Shulou(Shulou.com)06/03 Report--

This paper briefly introduces the principle, technical classification and technical characteristics of artificial intelligence technology, enumerates some applications of artificial intelligence technology in typical industrial scenes, and puts forward some suggestions on how to build such a set of tools or platforms.

AlphaGo was born in 2016, and artificial intelligence and machine learning technology were famous for a time.

With the development of time, the application of artificial intelligence in engineering has become more and more mature, and a series of applications such as automatic license plate recognition, intelligent customer service robot, advertising recommendation and so on have been successful in engineering and business.

In the industrial field, artificial intelligence has not stopped the pace of development, predictive maintenance, quality control, intelligent scheduling and other fields have been exploring the process of project landing and commercial feasibility has made considerable progress. A large number of technology giants and experts predict that artificial intelligence will bring about the fourth revolution, which will change our work and life from the bottom after the agricultural revolution, the industrial revolution and the information revolution.

As one of the most popular tuyere areas, there is still a big gap between the depth and breadth of artificial intelligence application in the industrial scene and the consumption field.

Because of the poor versatility of the scene, the high cost of investment and construction and other reasons, commercial use is still limited to some high-end manufacturing scenarios, unable to achieve the same level of popularity of systems such as OA, ERP, CRM and so on.

For the product explorers of AI+ industrial Internet integration, the next key issue is the landing and commercialization of specific scenarios.

In this paper, from the perspective of popular science, I summarize and sort out the artificial intelligence technology I have learned and the application in the industrial scene. If there is something wrong, I also ask all kinds of experts to correct it.

First, understanding artificial intelligence 1. Take machine learning as an example

From the operation of the economy to the fact that the apple falls on the "hapless" Newton, the development of everything in the world is driven by objective laws. It's probably a model like this:

As far as the apple smashing on the head is concerned, gravity and Newtonian mechanics are the objective laws that drive the development of things, through which we can accurately judge the specific results such as how many seconds it takes for the apple to fall and how much momentum it takes to hit the head.

For economic problems, this model will be much more complicated.

For example, under the influence of the epidemic, we want to discuss the rise and fall of house prices in Chengdu. Although we can immediately judge that house prices will be dominated by objective laws, including economic models such as the balance of supply and demand, it is obvious that even if we know all the detailed data such as the progress of the epidemic, government investment plans, and the new registered permanent residence population in the five districts of the main city, we are still unable to accurately judge how much house prices will rise or fall in the next six months and one year.

This reflects a problem: the more complex the law of system operation, the more difficult it is to summarize the objective law by means of induction and derivation.

Machine learning technology enables computers to summarize their inherent laws through training from a large number of cases.

In the book artificial Intelligence written by Kai-Fu Lee, CEO of Innovation works, there is a definition of machine learning:

Machine learning. It is a process of using mathematical models to model specific problems in the real world in order to solve similar problems in this field. Such technical characteristics make it difficult to solve or solve high-cost problems with traditional methods in our life and production.

two。 Classification of artificial Intelligence

In the 1990s, the algorithm trend based on probability and statistical modeling, learning and computing began to occupy the mainstream. At the same time, the research of artificial intelligence is gradually divided into several major disciplines:

Computer vision: make computers understand the world; natural language understanding and communication (including speech recognition, synthesis, including dialogue): make computers understand and communicate with the world; machine learning (various statistical modeling, analytical tools and computational methods), such as today's popular deep learning and AlphaGo related reinforcement learning (Re-enforcement Learning) are all branches of this direction. Cognition and reasoning (including all kinds of physical and social common sense), let computers learn to think; robotics: including machinery, control, design, motion planning, task planning, etc.; game and ethics (mainly the study of multi-agent interaction, confrontation and cooperation, robotics and social integration and other issues). 3. Elements of artificial intelligence

Artificial intelligence needs at least three key elements to solve a specific project:

Huge amount of data: the computer can not understand and recognize the laws of the objective world, it needs huge amounts of data as samples for training: face recognition requires the computer to watch a large number of face photos, and machine fault identification and prediction also need to read massive monitoring data. Sufficient math: artificial intelligence has ushered in a craze again in recent years, thanks in large part to the development of computer chip technology. According to Nvidia CEO Huang Renxun, GPU performance increases 1000 times every 10 years, far exceeding Moore's Law. Even so, the efficiency of artificial intelligence computing is still far lower than that of the human brain, and a large number of chips are needed to provide computing support. Suitable model: there are dozens of algorithm models for machine learning, such as supervised learning, unsupervised learning and so on. Each model has a large number of parameters to configure. The accuracy of machine judgment highly depends on the selection of the correct algorithm and parameter configuration.

