The difference between commercial AI and industrial AI, and the architecture analysis of industrial AI

“Industrial AI must be based on industry and attach importance to the role of people in order to better leverage the value of software, make machines smarter, and truly improve users’ production and operation efficiency.”

Today, in an era of surging concepts, "intelligence" has become very popular. Some people believe that "intelligent manufacturing is AI + manufacturing", while others worry that "AI will replace people and make many people unemployed". Others from the manufacturing site remain skeptical of AI because "AI is just a last resort to solve problems", and if you don't understand the industry, the role of AI is no different from that of manual labor. For all these reasons, people seem to have extremely high expectations for AI but are full of doubts. To be precise, it should be like the popular saying that "people often overestimate the current application of technology and underestimate the huge potential of technology in the future".

Considering that there are a lot of "unpredictability", "nonlinearity", "uncertainty" and no physical laws or chemical equations to follow in the manufacturing site. In the past few decades, many experts are cross-border experts in both control and artificial intelligence. They continue to use AI methods to solve control problems in the manufacturing site. Therefore, the application of AI in industry has always been on the way, but in the past it has been restricted by computing power costs and mathematical methods and has always been in the exploratory stage.

Today, considering the economic efficiency of computing power and the urgency of demand, people are placing more hopes on AI to solve problems that are difficult to solve in traditional manufacturing. However, we must understand the similarities and differences between industrial AI and commercial AI in order to bring its value to the industry and bring innovation and efficiency improvement. 

The difference between commercial AI and industrial AI

Indeed, AI has been widely used in commercial scenarios. Since commercial AI often processes high-dimensional data such as pictures, voice, and text, it contains more data, so the space that can be mined is naturally larger. In contrast, the data processed by industrial AI is more low-dimensional data such as temperature, pressure, and vibration, or small data, which makes it inherently different from commercial AI.

Figure 1: Industrial AI requirements

而工业还有一些必须予以考虑的情况，则是商业A I 通常不特别的要求，图1 表明了工业A I 的一些特别的需求, 其中有几个显著的特征要求：

01. Real-time/periodicity:

制造现场的控制、边缘计算任务通常是周期性的任务，这意味着A I 在工业场景里，从数据的采集、处理、传输、分析、应用都必须考虑其周期性特征。而另一方面，实时性等级也将会影响生产的品质和效率，因此，数据需要打上时间戳，并经由时间顺序等进行分析，而其推理和执行也需要考虑实时性和周期性这显著的特点。

02. Interoperability:

In industrial sites, due to heterogeneous networks and differences in control and edge data, it is necessary to consider interoperability in the architecture, that is, to be able to recognize each other's syntax and semantics in order to perform "computing" under homogeneous data.

03. Human-machine collaboration:

Since AI is good at dealing with those things that cannot be regularized (theorems, formulas, physical and chemical equations), this kind of experience, knowledge that exists in an implicit form, must find an object of learning, that is, human participation. Therefore, in practice, a large number of industrial AI applications are basically carried out in a supervised learning manner.

04. Explainability:

This involves the interaction between machines and humans. Since algorithms such as deep learning are more based on a "black box" approach to data training and model formation, there are potential risks. It lacks explainability, which leads to potential risks. 99% accuracy but 1% inaccuracy is difficult for the industry to accept, because it may mean a large loss of good products, or even unsafe (functional safety).

05. Scalability:

Due to the particularity of the vertical industry attributes of the industry, learning in specific fields is possible, but if these experiences and knowledge cannot be extended to other fields, the cost of AI cannot be effectively diluted. Therefore, how to have high scalability is an issue that industrial AI must consider during training and packaging.

Other requirements such as distribution, modularization and robustness are relatively common in manufacturing sites.

 Industrial AI applications, people are the key 

In the application of AI, do not ignore the role of people. "Technology determinism" often magnifies the power of technology, software, and algorithms, while ignoring the importance of people. This includes several important reasons:

1. Machines are learning from people

In fact, AI has gone through a long development process, going through the connectionism of neural networks to simulate the human brain (Brain) and the symbolism (Mind) to simulate the human way of thinking, including the behaviorism (Action) of human behavior feedback and then adjusting strategies. However, in the end, it returned to the "machine learning" stage where machines learn from humans. Intelligence starts with learning from humans.

