|
| Algorithm used | AI model taxonomy | Depth, layer sizes, training time, testing time | Dataset | Framework, core language, interface | Ref. |
|
| The backpropagation algorithm trains the weights of feedforward NNs consisting of multiple layers to predict the mass loss quantities of A390 aluminum alloy. | Supervised learning | An NN has a first layer containing three neurons and a second layer containing two neurons; overall, a two-hidden-layer network was used. | Data were normalized in the middle of [0, 1]. | No data | [53] |
| An NN containing two hidden layers was used. The standard load, environment, and time are the three input variables. The amount of wear loss and microhardness are the two output variables. The LMA was used to train the ANN along with BP. | Supervised learning method: BP | Three layers, 3-2-2 topology | No data | No data | [54] |
| trainlm (network training function) was used to train a multilayer ANN. Two outputs were recorded. The LMA was applicable in adjusting the biases and weights. | Feedforward NN with BP | The ANN with three inputs and two hidden layers has the first layer containing 20 and 30 neurons.
The data used for training were 70% and 15% each for validation and testing. | No data | MATLAB 2016a for the ANN
Minitab 16 for visualizing the linear regression model | [55] |
The NNs were diversified and tested to discover the most acceptable results possible. Thrust, cutting speed, and force were the inputs, and tool wear was the output.
In contrast, the second is for predicting the surface roughness. | Feedforward NN using BP | NN topologies 3-5-1 and 3-4-1 were tested for tool wear, and 4-6-1 and 4-6-4-1 were tested for predicting surface roughness. | Laboratoire Génie de Production, ENIT Tarbes, France | No data | [56] |
| SVR (support vector regression) was applied to solve the regression problem, and here the least square error is also used; therefore, it is known as LSSVM. | LSSVM | The kernel chosen was the radial basis function. | No data | MATLAB 2013 | [57] |
| A model having a three-layer Taguchi coupled ANN was proposed. The input nodes were sliding distance, load, sliding velocity, and weight percentage. The hidden layer had seven neurons, whereas the output layer had a single neuron. The LMA was used to train this model. | Supervised learning | 3-7-1 architecture ANN | No data | MATLAB 2013 | [58] |
The network consists of three layers and four PCA-declared input nodes, whereas the hidden layer has three nodes, and the output layer having a single node was best out of all the networks.
The first two layers used the neural transfer function tansig, whereas the last layer used purelin. | Supervised learning | 4-3-1 architecture with the three-layer feedforward ANN | No data | A commercial Neural Network Toolbox | [59] |
| Adaptive neuro-fuzzy inference system combines the “Takagi–Sugeno fuzzy inference system” and the principles of ANN. | Supervised learning | Five layers in ANFIS | No data | MATLAB | [60] |
| Taguchi technique has been used here. A weight-loss model to make predictions was made using regression. Nonlinear regression was used to correlate control factors and weight loss. | Supervised learning | No data | No data | Minitab 15.1 | [61] |
| The AI algorithms used here are random forests, regression trees, MLP, and RBF. | Supervised learning: MLP, RBF, and random forests | RBF has one layer, and MLP has one hidden layer. | The experimental data were collected to provide a broad number of wear conditions and processing times while acquiring data on the power drive for a fixed machining process—the face milling of carbon-quality structural steel 45. | No data | [62] |
| “Kohonen’s self-organizing map” was used to evaluate the tool’s working status. Also, a triangular membership function applied neuro-fuzzy and fuzzy logic. The “centroid method of defuzzification” was used to obtain the flank wear. | Supervised learning: backpropagation NN | 2-3-1 architecture | The training data for the networks were collected through experimental studies. | No data | [63] |
One neuron represents each input parameter distinctively related to the coefficient of friction.
The input variables include applied load, sliding velocity, sliding distance, and material type, whereas the output is the coefficient of friction. It has 4-6-4-1 architecture. An MLP model was applied here because of its feedforward nature. | Supervised learning: MLP | 4-6-4-1 architecture | No data | No data | [64] |
| For evaluating the tool wear, a developed configuration system was applied. Also, using an expert system at different wear states helped clarify the output values of ANN. | Unsupervised learning: ART2 | Number of input neurons in SOM: 15, and number of neurons in an SOM layer: 36 | No data | No data | [65] |
The network used in this study was a generalized feedforward network. Input parameters were sliding time, sliding speed, load, and Al-Si%, whereas the output parameter was specific wear rate. The network consisted of three hidden layers with 16, 8, and 5 neurons.
