Abstract: FP-growth algorithm is very effective for mining frequent itemsets. However, the huge frequent pattern trees generated due to repeat F-list searching consume a large amount of memory and lead to low efficiency. In response to these drawbacks, this paper presented an improved algorithm, termed as UFP algorithm (upgraded FP-growth). First, it used support function to map the support rate with each item in order to improve the operating efficiency. Second, it took advantage of keyword filtering technology to devide the F-list into two parts, key-item list and non-critical list, ensuring the association rules which ultimately were obtained, were all valid information. Finally, it divided the whole database into subsets according to the first frequent itemsets and constructed condition pattern trees directly which saved lots of memory space. This paper applied UFP algorithm into TE process for fault diagnosis. The comparative experiments with PCA and KPCA algorithm under different process faults improve its superiority.