人的智能，即，人的思维方式包含了非结构信息处理能力、直觉判断、自组织学习这几个显著特性，而这都是A I 所不具备的，因为机器最大的特点是算力，如果了解计算机的原理就知道，计算机仅在利用非常高速的计算能力来处理数据，以及大量数据存储和不知疲倦的算力。

2. Machine learning applications themselves require expert intervention

The second aspect of human participation in machine learning is that the application project process of "AI" requires human participation, that is, in defining feature values ​​and supervising the machine learning process, in industry, basically a "supervised" learning method is required, and people calibrate the learning effect to allow the machine to obtain "direction" - to know whether it is learning correctly. Third, in data learning, data preprocessing and dimensionality reduction require human participation. Therefore, AI in the industrial field is even more so, and human participation is needed to better combine it with demand, effectively process it, and establish effective data association and causal analysis.

Figure 2: Comparison of capabilities in collaboration between machines and humans

 Industrial AI Architecture 

With the development of current controller technology, PLC can also implement machine learning applications - in essence, AI applications are also "software" and can use X20P LC. Of course, considering the computing power, storage, and operating system, B&R recommends industrial PCs using Hypervisor technology because it can combine general operating systems such as Linux/Wi ndows with embedded RTOS in industrial sites. Processing control and computing tasks separately in multi-core processors is the fusion of distributed computing and centralized control architecture. In Figure 3, we can see that B&R's PC can achieve such application integration.

The algorithm can monitor the original PID process and seek the best data. Of course, static or dynamic data monitoring and historical data analysis can be used to obtain an optimized model for reasoning operations. The reasoning results can be directly given to the control system for I/O logic, motion control, process control (temperature, pressure, etc.), and can drive robots and conveying systems to achieve intelligent execution.

Figure 3: B&R's overall architecture for implementing machine learning applications

 Application examples 

Through these two simple explanations, we can see that in industrial scenarios, machine learning has huge potential in defect detection, parameter optimization, quality improvement and other areas.

01. Tire defect detection and analysis

In traditional tire monitoring, since the tire uses an X-ray machine to detect the wire ring forming process and obtain X-ray films, it is usually determined by humans whether it is qualified, what kind of defects it has, and the cause of the defect. It is then classified into warehouses and fed back to the production process.

As shown in Figure 4, after obtaining the imaging information of the tire through the X-ray machine, the learning system will analyze the defects instead of humans, because during the winding process of the wire ring, there are many defects caused by the machine status or abnormalities in production, such as winding layer deviation, foreign matter, uneven arrangement, overlap, protrusion, warping, abnormal bending that interferes with other steel wires, bubbles, etc.; and because of the characteristics of the tire's internal cord, wire ring, and rubber material, there will be anisotropic textures after its X-ray imaging, and these textures will interfere with the judgment of the image, causing the system to misjudge.

Figure 4: Tire defect analysis based on deep learning

The deep learning-based method can extract features from the relevant elements of these different defects, use filtering algorithms to effectively obtain defect features from interference factors in the image, find defect points from these complex images, cluster the defect points, calibrate their size specifications, and ultimately judge the quality and grade of the tire, decide whether to classify it as scrap, and provide feedback to the manufacturing system for continuous improvement.

02. Intelligent adjustment of printing pressure

In flexographic printing, the embossing force is a factor that has a great impact on the quality, especially during the startup phase, which often relies on the experience of the master, which often results in a large startup waste (100-200 meters). The place where the machine relies on human experience is where the machine can learn.

After each color group, the product image quality is visually inspected to obtain the characteristic information of the image, which is then transmitted to the central controller via the POWE RLINK network. The intelligent reasoning algorithm on the controller can analyze these qualities and establish the relationship between the image and the control quantity through the convolutional neural network (CNN), and then infer the pressure adjustment amount between the publishing roller/printing roller.