–> Loop regulation – Loop regulation also lies in the algorithm.

ALR121-S00 In the final analysis, the essence of industrial control is mathematics, whether it is Gokonomov’s system theory – laid the various mathematical models of artificial intelligence today, Wiener’s cybernetics, Shannon’s information theory, in fact, are mathematical problems, the industrial field is a mathematical problem, this mathematical problem runs on the PLC, PC or the cloud. It doesn’t matter, what to do it with is determined by economy.

However, from another point of view, automation is constantly using IT technology to expand itself, including the use of Ethernet, the application of Web technology in the controller, new FPGA chips, MATLAB/Simulink, FMU/FMI are integrated into the automation platform to expand its own capabilities. It seems that the industrial Internet always wants to subvert the manufacturing industry, but it seems that the industrial people have not thought about who to subvert.

How many models are accumulated?

ALR121-S00 Many industrial Internet platforms think that the term “digital twin” is interesting and that artificial intelligence is promising, but I think they may have overlooked several issues:

(1) Do we have a model?

In a manufacturing industry that has developed for many years through “mapping” or “reverse engineering”, the original design is actually very few, because the original design needs to be modeled and tested and verified, and we simplify the process that costs the most research and development investment, and directly map the already verified design, which really reduces the cost and shortens the research and development cycle. However, this also makes us have no basic model, know not know Why – know “What”, but do not know “why” – this model foundation is weak, will lead to all learning algorithms must “idle”, can not be verified.

ALR121-S00 In 2018, two companies in Shanghai paid Dassault 10 million for infringement, they were unwilling to buy a legitimate software platform to use, thinking it was too expensive, but in fact, these modeling and simulation software design purposes are to reduce the risk of research and development, if you really follow the original design, build a physical model, prototype, test, modify the cycle. Then the real research and development is extremely expensive, this time you find that the use of building imitation can really reduce the cost, you will feel that buying these genuine software is very cost-effective, I think they are not willing to buy the reason is that they are only for the “drawing” function – because the supply chain needs to use this software drawings.

How do you implement a digital twin without a model? How to do machine learning?

ALR121-S00 Some people think that machine learning can solve many problems and intend to use machine learning to replace existing control models

According to the topic that Academician Chai once talked about before, your learning in the Internet application scenario is a lot of big data small problems, but in the industry is small data, big problems, and interpretability has become a barrier – you say that this learning can, but can it be explained? Are there any potential risks? If 0.1% is uncertain, it is also unacceptable for industry.

Many people claim to change the existing model of industry through learning, you must understand that the “traditional” models are actually the “most economical”, the cost of learning models is relatively high, and machine learning models mainly solve the “nonlinear” part, that is, the observer’s observation of the existing control model. The part of the task that cannot be differentiated is learned – so in the case that a single control system has reached a local optimal, the industry has been looking for more complex dynamics under the optimal, which was previously insufficient, but, Today’s chip, storage, and network technologies make this possible – so the contribution of IT is the contribution of tools and platforms – just as many people think that the biggest advances in Western medicine today come from improving the level of measurement tools, not just the contribution of models, because industry has been doing this for a very long time.

ALR121-S00 This is also my more strange, those who really provide platforms such as Microsoft, Huawei, these themselves have a very strong platform, research and development support of the enterprise is very low-key, positioning their role in the platform and tools.

However, many industrial Internet platforms – in particular, say they belong to the “platform” – I think these platforms may only be Microsoft, IBM, Huawei their “system integration” role is more appropriate, if it is this role, it must go deep into the application scenario in the industry, then there are many problems to be solved

ALR121-S00 (1) Do you know how to analyze complex parameters? You may not even know what these parameters are, but once you enter the field, you have no possibility to learn without a seasoned industry process expert.

(2) What is your model? You know, the open-source learning code downloaded from the Internet? If this can be played, then why don’t the manufacturing process experts play themselves, and you have to play for me?

Is data really that easy to get

This is a very real problem, and I have written before about the pseudo-proposition of predictive maintenance, that is, if the machine has a lot of failure data, then the machine may be eliminated from the customer.

The second problem is that if you can collect the fault signal, the fault data may be small – and the corresponding correlation data is not complete, the model can not learn the real cause, and this requires a lot of experience.

Therefore, digital learning must be combined with actual models, and in this respect, enterprises with the industrial genes of autonomous models are qualified, but they lack digital tools.

However, if you are a platform maker, you have a hammer but no nails can be more awkward

How long can the industrial Internet platform of three nothings live?

When the wind power market is more lively, there are 89 domestic host manufacturers (around 2012), at that time we said that the industry will eventually survive about 15, now, this is true, and the photovoltaic industry in the hot time, that scene is to do shoes, do socks, lighters are swarmed to do photovoltaic, Too profitable. Now what?

Right, we mainly want to say that the platform with “standards”, “tools” and “models” can really empower the manufacturing industry in the future and bring about transformation and efficiency improvement, otherwise, it is the “risk” in venture capital rather than “investment”.