The　fault　diagnosis　of　equipment　is　important　for　the　operation　and　maintenance　management　of　industrial　process.　However,　the　abnormality　of　key　equipment　operation　data　will　affect　its　fault　diagnosis　effect.　In　this　paper,　we　reduce　the　influence　of　abnormal　process　data　on　fault　diagnosis　results　by　an　improved　probabilistic　neural　network　(PNN).　Firstly,　the　data　used　to　build　the　network　is　analyzed　to　characterize　its　distribution　in　feature　space,　and　the　pattern　layer　output　is　calculated　to　reduce　the　effect　of　outliers.　Secondly,　in　order　to　obtain　a　robust　center-of-mass　representation　of　the　data　in　the　feature　space,　a　centroid　that　is　robust　to　outliers　is　designed　and　used　in　the　weighting　strategy　of　the　network.　Then,　the　input　weighting　strategy　of　the　network　is　designed　and　determined　by　the　similarity　between　the　pattern　layer　centers　with　high　confidence　and　the　network　input.　Finally,　experiments　have　been　conducted　in　a　classification　task　and　a　wing　damage　task,　and　the　results　show　that　the　method　is　effective　under　data　contamination.　©　2023　IEEE.