Aiming at the accuracy challenges faced by magnetostriction modeling under complex excitation conditions

such as different magnetic density amplitudes

frequencies

and harmonic combinations

this paper improves the traditional convolutional neural network (CNN) by nesting spatial attention mechanism (SAM) to develop an improved SAMCNN network. First

the SAMCNN-BiLSTM magnetostrictive model for electrical steel sheets is proposed by combining with the bidirectional long short-term memory (BiLSTM) network. The structure parameters of the SAMCNN-BiLSTM are optimized by using the grey wolf optimization (GWO)

which realizes the accurate characterization of the magnetostrictive effect under complex working conditions. Then

it establishes a database of magnetostrictive strains under complex working conditions such as single-frequency excitation and superimposed harmonic excitation

and carry out data preprocessing and characterization. Finally

the SAMCNN-BiLSTM model is compared and verified. Comparing the main components of the magnetostrictive strain spectrum under superimposed third harmonic excitation

the maximum relative error of the calculated values of the SAMCNN-BiLSTM model is 3.70%

which can more accurately characterize the magnetostrictive effect of electrical steel sheets than the Jiles-Atherton-Sablik (J-A-S) physical model

and the quadratic domain rotation model

etc.