The neural network (NN) model is an accurate and efficient tool that can be used to simulate manufacturing processes. Many industrial plants in the past have chosen to adopt artificial neural networks (ANNs) to optimize manufacturing processes. Typically, the adoption of ANNs gives the ability to predict processed products’ mechanical properties around some specific technological parameters. That means that neural networks in general are hugely beneficial to the manufacturing industry as they allow material and cost resources to be saved.
Recently, there have been experimental observations where ANNs were used to control, optimize and monitor manufacturing processes. The neural network can be trained, and it can then develop numerical relationships between mechanical features and process-related factors.
Applications of neural networks in manufacturing
NNs and ANNs have led to revolutionary advances in the manufacturing industry. Artificial intelligence (AI) has been utilized for a while in manufacturing, however manufacturers primarily focused on bespoke expert systems, and have not benefitted from the specific advantages that ANNs bring, such as:
- Their parallel computing architecture, which can affect many applications and disciplines by processing natural language and speech, images, biomedical engineering and issues in bioinformatics
- Their self-adapting features, which predict any environment changes to improve the networks’ ability to learn
This makes ANNs a perfect tool for computer integrated manufacturing and for use in smart factories. The manufacturing industry has been changing rapidly in recent years and is a far cry from the traditional manufacturing plants that come to mind from the Industrial Revolution, or even plants from 20 years ago.
Injection molding processes
Injection molding processes are defined by their dynamic characteristics as their process input variables are the temperatures that melting occurs at, cylinder velocity and the pressure that forces the polymer into the mold. The phenomena that occur during this process vary in the length of time they take, are complicated and are uncertain. The complex nature of the input operating variables make it extremely difficult to relate them to the surface smoothness and accuracy of the geometry. ANNs are used here to optimize these processes with the use of multilayer perceptron, which is the most used network in this application to model the process and predict the quality of the final product.
Gas metal arc welding processes
During gas metal arc (GMA) processes, an electrical current is produced by an electrical arc, which is kept between a wire electrode and the welding metal. The consumable electrode and filling metal are both fed automatically by a wire feeding device. A decent weld is determined by a high depth to width ratio of the pool of molten metal. So, monitoring and controlling the surface temperatures and weld geometry that determine the formation of the weld pool are very important, as they directly affect the back bead width and penetration depth. An ANN multilayer perceptron used recent studies and an infrared sensing system to control and detect these variables to a high degree of accuracy.
There are many variables that need to be controlled in order to maintain a certain level of quality, like vibrations, the state of the cutting tool, temperature and forces created during the machining process. One application of ANNs here is to use the previously mentioned process data to predict and classify the tool life expectancy, detect tool failure and identify the status of tool wear.
How do neural networks work?
ANNs are computational models that identify a relationship between output variables and process factors. Artificial neurons work as adjustable coefficients and can be combined through “weights.” There are many different frameworks and programs, either to simulate neural structure, specific functions or for general purpose. There is not one specific computational model that is being used by the whole industry, as different approaches are more suitable for various applications. Essentially, most ANNs can be utilized fully on the basis of the computer hardware’s features it is installed on.
The main features of any network are as follows:
- Architecture — identifies the connections between the neurons and layers
- Dataset splitting — determines the subset data to be used for training, validation and testing for the ANN’s development
- Learning algorithm — designates the weight of the links between the neurons
ANNs are garnering more interest almost daily, as their expansive list of benefits have become too much to ignore for companies in the manufacturing industry. It is growing in popularity so much in fact, that the algorithms and models that are used are quickly becoming standard tools in information engineering and computer science.
Now that they have progressed from adolescence, neural networks have put together a robust set of tools, including computation procedures that feature an undeniable effectiveness and strong theoretical base. For example, in this article the ANN simulation model showed it was efficient at predicting the friction stir welding (FSW) process. Adoption of simulation models like this one can be very beneficial to FSW, and the manufacturing industry as a whole.
In this case, this tool was able to predict the mechanical properties of welds and maintain control of the overall welding process. This decreases the number of repairs that are needed and the inevitable costs that repairs bring. With the constant growth and development of neural networks, the next question is, can we actually build a human brain? While this is still a long way off and an ambitious goal to say the least, the ANN can still imitate and support human behavior. This helps with simplifying the decision-making process, analysis of complicated processes, and increasing evaluation and cognition under stressful conditions, among many other things.