CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 IMPLEMENTATION OF ARTIFICIAL INTELLIGENCE IN WELDING TECHNOLOGIES D.Eng. IWE Bodea Marius Technical University of Cluj, Faculty of Materials and Environmental Engineering mbodea@stm.utcluj.ro Abstract. The paper presents some analysis regarding the artificial intelligence impact on economy, especially in the welding manufacturing. The artificial intelligence systems are increasingly present in our daily life, but even more in the nowadays-modern technologies. The AI systems are the ideal solutions for process control or in decisions actions for processes strongly characterized by non-linear variables. The welding processes are by excellence environments strongly characterized by non-linear variables, like material property’s function temperature, processes and technological parameters, human factors and others. In the paper is presented also a software program developed by author for interpretation and realization of TTT and CCT diagrams for welding. For this purpose, it was used the IA and neuronal network analysis of data collected from CCT scanned diagrams. During your lecture of this abstract, in about 1 minute, Siri which is the IA virtual assistant for Apple products, has given over 100.000 answers all over the world (accordingly to 2015 statistics). 1. Introduction 1.1. What is artificial intelligence? The modern society evolves extremely fast in an exponential manner, taking by surprise even the most renown specialists in the world. Robert Melancton Metcalfe, a professor from Texas University, Austin teaching the course “Innovation and Entrepreneurship”, one of the internet inventors in 1970, coauthor for Ethernet patent and cofounder of 3Com Company, is famous also for his prediction from 1995. According to his prediction, the internet would dramatically be collapsing in terms of maximum one year and it was so sure about that, thus he made a public statement that if this would not happen, he will eat its own words. Two years later, when he was invited with a keynote speech to the 6th Edition of International World Wide Web Conference, he brought himself a blender on the scene, he mixed the paper with his famous prediction and as he promised, he was eating its own words in the front of auditorium’s Conference. World experts from highly ranked company have made a series of economic analysis regarding the Artificial Intelligence (AI) evolution and its impact on the economy. Thus, Delloite company has estimated that until 2021 will be invested over 57.6 billion dollars in AI and Machine Learning (ML) development, that represent an increase of five times fold in respect with 2017. McKinsey Global Institute estimated that the AI impact on 19 fields of activity would reach 3.5 to 5.8 billion dollars [2]. The multi billionaire Elon Musk, founder of SpaceX and Starlink program, which is providing internet in all over the world by using a network of over 12.000 satellites, and planned to grow up to 40.000 satellites of communication, has expressing repeatably his concern on the AI evolution. According to Musk, different robots controlled by AI will take the control over the economic activities in the next 5-10 years and in the end, over the entire world economy and governments. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 Elon Musk is most worried by the DeepMind project who was sold by his company to the Google Group for 500 million dollars in 2014. Google has using the DeepMind AI program in all its data centers from all over the world, in order to manage the energy consumption for over theirs 900.000 servers. Each Google data center is consuming the same energy like a town and generate huge quantities of heat. In 2011, Google has reported a consumption of electric energy of 0.01% from all energy consumed in all data centers over the world. The AI controlled program DeepMind has optimized the cooling system efficiency with 40%, decreasing the energy consumption with 15% [3]. Aftermath of DeepMind acquisition by Google, they were using the AI in medical applications which were freely offered to the National Health System (NHS) from USA. Starting only from basic rules of chess, shogi or Go, the Google Alfa Zero AI program has learned all these games in a 4 hours’ time span by playing against himself. In the following competitions, the Alfa Zero has crushed everything, from chess masters to others previous software programs that have beaten Kasparov chess world champion (Deep Blue in 1997) or the Go master world champion Lee Sedol, who was defeated by AlphaGo in 2016 (previous version of Alpha Zero AI program). 1.2. Artificial intelligence basic notions The data presented before shows the framework for the IA as well as the possibilities of development in the near future. According to Oxford dictionary the AI represent an informatic system able to perform actions that normally requires human intelligence. The visual perception, object recognition, image interpretation, speaking recognition, process optimization and others are a few examples of AI application. In a simpler manner the AI refers to systems or equipment that copy the human intelligence in order to perform some activities that can be improved iteratively, using the collected information. Machine learning is a subset of AI, based on systems that can learn or improve their performances according to processed data. Machine learning and AI are used frequently interchange but they represent different things. Nowadays we are surrounded by machine learning applications. We deal with them in on banks webpages, on online shopping, in social media. All these interactions are sustained by the machine learning algorithms so we can have a personalized and efficient experience. The algorithms are the software engines behind the machine learning. We use today 2 types of ML algorithms: ML surveyed and ML not surveyed. The difference between them is made by their learning mechanism used to make predictions. 