The above conditions are the basic requirements for artificial intelligence to perform a successful task. Amazon's Principal Scientist Shawe likens this to:

If a successful artificial intelligence algorithm is compared to a combat force, the data is the food, the computing power is the force, and the model is the strategy of strategic peace and warfare # skill command; the importance of strategic peace and warfare is needless to say, but without food and troops, the best strategy is just a castle in the air. Computing power can be understood as a force, and only with a strong force can we have the mobility and possibility to realize the strategy.

4. What are the advantages and disadvantages of artificial intelligence in industrial applications

advantage

The first is compatibility with the problems solved. The machine repair fitter analyzes the equipment fault, and the heat treatment engineering controls the process parameters, which are not related to each other in skill and experience; while using machine learning methods to solve the above two problems, the development resources and technical realization path are almost the same. Secondly, AI is loyal to the objective. People's summary judgment is often limited, and it is easy to ignore some secondary factors. On the other hand, artificial intelligence does not explain the law from the perspective of physical meaning or social meaning, it is only responsible for the samples and results, so that it can objectively reflect the missing or misjudged contents of human beings through positive analysis.

Inferior position

First, it is highly dependent on the quantity and quality of data. Humans only need a few simple pictures to identify different breeds of cats, while a machine learning system needs to read tens of thousands of pictures to do so. In addition, the data we obtain in engineering is usually heterogeneous, which means that in order to form a machine learning readable data source, we need to invest a lot of engineering resources to manage the data, which will greatly increase the cost. Secondly, the quality of artificial intelligence tasks is highly related to the skill level of employees. Among the three elements mentioned above, in addition to computing power, data and models all need specific practitioners. However, due to the imbalance between supply and demand, a large number of algorithm engineers in the market still stay at the level and level of adjusting database and parameters. to exaggerate, "the parameters basically depend on the Mongolia, the results basically depend on the test, and the effect basically depends on blowing". Finally, there is the cost. In the current hot days of artificial intelligence, both the cost of setting up a data center and the cost of recruiting algorithm engineers remain high, which leads to a backward shift of revenue balance and reduces the possibility of commercial use. Second, the scenario of artificial intelligence in industrial application 1. Industrial scenes and pain points

Industry is composed of many scenes, which can be decomposed layer by layer from industry dimensions (machining, electronic products, chemical industry, etc.) and product life cycle dimensions (design, manufacturing, sales, operation and maintenance).

Even every subdivided scenario, such as the maintenance service for machine tools, is a complex task involving multiple people, and there are unique pain points in this scenario:

It is difficult to find the right time for equipment maintenance; it is difficult to determine the cause of equipment failure; the knowledge of equipment fault maintenance is in the minds of the old fitters, and when faced with equipment that can not be repaired, they can only find maintenance resources through contacts and experience; similarly, in the scenario of production planning, there is the problem of planning and scheduling; for the scene of safety management, there are hidden trouble identification and troubleshooting problems.

The enterprise is profit-seeking, and all the problems (including the problems of use experience, efficiency and accuracy) are the real pain points as long as they affect the income of the enterprise.

Now the above scenes are basically dealt with by people, and the maintenance cycle is judged by the workers' statistics desk, and the identification of hidden dangers also requires a large number of personnel to inspect the scene, or keep an eye on it in the monitoring room; otherwise, it is easy to miss the hidden dangers.

If the scheduling is unreasonable, it will lead to idle resources, scheduling time is too long, put in too much resources; if something goes wrong, small rectification, large shutdown responsibility.

two。 Scene classification

In the industrial scene, there are not only repetitive and common sense mental work, but also complex mental work such as optimization.

According to the goal of the problem that needs to be solved, we can divide the industrial scenario AI application into the following categories:

The optimal problems of the system: production planning and scheduling, structure optimization, process parameter optimization, etc.; problems of classification or identification: identification of hidden dangers of safety, judgment of fault types, identification of quality problems, etc.; problems of recommendation and prediction: prediction of machine maintenance, situation prediction, auxiliary decision-making, etc.; problems of knowledge management: knowledge identification and maintenance, application of customer service robot, etc. 3. Analysis of typical scenarios

Because the ways of technology implementation are similar, take a brief analysis from each category, and the others will not be expanded.

