Aiming at the nonlinear and non-Gaussian characteristics of batch processes, an improved genetic algorithm is proposed to optimize the parameters of separable convolution network and temporal convolutional network ( TCN ) for fault diagnosis of batch processes. Firstly, the SeparableConv1D-TCN network is constructed to extract features from the original intermittent data and perform fault diagnosis by standardizing the intermittent data. In order to find the network parameters with high diagnostic accuracy, the genetic algorithm with good point set method and optimal domain search is used to optimize. The effectiveness of the method is verified by simulation experiments and comparative experiments with penicillin experimental data.