1.3. Predictions for AI At National Governors Association Elon Musk warned us about significant changes on the work market in the near future, further to AI development and the replacement of the human workforce. This change is unavoidable because the AI is working faster, better, more precise than any human on the planet, and also cheaper, with no breaks or holidays. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 UE is aware of this concern and within CEDEFOP (The European Center for the Professional Development Learning) they underlined that the creation of the new innovative jobs remained behind the innovative cycles in which the AI/ industrial robots/ replaces easily the human work (5). If in the past the unqualified work was replaced by industrial development, today even the highest qualified tasks can be better performed by AI systems in different sectors such as health, law, finance or education. According to CEDEFOP, 4 jobs out of 10 in EU will be automatized and they will require digital competencies and interactions with AI systems. The jobs based on routine tasks will be the first to be automatized. CEDEFOP analyze the way to adapt the professional education to the actual needs in the industry [5]. Fig.1 The UE workers threatened by automatization risks, by industries [5]. All countries in EU will have to invest in life-time “robot-proof” programs for the entire population. Each person is responsible for his learning process and for being updated with new technology competencies. Today over 80% of the adult work force in EU need specific digital competencies in their jobs. However, 43% of them don’t have the basic level and one third of them don’t have any digital competence. According to Cedefop study on online vacancies, the most searched ability by European recruiters is the adaptability to change, mentioned in 3 out of 4 job announcements, study based on over 30 million of jobs posted. In this world where humans and robots have to work together the capacity to adapt to this environment, to embrace the change, is crucial for people. 2. Applications of Artificial Intelligence 2.1. Applications of AI in essential domains CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 The AI applications are very diversified and in a continuous expansion. Hereafter are some examples of the AI expansion in the modern society. Interaction Clients - Virtual assistants. More and more companies introduced the chatbots in their webpages. They process the natural language in order to understand the clients, to ask questions or give answers. They can improve themselves during these interactions. Netflix offers video streaming in over 190 countries and they implemented ML to personalize the interactions with 125 mil of users, and so they increased with 25% their clients in 2017. Text editing: Applications uses AI for short press releases. The AI created totally autonomous a short movie screen play, musical partitions or spot advertising for movies (Morgan film 2016). According to Century Fox IBM by using AI the time to make a movie trailer was reduced from 10-30 days to 24h. Cybersecurity. A report made by experts from 26 world institutions underlines the importance of cybersecurity in the main fields of health, transport, energy, as well as in National Security or military domains. Admiral Mike Rogers, Director in NSA- USAstated that AI and ML represent the foundation of cybersecurity in the future [6]. Using AI (deep reinforcement learning) we can detect malware attacks, Distributed Denial of Service that blocks websites (Ddos), spam, botnets e-mail phishing campaigns, ransomware, fake news etc. AI based programs can perform face recognition in real time on public domain, using survey cams. Up today the image recognition rate of success is up to 98% for AI, compared to 95% human capability, see Fig.2. Supasorn Suwajanakorn from Washington University developed with AI algorithms an application for 3D facial construction, capable of talking any text and having real facial expressions, so similar to humans that can hardly be differentiated from real people. For these video clips they used only photos or short video records, the edited material is showing the former US president Barrack Obama, keeping an invented speech [7]. Fig.2 The AI systems performance for image identification [6]. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 Medical field: the program Deep Patient made in Icahn School of Medicine in Mount Sinai, uses AI to identify the high risk at 80 diseases, in patients, one year before the disease starts. Medical Researches made in Lawrence J. Ellison Institute for Transformative Medicine made possible early detection in cancer using AI programs. NYU Langone’s Perlmutter Cancer center approved an AI program that cand detect the early signs of cancer, through radiology, before even the specialist could do it. There are many other domains changed by AI: in transport, in USA trucks operate without drivers, using instead AI. In the next 10 years 1,7 million drivers could lose their jobs in USA, being replaced by AI. In banks AI is used for investments analysis, to evaluate clients for loans. Malta uses drones to capture images from road accidents, etc. 2.2. Applications of AI in the welding field Welding technologies are key enabling technologies for over 85% products realized in the industrial sector. In Romania, the industry has a 24.1% contribution to the Gross Domestic Product (GDP), assuring jobs for about 22% of the national workforce (2019 national statistics), see Fig.3 [9]. The International Federation of Robotics (IFR) reported every year an increase of industrial robots’ number with about 15% and for 2020 is estimated that 521 thousand industrial robots will operate globally. Based on the data presented in Fig.3 it can be observed that Romania has the lowest hourly wage in UE, after Bulgaria. These statistics do not stimulate the companies to invest in automatization and AI, but in the future, this is expected to change, because the market competivity will force the companies to adapt to the new conditions. Fig. 3 The hourly wage in EU and industry contribution to Romania GDP [9]. During the welding operations are unfolded a series of metallurgical reactions and physical-chemical processes controlled by over 20 non-linear variables, like: chemical composition of parent and filler material, dilution factor, heat energy, t 8/5 time, metallic structure design, welding order, protection atmosphere properties etc. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 As result, the analytical mathematical models used to describe the welding phenomena are extremely complicated, having limited applicability impose by the initial conditions, accuracy and the applicability range for different material constants etc. The correlation between the welding process parameters and the output data like, mechanical properties of the weld, the microstructural features in the deposited metal and in HAZ, the welded structure behavior under load or different working conditions etc. represent an ideal environment for AI applications. One field of applications for AI in the welding area is the identification and classification of the welding imperfections in metallic welded structures using nondestructive testing methods (NDT). The correlation of welding parameters with the incidence of the welding imperfections represents a very important problem that affect the welding productivity and economics, but also it is important for the security and integrity of the metallic structures in exploitation. The process of welding imperfection analysis and classification using AI tools, present four stages: the image acquisition, the image pre-processing, the extraction of visual characteristics and the classification of the data obtained. The AI program has a neural network which is trained initially using several hundreds or thousands of images that contain different welding imperfections. After that, when in the network are introduced new images, different from the initial ones, the AI program is able to identify and classify new welding imperfections using the ML algorithms (ZFNet, VGGNet, GoogLeNet, ResNet etc.). The AI program performance is directly proportional with the initial size and quality of the database (images) used for training process. The self-learning ability of the neural network can be improved in time, based on the feedback of a qualified personnel that supervise the entire process. Another’s applications for AI in the welding area are based on the data processing in order to find different correlations between welding parameters using multidimensional regression. For instance, we are interested to find correlations between welding conditions and different weld features, like: mechanical properties, geometric parameters in weld cross section, depth of penetration, etc. These researches are performed more frequently in the last years, if a search with the “AI in welding” keywords is executed on the Elsevier or Springer database, we obtain over 3.000 articles [12]. Also, there are some studies regarding the implementation of AI systems designed to assist the welding personnel for coordination or even the welding workers in order to increase the welding control and quality by using optimized welding parameters: welding current, welding speed, arc length etc. [11,12]. The KEMPER GmbH company has implemented the AI systems in order to improve the welding working conditions and health securities. The company extended the concept “delayed cleaning start” used in the gas and fume exaustation from the welding space working. By these innovative measures, the KEMPER GmbH has prolonged the efficiency and time life of the filters used in specialized equipment’s, the air being purified and cleaned from dangerous fume gas and fine particles emissions [13]. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 3. Case study. CCT welding diagrams obtained by using AI programs The TTT diagrams, but especially the CCT diagrams are very useful resources in welding. Based on these diagrams, we can estimate the microstructure and mechanical properties of the deposited metal, in HAZ according to welding conditions. The welding processes are characterized by different heat input energy, the final chemical composition of the deposited material being determined by dilution that is usually between 10-35% and by chemical composition of the parent/filler materials, protection atmosphere properties, flux characteristics etc. The CCT diagrams are very dependent on the material chemical composition, therefore we do not have all the time a CCT diagram for any particular composition. Using AI programs, is possible to build CCT diagrams for a large range of chemical compositions. The CCT Editor is a software program designed to implement the neural network analysis in order to build new CCT/TTT diagrams. In the current stage of development, we can use the program to convert the data from scanned diagrams to numeric data, which can be used further to train the neural network, Fig.4. The input data from scientific literature, books or articles can be converted in a format that a neural network can work with. The data are organized in a large database, that contain information about chemical composition, hardness, microstructure characteristics and other welding parameters. Fig.4 Screen capture of the CCT Editor program. After the CCT Editor database has been populated with data from TTT and CCT diagrams for structural and others steels used in welding fabrication, we can use the CCT Editor AI module in order to train the integrated neural network, structured in three layers with 16, 8 and 1 neuron for output data (time). CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 Another solution is to use the data from CCT Editor database in TensorFlow, which is a more powerful application from Google, which is an open-source platform for ML and AI projects. Tensor Flow can be accessed on-line, is free and has the advantage that is very flexible in implementation, requiring only a few lines of codes. Practically, the CCT Editor has built-in several tools for processing the images obtained by scanning and is converting graphical format in pairs of data temperaturetime for each type of transformation. The start/end transformation curve for each domain is transposed in numerical data using the CCT Editor Toolbox. First, is performed an image calibration and then by clicking with the left mouse button on transformation curves, new points of data are added in the program database. The points defined can be relocated in new positions by click and drag technique, the program storing the new coordinates. When the mouse is moved over the diagram, the temperature and time in logarithmic coordinates are displayed in real time in the toolbox. The program will display on the working diagram the position of all points defined for each type of transformation. Each point has a label and a specific color, according to the transformation type. For martensitic transformation are allocated only three points, that defines two lines of transformation. For the others curves are allocated ten points for each transformation. Thus, in the first layer of the CCT Editor neural network are 16 neurons, 15 for chemical composition (C, Si, Cr, Co, Ni, V, P, Mn, Al, Mo, Cu. Ti, Nb, S, B) and the last input neuron is assigned for temperature. The hidden layer has 8 neurons and the third layer just one neuron for time, which stand for output validated data. In the Tensor Flow program is possible to build a different layers stucture, with more parameters according to the data collected in the CCT Editor. x1 Input data x2 w1 w2 x3 w3 ... xn wN Output data Σ - yp Fig.5 The analogy between the natural and artificial feedforward neuron. Model of signal propagation for the Npj neuron, where j= 1...Nh The signal that travels within a neuron of the neural network is presented in Fig.5. The input data or the input signal is collected from all exit neurons signals that belong to the previous layer. For the first layer, the input data is the chemical composition and temperature for each point stored in the database. The input signals in a neuron are summed, adding also a bias bk in the activation function argument, in order to compensate the null entries, relation (1). The neurons weights are adjusted during the training process, each neuron net input being obtained by sum of all entries using the relation (1): CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 𝑝 (1) 𝑦 = 𝑓 (∑ 𝑤𝑖 ∙ 𝑥𝑖 + 𝑏𝑘 ) 𝑖=1 The complex neural networks are structured in multilayers, with one of several intermediate layers. When the information or the signal is propagating from the input layer to the output layer, the network is called feed-forward. The activation function performs complex transformation of the input data, or we can have step functions. The activation function used in CCT Editor is a sigmoidal type, expressed in relation (2). 𝑦𝑖 (𝑁𝑝+1 ) = 1 1+𝑒 (2) −𝛽∙𝑦𝑖 (𝑁𝑝 ) From mathematical point of view, the network is adjusting the weights for each neuron by iterative operations on all layers of the network. The output signal of the network is compared to the reference value introduced in the initial stage. If the error in the output data is bigger than admitted one the adjusting process is repeated. The training stage of the neuronal network can be further optimized in order to obtain the best combination between the learning rate and the result’s accuracy. It is important to make a distinction between the real tendency of data variation and the noise in every phenomenon from real process. The neuronal network must identify through a proper selection of the network parameters the correct correlation within the input variables and ignore the noise from recordings. The back-propagation error method is an efficient way to train the neuronal network in order to find the neurons’ weighs. This method contains several stages: in the first one the propagation signal goes through the network from one layer to another until the exit, see relation (3). 