1) Optimization of process parameters

In terms of demand effectiveness

Process parameter optimization is an urgent need with high priority for many enterprises, especially in the field of process chemical industry.

Take chestnut as an example, the production process of chemical products such as nitrocellulose is similar to "cooking", the amount of "material" (raw material) and the control of "heat" (temperature, pressure, stirring, etc.). May have an impact on the results (uniformity, composition content, etc.), but it is difficult to accurately judge the quantitative impact of various factors on the results, it is essentially a "black box" system.

The traditional practice is to keep trying, to try a new product, to try another batch, and to put into production according to this method until the quality loss of one batch is acceptable.

Therefore, the problem of process parameter optimization needs to be solved urgently by artificial intelligence technology.

In terms of technical realizability

Usually a large number of batches of past production are analyzed, in which the feeding and process parameter data are taken as independent variables and the quality data as dependent variables; training is carried out and a model is established to find its inherent quantitative relationship.

From a commercial value point of view

Even if it is the same product and the same process, there are great differences in temperature and humidity in different environments, such as the chemical enterprises in Hainan and Gansu. Some of the processes may reduce the accuracy or failure of the algorithm due to the change of the external environment, and the model needs to be customized for each specific scene, so it is difficult to reuse.

At the same time, the optimization of process parameters is obviously restricted by production scale and batches-the larger the production scale, the higher the output of a single batch, the higher the unit value of the product or the higher the risk of quality loss, the more feasible it is to use AI for process optimization.

Many related success stories can be found on the Internet.

The following case comes from a speech by Wu Hequan, academician of the Chinese Academy of Engineering.

Taiwan Sinosteel, they introduced IBM's Power AI solution to analyze defects in the rolling process. In order to roll 27 tons of billets to 0.5mm finished products and predict and analyze defects in the process, they collected more than 7000 batches of product data in the past year. After data cleaning, characteristic data that may affect product quality are selected and converted into data that can be used by machine learning.

Of these data, 80% are used for study and 20% for testing. Then they designed four mathematical models to see which model is more in line with the actual situation. Finally, they analyzed more than 2000 data generated by a product line according to the model and found that the pressure in the furnace had the greatest impact on the defects. Finally, Sinosteel has made good improvements in human resources and billet quality, and the cost has been greatly reduced. "

2) Identification of security risks

From a demand point of view

Many dangerous behaviors in the factory (such as illegal operation of equipment, fighting, etc.) can only be detected through on-site inspection and video surveillance. If they can be identified by video images, the hidden dangers can be found and dealt with in real time.

In terms of technical realizability

The general flow of its implementation is to do an example of → matching algorithm → training → embedded framework program for application. There are two kinds of embedded framework programs, one is the embedded program embedded in the hardware, and the other is the server-side web program. In addition, license plate recognition, face recognition and security hazard recognition are highly dependent on the scene.

The same behavior, different angles and different backgrounds, may lead to misjudgment, which means that it needs to be customized according to the scene.

From a commercial value point of view

Limited to dangerous behavior scenarios are extremely complex, can not be reused and need to be customized, it is difficult to achieve similar license plate recognition system production in the short term, resulting in its available scenarios are actually very limited.

3) equipment maintenance forecast

From a demand point of view

The equipment of the enterprise is similar to the vehicle, which needs regular maintenance and replacement of parts. In order to avoid liability, general car manufacturers require maintenance of 5,000 or 10,000 kilometers once, but most old drivers do not follow this, and they will choose the appropriate maintenance time according to the driving conditions.

The machinery and equipment used in production are usually maintained according to man-hours, and some enterprises can not even count the working hours (different upstream and downstream standards or lack of means), resulting in missed maintenance time and frequent equipment failures. If the maintenance suggestions can be given according to the actual working conditions, this problem can be solved effectively.

In terms of technical realizability

The technical realization path is roughly the same as the process parameter optimization, except that the independent variable becomes the detected working condition parameter of the equipment, and the dependent variable becomes the fault statistical data set.

From a commercial value point of view

The problems it faces are the same as the process parameter optimization, which is limited by the non-reuse, so it needs a large number of similar equipment to reduce the marginal cost. Once successfully applied, the cost of equipment operation and maintenance can be greatly saved.

