Temporal-frequency　characteristics　in　fMRI　data　are　key　to　distinguishing　Autism　Spectrum　Disorder　(ASD)　from　neurotypical　individuals.　However,　the　non-linearity　and　multidimensionality　of　fMRI　data　pose　significant　challenges.　To　address　these,　we　introduce　a　Deep　Non-linear　Factorization　method　with　a　Wavelet　Temporal-Frequency　Attention　module　(Deep　WTFAF)　tailored　for　multidimensional　fMRI　analysis.　By　leveraging　the　wavelet　domain,　our　approach　applies　temporal-frequency　attention　to　assign　weights　to　significant　features,　enhancing　critical　data　while　reconstructing　incomplete　fMRI　data.　This　method　enables　deep　nonlinear　factorization　and　effective　feature　representation　for　subsequent　classification　tasks.　Validated　on　ASD-related　fMRI　datasets,　Deep　WTFAF　outperforms　traditional　methods,　maintaining　essential　information　and　ensuring　robustness　against　high-dimensional　and　incomplete　data.　Stability　theory　proof　further　confirms　the　model＇s　reliability,　crucial　for　clinical　applications　like　neurological　disorder　classification.