Application of Convolutional Neural Network for Classification of Automatic Cab Signalling Codes

Abstract

Purpose. Justification the selection of architecture and hyperparameters for an artificial neural network designed to classify Automatic Cab Signalling (ALSN) code signals under conditions of intense interference and temporal distortion. Methodology. To achieve the purpose, a comparative analysis of different types of artificial neural networks was performed. It was found that for the classification of ALSN codes, a one-dimensional convolutional network (1D-CNN ) is the most suitable, as it effectively detects the characteristic temporal features of signals, requires less data for training, is resistant to noise, and provides stable classification even with a signal shift in time. Calculations of the number of network parameters were performed based on its configuration, including the number of layers, convolutional filters, and the number of neurons in fully connected layers. For model training, synthetic data were generated by adding noise and stochastic timing variations to reference signals («Z», «Zh», and «ChZh»). 1000 implementations were synthesized for each ALSN code signal. The program code for data preparation, network training, and signal classification was written in Python using the TensorFlow, Keras, Scipy.signal, and NumPy libraries. Findings. Based on the observed correlation between network configuration and recognition accuracy for the ALSN codes, an optimized model was selected featuring three convolutional layers, a 16-neuron Dense layer, and a 3-neuron output layer. The robustness of the proposed network was validated against both heavily distorted synthetic waveforms with low signal-to-noise ratios and empirical ALSN signals. Originality. This work represents the first comprehensive study of artificial neural networks efficacy for ALSN code classification under severe interference and temporal distortions. It justifies the specific 1D-CNN configuration and establishes the dependencies between model structural parameters (number of layers, filters, and neurons) and classification accuracy. Practical value. Replacing legacy relay-based equipment with the proposed 1D-CNN-based digital receiver will reduce operational expenditures while significantly enhancing the noise immunity and reliability of the ALSN system through stable code recognition in adverse settings, i.e. intensive noise and temporal parameters disturbances of the code signals.