Use of neural networks in the technical operation of self-propelled machinery

The actual service life of tractor engines of traction class 1.4 in the Tomsk region falls significantly short of the warranty period and exhibits high dispersion. Statistical analysis reveals that the mean operating time until the first overhaul does not exceed 7,000 engine hours, with a standard deviation of 1,707 hours and a coefficient of variation of 0.24. The service life of new engines until the first overhaul varies by more than a factor of 2.8. To address the challenge of predicting failure modes based on cumulative operating time, this study employs artificial neural networks (ANNs). The research objective was to train an ANN to identify the most likely cause of engine failures using durability data collected under routine operating conditions of traction class 1.4 tractor engines. The authors developed an intelligent failure diagnostics system using Python and the PyTorch framework. The Matplotlib module was used for visualization, NumPy for matrix operations, and sklearn for input data normalization. The ANN uses a fully connected (dense) architecture consisting of an input layer (one neuron), a hidden layer (10 neurons), and an output layer (four neurons). The model was trained on a dataset from 25 Minsk Motor Plant engines (type 4Ch(N) 11/12.5). Based on the “operating time” input parameter, the model generates a probability distribution across four failure categories: the crank mechanism, the lubrication system, the fuel system, and the cooling system. Initial testing yielded a prediction accuracy of 60%. Future research will focus on fine-tuning the artificial neural network by expanding the training dataset to achieve a target accuracy of 80%.