagnosis process. Bond graphs represent the system behavior from the viewpoint of energy flow, which makes it easy to model not only physical phenomena or components, but also abnormal events. Though the relations among state variables represented by the system bond graph can support the detection of an abnormal event and the identification of its cause, they cannot support the evaluation of its effect on the system function explicitly. Using means-ends hierarchy structure, the diagnosis result in the state variable space is transformed into the system functional space, which can help the system analysts select an preventive action and predict its effect. An illustrative example of a water flow control system shows the details and merits of the proposed method.

This paper proposes a generic design method to develop an optimal state feed-back control system using multilayered neural networks. Because Back-Propagation can not be applied to this case, two new methods are proposed for training neural networks: one is based on gradient method and the other on Powell's conjugate direction method. These two methods can be applied to various problems such as ones for non-linear systems where the feed-back controller can not be obtained by conventional methods. Especially, the method based on the direction method can deal with criterion functions whose derivatives can not be obtained analytically. Illustrative examples of optimal regulator and minimal time control problems show the effectiveness of the proposed methods.

To prevent the effect of human errors in a team of multiple members who execute a given task, how they take actions to achieve it must be clarified. This paper proposes a framework for qualitative analysis of a team performance based on a human behavior model. Each member goes around a hierarchical problem-solving structure to achieve his current task goal. The effectiveness of a team performance depends on a kind of collaboration, communication. A communication is assumed to occur when a member needs help from others to achieve his current goal. An illustrative example of a tank filling task show the representation of a team performance model and its simulation result.