Understanding Machine Learning for Friction Stir Welding Technology

Abstract achine learning process drastically decreases the time it takes to develop stronger, lighter materials. This is important to the aerospace, automotive and manufacturing sectors. Machine Learning techniques like Artificial Neural Networks and Image processing are used in a solid state joining process such as Friction Stir Welding process (FSW) for optimization of mechanical properties like Ultimate Tensile Strength (UTS), Fracture Strength , Elongation percentage and microstructure properties like grain size and understanding defects formation. In the recent paper, application of Machine Learning technique in Friction Stir Welding technology will be discussed. 1. Understanding Human Nervous System The human nervous system can be classified as a three stage system which is shown in the block diagram of Figure 1. The brain is in central position to the nervous system which is represented by the neural net. The function of the neural net is to receive the information continuously, perceive it and then to make appropriate decision.


Understanding Human Nervous System
The human nervous system can be classified as a three stage system which is shown in the block diagram of Figure 1. The brain is in central position to the nervous system which is represented by the neural net. The function of the neural net is to receive the information continuously, perceive it and then to make appropriate decision. The green arrows which are pointing from left to right represent the forward transmission of information carrying signals through the nervous system. The red arrows in the diagram show the presence of feedback in the nervous system. Stimuli from the external environment or the human body are converted to electrical impulses by receptors which further convey the information to the brain (neural net). Electrical impulses which are generated by the receptors are further converted to a distinguishable response as a nervous system outputs by the effectors [1]. Figure 2 represents the schematic diagram of the human nerve cell.

Figure 2. Representation of human nerve cell
Neuron is the fundamental unit of the neural network architecture. Dendrites, soma and axon constitute the neuron structure. Signals from surrounding neurons are received by a tree like structure called Dendrites. Each of the line is connected to one neuron. The signal from one neuron to others is transmitted by a thin cylindrical structure is called an Axon. Synapse is responsible for making contact between at the end of axon and dendrites. At the synapse, the inter neuronal signal is propagated through chemical diffusion or by electrical impulses. Only if certain condition is met, the neuron fires an electrical impulse [2]. If the total weight of the synapses which receive impulses in the period of latent summation become more than the threshold then it causes neuron fire [3].

Modelling the architecture of Artificial Neural Network
The information processing unit of the basic neural network architecture is a neuron. Non-linear model of a neuron is depicted in the Figure 3. The given neural model has three basic elements i.e. an adder, set of synapses and an activation function. The characterization of the connecting links or set of synapses are done by its own weight or strength. Signal at the input of synapse j which is connected to the neuron k is further multiplied by the synaptic weight . The synaptic weight of the artificial neuron lies between the both positive and negative values. An adder performs the operation of the linear combination. It sums up the input signals which are weighted by the synaptic strengths. The main application of the activation function is to limit the amplitude of the output of neuron. The bias which is shown in the Figure 3 is applied externally and its main function is to decrease or increase the net input of the activation function.
In the form of mathematical expression, we can define the neuron k shown in the Figure 3 with the help of the pair of equations: and, = (

Machine Learning and its classifications
Machine learning technique give priorities to the development and execution of the computer programs that can process data and use that particular data to learn for themselves. It's just like before solving calculus exercise numerical problems, you go through the various examples given in that particular chapter i.e. you are training your mind and when you solve new question it is called testing your mind and the result you get for that particular question measures the accuracy of your performance. So we can define Machine learning as a subset of artificial intelligence (AI) which provides the given systems or state the ability to learn by itself and further modify from the experience without being explicitly programmed [4][5][6][7][8]. Machine learning classified into three main types i.e. reinforcement learning, supervised learning and unsupervised learning which will be discussed one by one in the following sub sections.

Supervised Learning
Supervised learning can be referred to as the situation of learning a concept with a teacher. The working of supervised machine learning is shown in the Figure 4. We may assume that the teacher possess the knowledge of the given environment which is unknown to the neural network architecture. The representation of the knowledge is done by a set of input-output examples.  Figure 4 shows that a training vector which is dragged from the same environment is exposed to neural network and the teacher. So, for that training vector, a desired response is given by the teacher with the assistance of the possessed knowledge. Error signal shown in the Figure 4 is the difference of the actual response and the desired response. Under the combined influence of the error signal and the training vector, network parameters are adjusted. This step by step adjustment is carried out further which makes proper transfer of information of the environment possessed by the teacher to the neural network.

Unsupervised Learning
Unsupervised machine learning is a self-organized learning process in which there is no involvement of external person or teacher to look after or observe the training process. The working of the unsupervised machine learning process is shown in the Figure 5. In order to carry out unsupervised machine learning task, a competitive learning rule is used. Just take an example of a neural network consisting of two layers i.e. an input layer and a competitive layer. The available data is received by the neural network. If we talk about the competitive layer, it consists of various neurons which compete against each other for the opportunity for responding the features contained in the input data as per the given learning rule.

