Rolling　bearings　are　extremely　critical　rotating　mechanical　components,　and　when　they　fail,　they　can　damage　the　equipment,　causing　safety　threats　or　economic　losses.　Collecting　and　analyzing　vibration　signals　of　faulty　bearings　is　a　common　fault　diagnosis　method.　Modulation　signal　bispectrum　(MSB)　can　utilize　the　inherent　modulation　characteristics　of　the　signal　to　achieve　fault　feature　recognition,　but　it　is　easily　affected　by　strong　background　noise　and　other　modulation　signals,　making　it　difficult　to　detect　the　fault　feature　frequency　in　the　end.　Based　on　this,　a　time　-　frequency　modulation　bispectrum　(TFMB)　method　is　proposed　in　this　paper.　The　signal　is　first　transformed　into　the　time　-　frequency　domain　through　short　-time　Fourier　transform,　retaining　both　time　and　frequency　information.　It　reduces　the　amplitude　of　bispectral　noise,　highlights　spectral　lines　related　to　fault　information,　and　displays　modulation　information　more　clearly,　making　the　demodulation　results　more　accurate.　Based　on　the　characteristics　of　bispectrum,　a　two-dimensional　singular　value　decomposition　(2D　SVD)　strategy　is　designed　to　reduce　the　amplitude　of　the　noise　components　in　the　bispectrum,　and　the　relative　change　rate　is　proposed　to　automatically　determine　the　number　of　effective　singular　values.　At　the　same　time,　considering　that　the　Gini　index　has　good　sparse　property,　the　modulation　Gini　index　(MGI)　is　proposed　based　on　the　Gini　index,　which　can　identify　the　repetitive　pulses　in　the　signal,　suppress　the　random　pulses　and　noise　interference,　and　extract　the　modulation　component.　MGI　can　be　used　to　evaluate　the　proportion　of　fault　information　contained　in　TFMB　slices　and　select　the　optimal　frequency　slice.　The　simulation　signal　verifies　the　effectiveness　of　the　method　in　fault　diagnosis　of　rolling　bearing,　and　the　proposed　method　is　applied　to　the　diagnosis　of　rolling　bearing　outer　ring　fault,　inner　ring　fault,　and　compound　fault　respectively.　The　experimental　results　show　that　the　method　can　effectively　reduce　noise　amplitude,　enhance　fault　information　in　the　bispectrum,　and　extract　the　optimal　frequency　slice　to　accurately　detect　bearing　fault　characteristic　frequencies.　The　proposed　method　is　compared　with　MSB　detector,　Fast　Kurtogram　and　Autogram,　which　proves　the　superiority　of　TFMB　method　in　extraction　of　rolling　bearing　fault　characteristics.