𝑁𝑖𝑛𝑝𝑢𝑡 (3) 𝑦𝑝,𝑗 = 𝑓( ∑ 𝑤𝑖,𝑗 ∙ 𝑥𝑖 + 𝑏𝑗 ) 𝑖=1 𝛿𝑝𝑘 = (𝑡𝑝𝑘 − 𝑦𝑝𝑘 ) ∙ 𝑓 ′ (𝑦𝑝𝑘 ) (4) Where: Ninput number of inputs in neuron network on p layer Nh number of neurons on p layer Noutput number of neurons on exit layer The ypk output is compared to the tpk control value and the result is an pk error, see relation (4) where the f' is the derivative of the neuron activation function. In the next stage the errors are propagated from the exit layer to the input layer (back-propagation), resulting the change of neuron weights, minimizing the pk error for each neuron. CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 The learning rule for the connection between output neurons and the hidden ones is presented in equation (5): ∆𝑝𝑘 ∙ 𝑤𝑘𝑗 = 𝛽 ∙ 𝛿𝑝𝑘 ∙ 𝑦𝑝𝑗 𝑁 𝛿𝑝𝑗 = (∑ 𝑡𝑝𝑘 ∙ 𝑤𝑘𝑗 ) ∙ 𝑓 ′ (𝑦𝑝𝑗 ) (5) (6) 𝑘 ∆𝑝 ∙ 𝑤𝑖𝑗 = 𝛽 ∙ 𝛿𝑗 ∙ 𝑦𝑝𝑖 (7) - is the learning network parameter with values between 0 and 1 The pj errors for the neurons situated on the previous layers are calculated considering the errors on the output layers, pk using relation (6). After all errors are calculated all weights are readjusted using the relation (7) and the entire process is repeated forward. The iteration process of the weights adjusting is continued till the value of the final error is smaller than the set one. 4. Conclusions Based on data from specific literature and from those presented in some reports of specialized institutions like Cedefop, the following conclusions result: 1. The AI becomes an indispensable instrument in all sectors of activity. The automatization and the AI implementation in all industries and other fields will cause a significant loss of jobs; 2. The learning programs “robot proof” become a critical necessity in all developed countries. The work market requires a higher level of digital competencies necessary in all jobs; 3. One of the most required soft skill at hiring is the adaptability of the candidates, able to work in an environment where robots and humans must cooperate; 4. The IT technologies assisted by AI need a higher volume of relevant, accurate quality data; 5. The AI and ML solutions have to be scalable in order to answer to the changing needs of the companies. As for now the greatest part of the IA solutions are implemented by IT experts, requiring maintenance and configuration of the informatic systems. The next generation of AI programs will allow access to less experimented users; 6. The AI programs require a high level of computation power and in the absence of a cloud computing their access is limited; 7. AI application in the welding area is already a reality for many companies; 8. The CCT editor software is using AI to build TTT and CCT diagrams for welding, heat treatment and other applications; CONFERINŢA SUDURA 2021 Reşiţa 22-23 aprilie 2021 9. The data obtained by CCT Editor can be used in Google TensorFlow AI program; 10. It is important to collect data and make database in all industrial sectors so that they could be processed with AI programs. Bibliography 1. WhatIs, Robert Metcalfe, https://whatis.techtarget.com/definition/Robert-Metcalfe 2. McKinsey Global Institute, Notes From The AI Frontier. Insights From Hundreds Of Use Cases, April 2018. 3. Insider Intelligence, Sam Shead, Google's $500+ million purchase of DeepMind just got very interesting, https://www.businessinsider.com/googles-400-millionacquisition-of-deepmind-is-looking-good-2016-7, Jul 21, 2016. 4. Garry Kasparov, Chess, a Drosophila of reasoning,Science, Vol. 362, No. 6419, DOI 10.1126/science.aaw2221, 2018. 5. CEDEFOP, Centrul European pentru Dezvoltarea Formării Profesionale, Inteligența artificială sau cea umană?, Notă de informare Iunie 2019. 6. Future of Humanity Institute, University of Oxford, University of Cambridge, Center for a New American Security, a.o., The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation, Febr.2018. 7. Supasorn Suwajanakorn, Fake videos of real people — and how to spot them, TED talk, Jul.2018, 8. Natalie Kitroeff, Robots could replace 1.7 million American truckers in the next decade, Sept. 25, 2016. 9. Ziarul Financiar, Industria a adus în 2019 un sfert din valoarea adăugată în economia românească şi a angajat o cincime din forţa de muncă a României, Răzvan Botea 30.10.2020. 10. Haixing Zhu, Weimin Ge, Zhenzhong Liu, Deep Learning-Based Classification of Weld Surface Defects, Appl. Sci. 2019, 9, 3312; doi:10.3390/app9163312. 11. Martin A. Kesse , Eric Buah, Heikki Handroos, Godwin K. Ayetor, Development of an Artificial Intelligence Powered TIG Welding Algorithm for the Prediction of Bead Geometry for TIG Welding Processes using Hybrid Deep Learning, Metals 2020, 10, 451; doi:10.3390/met10040451. 12. E. A. Gyasia, H. Handroosa, P. Kaha, Survey on artificial intelligence (AI) applied in welding: A future scenario of the influence of AI on technological, economic, educational and social changes, Procedia Manufacturing 38, (2019) 702–714. 13. Andreas Effing, Self-learning extraction technology: Has AI already arrived in welding fume extraction?, https://safe-welding.com/self-learning-extractiontechnology-has-ai-already-arrived-in-welding-fume-extraction/