Failure Diagnosis Using Bayesian Networks for Multifunction Devices
When a customer requests the repair of trouble that occurred on their multifunction device, our customer engineer will visit the customer to repair the device. The time taken to repair a device differs depending on the experience and skills of each customer engineer. This difference in time is due to the fact that some engineers take the time to search the troubleshooting database and service manuals, while others don't. Furthermore, it is difficult to identify the cause of trouble in a short time if the trouble is related to multiple parts of a device.
To work on such problems, Fuji Xerox has developed failure diagnosis technology that uses Bayesian networks. With this technology, failure diagnosis equivalent to that performed using the accumulated knowledge of experienced customer engineers has become possible. A Bayesian network illustrates the relationship between cause and effect using a directed acyclic graph (a directed graph that does not contain cycles). As shown in Fig. 1, a directed acyclic graph (DAG) consists of a conditional probability table and nodes (shown as circles in Fig. 1) connected by arrows that show their causal relationships. A Bayesian network is a probability model that can be used for such purposes as estimating events that contain uncertainties, rational decision-making, and diagnosing failure. As the causes of failure are shown with probability in failure diagnosis that uses Bayesian networks, even an inexperienced customer engineer can identify and start working on the most probable cause of failure, thereby enabling the engineer to remedy the failure in a short time.
Usually, special knowledge is needed to create a diagnosis model that uses Bayesian networks. However, our failure diagnosis technology makes it possible to easily create a diagnosis model from a table.
For example, in order to create a diagnosis model for image quality trouble, we first create a two-dimensional table (Table 1 below) that shows the causal relationships between the causes of image quality trouble and the characteristics of each trouble, as based on past trouble and the information collected from experienced customer engineers. Then we enter the table into a software tool that automatically creates failure diagnosis models (shown in Fig. 2). The software tool automatically creates a failure diagnosis model by analyzing the causal relationship between the cause of image quality trouble and the characteristics of trouble from the input two-dimensional table (in the two-dimensional table analysis part), and then creating a Bayesian network (in the Bayesian network creation part).
Figure 3 below is an example of a created diagnosis model for image quality trouble. Diagnosis models are created for each major type of trouble such as streaks, spots, and deletion. Figure 3 shows a part of the diagnosis model for streaks.
Table 2 shows the performance evaluation results of a diagnosis model for image quality trouble that occurred at customers' sites. Even though the trouble included many cases of untypical trouble, the diagnosis model identified the causes of trouble with high accuracy. The percentage of correct identification was 82% for causes diagnosed as being most probable. When the second most probable causes are added, the percentage increases to 91%.
With the use of Bayesian networks, it is possible to develop a system capable of making comprehensive judgment based on the knowledge and observation of various events in a manner similar to that of human beings. Bayesian network technology offers the potential to bring about significant changes.