The first two layers used the TanhAxon function, whereas the last layer applied the BiasAxon function. | Supervised learning | 4-11-5-1 architecture and two hidden layers with four inputs and one output layer were applied. | No data | No data | [66] |
| The LMA along with BP was applied in this study. Load and speed are the two nodes of the input layer, whereas the friction and mass loss coefficient are the two nodes of the output layer. The minimal fault was observed in the output due to ten neurons in the hidden layer. | Supervised learning | Two output and ten hidden neurons | No data | MATLAB | [67] |
| The proposed reduction model here is a combination of POD and RBF. | Supervised learning | The network consists of two layers, one with RBF neurons and the other with output neurons. | No data | MATLAB | [68] |
| The ANN was trained with the Levenberg–Marquardt algorithm (LMA), Bayesian regulation (BR), resilient backpropagation (RP), scaled conjugate gradient (SCG), and gradient descent (GD). | Supervised learning | Five distinct training algorithms were used, along with eighteen different architectures.
The Bayesian algorithm trained in a two-layered neural network has reached the best results (26 10 5 1). | Training data were obtained by 360 randomly distributed data collected from testing of four friction materials.
Training data were acquired by testing eight different friction materials, only predicting fade performance. | No data | [69] |
| The ANN and Sugeno FIS have been applied, and BP having 4-3-1 architecture and LMA is adopted here. | Supervised learning: backpropagation | The network has 4-3-1 architecture and one hidden layer. | No data | MATLAB R2015a using NN Toolbox | [70] |
| FZM and ANN, along with a neuro-fuzzy ANFIS, are adopted here. | Supervised learning: backpropagation for ANN
Unsupervised learning for fuzzy c-means clustering | The ANN having 4-3-1 architecture and ninety-cluster C-mean clustering gave the best performance. | No data | No data | [71] |
| The Elman-inspired RNN was applied. The sensor uses the relationship between the variables to be measured and the power consumption. | Bayesian regularization | The best model HU55 implies five hidden units and a delay of 5. | A Training and Test Data Set (TTDS) is generated with a specific combination of the grinding experiments collected. | MATLAB | [72] |
| RF, MLP, RBF, etc., were used in this study to predict surface roughness and mass loss. | Supervised learning: regression trees, MLP BP | A network having a three-layer architecture and a hidden layer consisting of RBF was used. | No data | No data | [73] |
| Output, i.e., tool wear, is predicted with the help of residual errors as the basis of decision-making. | Supervised learning: MLP | MLP has 6-12-1 architecture, and one hidden layer was used here. | No data | MATLAB | [74] |
| Volume loss is predicted using LR, SVM, ANN, and other extreme learning methods. | Supervised learning: ANN, SVR, and LR | The ANN has a 3-4-1 architecture and a quadratic function as the SVR kernel, whereas ELM used here is a feedforward NN having a single hidden layer. | Experimentally obtained data | MATLAB | [75] |
| Supervised learning methods such as SVR, RF regression, decision tree regression, GBR, GPR, MLP, and KNN are used. | Supervised learning | SVR uses an RBF for the kernel.
MLP having five hidden layers and ten neurons in each layer with ReLU activation was used here. | Collected from 13 references of 316L SS parts processed by SLM | Python TensorFlow, scikit-learn, Google Colab | [76] |
| The ANN with BP is applied along with ANOVA to decide the potential parameters to predict the specific wear rate reduction. | Supervised learning | The ANN has the 2 : 5:1 architecture with sigmoid activation. | No data | Python, Minitab 19 | [77] |
| Analysis of the erosion process is done using the ANN model along with LMA. | Supervised learning | The network having three layers and 2-6-3 architecture is used here. | No data | MATLAB 2017a Neural Network Toolbox | [78] |
| ANN and RSM models were compared based on their predictive capacity of wear behavior of fabricated composites. | Supervised learning | Three inputs, ten hidden layers, and two outputs | No data | MATLAB | [79] |
|