The following cases are from Foxconn:

Because the milling cutter carries out discontinuous cutting at high speed, the tool wear is rapid and difficult to monitor, and the tool wear seriously affects the machining accuracy and product quality. In view of the difficulty of on-line prediction of tool wear in high-speed milling, a new method of tool wear prediction in high-speed milling based on depth learning is proposed. The milling force signal is extracted by wavelet packet transform, and the energy distribution in different frequency bands is used as the initial eigenvector. The sparse self-coding network is studied by unsupervised learning, and the single-layer network stack is formed into a deep neural network. Finally, the whole depth network is fine-tuned by supervised learning, and the prediction model of milling tool wear is established. The experimental results show that the accuracy of the proposed method for tool wear state prediction is 93.038%.

4) knowledge identification and maintenance

From a demand point of view

The traditional knowledge management software depends on the design of the database, and the system can not recognize the semantics, so we can only add, delete, modify and search in a system according to the preset data structure. The time at which knowledge is generated is often different from the time it is entered, and we often have to repeat it.

For example, if a product fails the test and needs to make a design change, it will usually take a change application, which will be approved by the project leader, the chief engineer and so on. If we can directly analyze the semantic of the approval materials, directly abstract the elements that failed the test, the ways and methods of change and other elements, there is no need for artificial input. The next time you encounter a similar problem, the relevant engineer can query the solution.

In terms of technical realizability

We can use natural language recognition technology to obtain a large number of texts directly from the business system, parse out the main elements of event execution subject, execution object, execution method, execution result, and then quantify the semantics of each element.

From a commercial value point of view

Both the technical knowledge base of enterprises and the application of service robots can reduce the dependence on personnel training and quality level. Combined with PDM, GIT and other software, it can form a complete knowledge management system of the enterprise.

Third, how do we build the application of artificial intelligence in industry

Attach importance to the construction of basic conditions, data first.

As has been analyzed in the previous case, the quantity and quality of sample data is a prerequisite to ensure the completion of artificial intelligence tasks. Fault prediction needs a lot of working condition data, and dangerous behavior identification needs a large number of image examples.

For enterprises, it is necessary to have basic conditions such as automation and digitalization.

For the industry and the government, it is necessary for the whole industry or the whole region to have the basic foundation of automation and digitalization. The key to the success of AI in the consumer field is that a large number of applications have been buried and a large number of behavior data have been obtained.

Define their respective positions and participate in the division of labor. The many industrial scenes, the deep complexity and the high industry barriers mean that this is not a winner-take-all monopoly market, which requires the construction of ecology and multi-party participation.

The era when everyone engaged in AI is not because human evolution has become smarter, but because of the birth of AI tools such as tesorflow and opencv. The high encapsulation of the algorithm enables us to focus on solving business problems, and the algorithm has changed from an esoteric mathematical problem to a basic operation of "adjusting database and parameters" + business insight.

The following figure shows the operation page of tensorflow-the front-end configuration page is highly mature, and the configuration of the algorithm can be achieved by dragging and dropping the module, even without the need to call the library.

In the future, the division of labor in the industrial AI field will be a high probability event, and the participants mainly include the underlying framework developers and scene contributors.

Framework developers have stronger software and AI technology background, they build industry-wide take-all application frameworks, and traditional service providers in various industries build scenarios based on the framework. In this way, the enterprise can obtain the scene at a lower cost, and the market does its own job to maximize the value.

Respect the business and deepen the business.

Compared with the rapid development of technology and business model of AI and the Internet, the development of industry is relatively slow and cumbersome; however, the knowledge and experience of industrial design, process and manufacturing process management are summarized by various industrial countries after hundreds of years of practice, and the processing methods of demand and business in the consumer field can not be directly reused in the industrial scene.

Whether industrial Internet or industrial AI, product managers need to have an in-depth understanding of industrial scenarios and businesses.

IV. Summary

Artificial intelligence technology has unique advantages in solving many industrial problem scenarios, but at present, there is a lack of in-depth integration of industrial scenes and artificial intelligence, and the scope of application is even less than 1% of the consumption field.

For industrial Internet practitioners, with the goal of engineering and commercialization, they should also pay close attention to the landing of business and scenarios, and also need to explore ways to reduce the marginal cost of scenario development.

Http://blog.sina.com.cn/s/blog_cfa68e330102zvrg.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.