Researchers have studied approaches to detect fault-prone modules for a long time. As one of these approaches, we proposed an approach using a text filtering technique. In this approach, we assume that faults relate to words and contexts in a software module. Our technique accepts inputs as a text information. Based on a dictionary that was learned by classifying modules that induce faults, the fault inducing probability over a target module is calculated, and it judges whether the given module is a fault-prone module. Although our approach targeted the source code of software, especially in embedded software, the analysis of byte-code is also required. The source code based fault detection suffered from noises such as the way of writing, the used name of identifiers, and so on. Eliminating such noises may improve the accuracy of prediction. In this study, we aimed at fault detection from the byte-code of Java. Specifically, we tried to detect faults from the disassembled intermediate code of Java class file. To show the effectiveness of our approach, we conducted an experiment and compared our approach with source code based approach.