Reinforcement Learning
Reinforcement learning is a semi-supervised machine learning method which trains the models to make a sequence of decisions. The working of the Reinforcement learning is shown in the Figure 6. In an uncertain, potentially complex environment agent learns to achieve a particular goal. An artificial intelligence is subjected to a game-like situation where the computer employs trial and error method to come up with a solution to the given problem. The main objective of the Reinforcement Learning is the maximization of the total reward. Reinforcement Learning works on the rewards and penalties scheme. The Reinforcement Learning model tries to figure out the method to perform the given task in order to maximize the reward, starting from totally random trials and finishing with advanced tactics and superhuman skills.

Application of Machine Learning in Friction Stir Welding Process
Verma et al. [9] used various sophisticated machine learning approaches like . Support Vector Machining (SVM) , Gaussian Process (GP) regression, and multi-linear regression (MLR) for evaluating the friction stir welding process. As shown in Figure 7 it is observed that in comparison to the SVM and MLR models, the GPR approach works better. Therefore, GPR approach is generally used successfully for prediction of the UTS of Friction Stir welded joints.
Debroy et al. [10] studied the conditions for void formation using a Bayesian Neural Network and a decision tree. Schematic representation of the research is shown in the Figure 8. The study showed that the neural network and the decision tree predicted void formation with 96.6% accuracy by computing the causative variables like temperature, strain rate, torque, and maximum shear stress on the tool pin.
Celik et al. [11] investigated the correlation between the friction welding parameters and tensile strength of both AISI 316 austenitic-stainless steel and Ck 45 steel by developing an Artificial Neural Network (ANN) model. Figure 9 represents the artificial neural network architecture used in the study.

Actual UTS (MPa)
Predicted UTS (MPa) Figure 8. The components are FSW process, mechanistic models, and machine learning methods (neural network and decision tree) [10].  Figure 10 shows that a good correlation was obtained between the Artificial Neural Network predicted values and the experimental values. Figure 10. Comparison of measured and predicted outputs for tensile strength [11] Maleki et al. [12] developed Artificial Neural Network based on backward propagation algorithm for predicting the yield strength, tensile strength, notch-tensile strength and hardness of friction stir welded AA 7075-T6 joints. The Neural Network architecture used in the study is shown in the Figure 11.  Mishra et al. [13] implemented the Convolutional Neural Network for identification of the texture of Friction Stir Welded joints and Conventional Welded joints. Macias et al. [14] established a correlation between the acoustic emission signals and the various main parameters of friction stir welding process based on artificial neural networks (ANNs) trained on Levenberg-Marquardt algorithm. Figure 13 and 14 shows the methodology and the development of Artificial Neural Network architecture respectively.

Designing a simple Artificial Neural Network on Google Colab by using Python coding
In this section, development of an Artificial Neural Network model by using Python Programming on Google Colaboratory platform as shown in Figure 16 will be discussed. Table 1 shows the experimental datset which shows the Ultimate Tensile Strength (MPa) value of Friction Stir Welded joints against the tool rotational speed of the tool (rpm). First we will start with our imports. Here we are importing TensorFlow and calling it tf for ease of use. We then import a library called numpy, which helps us to represent our data as lists easily and quickly. The framework for defining a neural network as a set of Sequential layers is called keras, so we import that too as shown in the Figure 17. Next we will create the simplest possible neural network. It has 1 layer, and that layer has 1 neuron, and the input shape to it is just 1 value as shown in the Figure 18. Now we compile our Neural Network as shown in Figure 19. When we do so, we have to specify 2 functions, a loss and an optimizer. The LOSS function measures the guessed answers against the known correct answers and measures how well or how badly it did. It then uses the OPTIMIZER function to make another guess. Based on how the loss function went, it will try to minimize the loss. It will repeat this for the number of EPOCHS which you will see shortly. But first, here's how we tell it to use 'MEAN SQUARED ERROR' for the loss and 'STOCHASTIC GRADIENT DESCENT' for the optimizer.  The process of training the neural network, where it 'learns' the relationship between the Xs and Ys is in the model.fit call which is shown in Figure 21. This is where it will go through the loop we spoke about above, making a guess, measuring how good or bad it is (aka the loss), using the opimizer to make another guess etc. It will do it for the number of epochs you specify. When you run this code, you'll see the loss on the right hand side.

Figure 21.
Training and Testing the Neural Network

Conclusion
There is a loss of time and materials if the optimization of the Friction Stir Welding parameters is done through experimental studies which further leads to increase in the cost of the experiment. Machine Learning approach like Artificial Neural Network and image processing overcome these issues. So, it can be concluded that the mechanical and microstructure properties can be predicted and also the defects formation can also be observed by the implementation of various Machine Learning tools in the Friction Stir Welding process.