3). Industry needs AI with high precision

For industrial applications, model accuracy can be a more important issue – hyperparameter problems, model structure, data preprocessing, etc., can affect model accuracy. And this precision will make inference, prediction bias – industrial tasks it is often fault tolerant space is relatively small. For example, the deviation of time and position, such as: 0.1mm is too harsh for the positioning training of commercial humanoid robots – but it is entry-level for the coordination processing of industrial robots. The same is true in time granularity, for time-critical tasks, time in precision, real-time, are relatively high requirements.

05701-A-0301 Compact is the model to be small, because there are hardware computing power and cost limitations – usually in embedded systems running model inference, it is impossible to use too strong a processor.

4). Small sample model training

In industrial scenarios, this problem may be more prominent, because usually machine learning requires a large sample size, but if a wind turbine has a large number of fault signals, the machine should not be sold to users. Including quality issues, there are fewer parameters to learn. Therefore, the industry must consider this scenario, using more small sample training methods, such as contrast method, τdistributed sample processing.

5) Combining data and mechanism

This should be a good place for automation companies. However, this requires a relatively easy-to-implement interaction on the interface of the software.

In fact, OPC UA is also establishing this connection and model interaction specifications, AAS asset management shell, industry information model, respectively, used to deal with how to establish interactive interfaces between different software platforms, its main purpose is to reduce the convenience of the interface, to avoid the need for manual interface development. AAS and industry information model are in data collection. These are mainly to solve the engineering creation, operation and maintenance of the integration of data and mechanism.

05701-A-0301

The combination of core data and mechanism is to give play to their respective advantages. Because, in real engineering development, and even in real production, the mechanism does not fully understand the optimal system, or the more efficient parameter set – the operator itself lacks this knowledge. Therefore, AI intervention is mainly in the convergence of better parameters, including the convergence direction of time and cost, which will have a certain effect. However, AI does not directly participate in control, but only gives better parameters as offline learning, of course, back to the issue of interpretability, which also needs to be judged by people to confirm whether it is run.

6) Human-machine combined AI

05701-A-0301 Human-machine integration, because there is an important context here, namely, what is the AI learning?

Because AI actually, in addition to AI mining the potential in the data, in some scenarios, AI itself is also learning from people – for example, in the offset press, by sampling the quality of printed matter, digital scanning method to the computer. Then, the computer simultaneously observes the process of how the technician adjusts the ink balance, and the feedback parameters learn how to improve the printing quality.

This demand will also be more in the future, because in many fields such as aerospace, semiconductor field of senior technical personnel is extremely expensive – even to reach more than a million income, and such people are particularly few, how to refine their rich experience, which is also the industry needs to use AI to solve the problem.

Therefore, in summary, industrial AI needs to combine the characteristics of the industry itself, and develop specially focused functions in tool development, data connection, processing, feature selection, training, parameter adjustment, interface with real-time tasks, cloud connection, etc., to make users easy to use and meet the industrial characteristics of AI.