In this paper, we propose a method for analyzing failure reports written in natural language. We introduce text coding methodology, which is mainly used in the field of sociology. Although this methodology is useful to improve the reliability of analysis, it is very costly and tedious work to assign codes to a large amount of sentences. Therefore, we introduce two methods to support coding work. One method is the clustering of sentences in failure reports. This clustering method supports users to generate adequate codes as the metric of failure reports and to systematize them. The other is a rule-based code assignment method, which correctly handles synonymous words and conceptually similar words using the ontology of failure reports. We applied our coding methods to nearly one thousand reports and analyzed their failure status. The result shows our system is promising to realize the reliable failure analysis.