Multiscale Time-Frequency Sparse Transformer Based on Partly Interpretable Method for Bearing Fault Diagnosis
Algorithm 1
Training of MTFST.
Input: Three multiscale TFRs ,, where ,,, and , which denote the fault types.
(1)
Set training batch , training epoch max_epoch, token embedding dimension , self-attention weight matrix size , number of head , positionwise forward network weight matrix size , block number of encoder , block number of decoder , and number of fault types .
(2)
Initialize trainable parameters of MSTFT
(3)
for epoch in 1, 2, …, max_epoch do
(4)
for step in 1, 2, …, max_step do
(5)
//Tokenizer
(6)
for each in , in and in do
(7)
Reshape ,, to = , = and = then slice into patches sequence [], [], [];
