OBRABOTKAMETALLOV Vol. 23 No. 3 2021 MATERIAL SCIENCE EQUIPMENT. INSTRUMENTS 7 2 5 Multilayer perceptrons (MLP) are recognized as the most widely used ANN models. A MLP consists of several layers: an input layer that receives raw data; one or more hidden layers that process the data by applying weights and activation functions; and an output layer that produces the final output or prediction based on the processed data. Neurons in each layer are connected only to neurons in the next layer, with no feedback or connections between neurons within the same layer. Additionally, a typical feature of MLP is the full connectivity between layers. An example of a network structure consisting of four layers: input layer, two hidden layers, and output layer, is shown in Fig. 1. Fig. 1. Neural network structure for predicting the roughness parameter Rz In this structure (Fig. 1), the input layer has 8 nodes, each hidden layer has 8 nodes, and the output layer has 1 node. The nodes in the input layer represent the following factors: feed per tooth (fz, mm/tooth), angle of inclination (γ, °), tool diameter (D, mm), cutting speed (V, mm/min), cutting depth (ap, mm), side step (ae, mm), coolant supply (W, l/min), and tool wear (r, mm). The node in the output layer represents the predicted surface roughness parameter (Rz, μm). The network shown is fully connected, meaning that each neuron in any layer is connected to all neurons in the previous layer. Signal flow through the network is from left to right, layer by layer. Considering a multilayer network with j and k nodes in each hidden layer, the example structure shown in Fig. 1 can be described by an 8–j–k–1 configuration. In general, the operation of this type of network is described by two main phases: forward propagation and back propagation. The process of training MLP networks using the back propagation (BP) method follows the sequence: Forward Propagation → Loss Calculation → Back Propagation → Weight Update. An essential feature of MLP networks is the nonlinearity of neuron outputs, achieved by using the activation function. Successfully building an ANN model based on the Rz response requires multi-factor experimentation and tuning. Although many researchers have applied ANN for modeling in various fields such as machine learning [12–14], there is still no clear guideline for building a predictive model. This study examines elements that can affect model performance and the Rz response using the capabilities of the TensorFlow Python library to reduce uncertainty and improve prediction quality. Four performance indicators (metrics) were selected to evaluate the accuracy of models for predicting surface roughness [15]. These metrics include the coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), and root mean squared error (RMSE). The coefficient of determination
RkJQdWJsaXNoZXIy MTk0ODM1