Bearings　are　an　important　component　in　rotating　machinery　and　their　failure　can　lead　to　serious　injuries　and　economic　losses,　therefore　the　diagnosis　of　bearing　faults　and　the　guarantee　of　their　smooth　operation　are　essential　steps　in　maintaining　the　safe　and　stable　operation　of　modern　machinery　and　equipment.　Traditional　bearing　fault　diagnosis　methods　focus　on　manually　designing　complex　noise　reduction,　filtering,　and　feature　extraction　processes,　however,　these　processes　are　too　cumbersome　and　lack　intelligence,　making　it　increasingly　difficult　to　rely　on　manual　diagnosis　with　large　amounts　of　data.　With　the　development　of　information　technology,　convolutional　neural　networks　have　been　proposed　for　bearing　fault　detection　and　identification.　However,　these　convolutional　models　have　the　disadvantage　of　having　difficulty　handling　fault-time　information,　leading　to　a　lack　of　classification　accuracy.　So　this　paper　proposes　a　transformer-based　fault　diagnosis　method,　using　the　short-time　Fourier　transform　to　convert　the　one-dimensional　fault　signal　into　a　two-dimensional　image,　and　then　input　the　two-dimensional　image　into　the　transformer　model　for　classification.　Experimental　results　show　that　the　fault　classification　can　reach　an　accuracy　of　98.45%.　©　2021,　Springer　Nature　Singapore　Pte　Ltd.