This article presents a novel approach to machine learning using orbital angular momentum (OAM) states, leveraging all-optical convolutional neural networks (CNNs) based on Laguerre-Gaussian (LG) beam modes. The proposed CNN architecture utilizes a trainable OAM mode-dispersion impulse as a convolutional kernel for feature extraction and deep-learning diffractive layers as a classifier. The OAM mode-dispersion selectivity enables high-accuracy mode-feature encoding, achieving 97.2% accuracy on the MNIST database by detecting the energy weighting coefficients of encoded OAM modes. The system also demonstrates resistance to eavesdropping in free-space optical communication. By extending the target encoded modes into multiplexed OAM states, the system achieves all-optical dimension reduction for anomaly detection with 85% accuracy. The CNN architecture is designed to encode input data into OAM states, enabling applications such as image classification, end-to-end switchable image display, and all-optical abnormal detection. The system uses OAM modes for information encoding, which can be decoded using an appropriate OAM-dependent hologram. The OAM-mediated machine learning approach provides a deep insight into the mechanism of machine learning with spatial modes basis, offering potential for improving various machine-vision tasks through unsupervised learning-based auto-encoders. The study demonstrates the feasibility of all-optical information mode-feature encoding for anti-eavesdropping wireless image transmission. The system uses OAM encoding to enable high-density data transmission and enhance security, particularly suitable for atmospheric turbulence-free links. The all-optical CNN architecture is implemented using four spatial light modulators (SLMs) and achieves high encoding accuracy. The system is also applied for all-optical dimension reduction in abnormal detection, achieving 85% accuracy through principal component analysis and spectral clustering. The research highlights the potential of OAM-mediated machine learning for various applications, including high-capacity holographic communications, LIFI, cellular deformation classification, and face similarity recognition. The study provides a comprehensive framework for all-optical machine learning, demonstrating the effectiveness of OAM-based approaches in enhancing the performance of optical communication systems. The results show that the proposed CNN architecture can achieve high accuracy in image classification and anomaly detection, with potential for further improvements through experimental realizations of free-space ONN-based OAM sensing technologies.This article presents a novel approach to machine learning using orbital angular momentum (OAM) states, leveraging all-optical convolutional neural networks (CNNs) based on Laguerre-Gaussian (LG) beam modes. The proposed CNN architecture utilizes a trainable OAM mode-dispersion impulse as a convolutional kernel for feature extraction and deep-learning diffractive layers as a classifier. The OAM mode-dispersion selectivity enables high-accuracy mode-feature encoding, achieving 97.2% accuracy on the MNIST database by detecting the energy weighting coefficients of encoded OAM modes. The system also demonstrates resistance to eavesdropping in free-space optical communication. By extending the target encoded modes into multiplexed OAM states, the system achieves all-optical dimension reduction for anomaly detection with 85% accuracy. The CNN architecture is designed to encode input data into OAM states, enabling applications such as image classification, end-to-end switchable image display, and all-optical abnormal detection. The system uses OAM modes for information encoding, which can be decoded using an appropriate OAM-dependent hologram. The OAM-mediated machine learning approach provides a deep insight into the mechanism of machine learning with spatial modes basis, offering potential for improving various machine-vision tasks through unsupervised learning-based auto-encoders. The study demonstrates the feasibility of all-optical information mode-feature encoding for anti-eavesdropping wireless image transmission. The system uses OAM encoding to enable high-density data transmission and enhance security, particularly suitable for atmospheric turbulence-free links. The all-optical CNN architecture is implemented using four spatial light modulators (SLMs) and achieves high encoding accuracy. The system is also applied for all-optical dimension reduction in abnormal detection, achieving 85% accuracy through principal component analysis and spectral clustering. The research highlights the potential of OAM-mediated machine learning for various applications, including high-capacity holographic communications, LIFI, cellular deformation classification, and face similarity recognition. The study provides a comprehensive framework for all-optical machine learning, demonstrating the effectiveness of OAM-based approaches in enhancing the performance of optical communication systems. The results show that the proposed CNN architecture can achieve high accuracy in image classification and anomaly detection, with potential for further improvements through experimental realizations of free-space ONN-based OAM sensing technologies.