Abstract: As the complexity of tensor data structures continues to increase, the computational complexity and workload within Optical Neural Network (ONN) have soared, accompanied by a rise in the number and sophistication of the required optical components. To reduce the complexity of neural networks and enhance their applicability, the dimensionality reduction of tensor data emerges as an effective strategy. In this paper, we construct ONN based on GridNet and FFTNet architectures, where the linear optical processing units are implemented using Mach-Zehnder Interferometer (MZI). The Principal Component Analysis (PCA) method is employed to reduce the dimensionality of ONN inputs, enabling the simulation and training of highly robust and compact ONN. We utilize four training algorithms, Stochastic Gradient Descent (SGD), RMSprop, Adam, and Adagrad, along with various learning rates, to complete the forward propagation and backward training of the ONN on the MNIST handwritten digit dataset. Subsequently, we characterize and analyze the performance and robustness of the respective networks. The research results reveal that reducing the input dimensionality of ONN using PCA can effectively reduce the network size while maintaining a certain level of recognition accuracy and robustness against MZI device errors. Furthermore, to achieve superior training results (the recognition accuracy greater than 90%), it is crucial to select appropriate training algorithms and learning rates for ONN with different input dimensions, scales, and architectures. Furthermore, the performance of the FFTNet architecture exhibits low sensitivity to changes in input dimension, and FFTNet with varying input dimensions and numbers of hidden layers demonstrates high robustness in handling phase errors and splitter errors. Thus, assisted by PCA technique, the ONN with FFTNet architecture is compact, capable of maintaining stable performance on datasets of different sizes, and more reliable.