Embedded Machine Learning enables Artificial Intelligent Machines - 4 / 10

This is the fourth in a series of 10 articles inspired by a Quick Guide on Machine Learning, initially published by the German Mechanical Engineering Industry Association (VDMA). The Quick Guide contents is based on articles from several persons, eventually edited by a team of 4, including the writer of this series.

Applications in mechanical engineering and plant construction

In this article we present some typical ML applications in mechanical engineering. Each application is briefly described and a possible implementation strategy is explained. Benefits, required skills as well as effort, opportunities and risks are listed. The technical implementation of these use cases is discussed in more detail in a later chapter.

"Human-like Machine Vision" 

The evaluation of surfaces with textures is one of the tasks in which classical image processing systems reach their limits. The human eye, on the other hand, can recognize textures, patterns, objects and structures and reliably evaluate and classify them visually after only a short training period. Using just a few examples, people learn to distinguish permissible variations from defects - even with natural products in which no two parts are the same. 

(c) i-mation GmbH

In combination with human-like image processing, all types of sensors can be used in the imaging process: 2D, 3D, ultrasound, X-ray and shape from shading. The ML application is based on a training phase with good parts; conventional image processing applications, on the other hand, usually require comprehensive error catalogs to be taken into account. With ML, the desired result itself is the measure, not the deviation from it. 

The process-safe solution for such tasks are ML-based image processing systems, which are specially developed and optimized for the industrial analysis of images.

The use of such systems based on ML opens up further application possibilities for process-reliable, automated inspection with very high detection performance. Where classic vision systems reach their limits and human evaluation is the best solution despite all limitations and risks, "human-like image processing", based on MLA algorithms, currently offers a state-of-the-art solution. Furthermore, new products can be learned without great effort, and even new, unknown features can be recognized without costly error libraries. This results in significantly reduced development and product launch times.  

In addition to experience in conventional image processing and the design of optical camera systems, modeling does not require any further software development or an understanding of the algorithms. 

Despite all possibilities ML has limits and is not always the means of choice.  

Limiting factors can be in particular:  

• Accessibility of the required training data 
• Availability and degree of determination of expert feedback in the evaluation of results 
• Image resolution and the size of the files to be transmitted or stored 
• Susceptibility to targeted manipulation or sabotage of the ML system.  

 In some cases the combination with conventional image processing is also necessary. In order to use the potential of ML systems profitably, interdisciplinary know-how, a comprehensive portfolio of image processing solutions and careful consideration of opportunities and risks are required. 

Adaptive control for process optimization 

In this application, the optimization of the start-up of a web offset printing press is explained - representing the optimization of complex physical machine processes. The ramp-up is a complex process due to numerous parameters and influencing variables. The most important factor here is the fine adjustment of the optical solid density, which must be manually parameterized before each production start. The effect of changing the parameters, however, can only be assessed after the press has been run through completely; this process is therefore associated with a dead time during which reject quality may be produced.

At the same time, other parameters influence the print result in unknown form: Consumables, physical parameters and machine condition. 

(c) Fraunhofer IGCV

Such technical systems, whose behavior is influenced by numerous variables and unknown correlations, can only be modelled with physical formulas with difficulty - they often elude process optimization. In this case, machine learning methods can be helpful, which can learn the system behavior and then make predictions about the process - this would be the so-called adaptive control. 

The dead time is bridged with a model-based adaptive control. The relationship between sensor data and the quality of the solid tone density is learned from historical data and used as a feedback variable in the control system. This allows the process parameters to be adjusted even before density measurement values are available. The use of predictive control significantly increased productivity and resource efficiency. On average, it was possible to reduce scrap in the start-up process by 37% and the time required by 39%. This type of control is called predictive control. 

The development of an adaptive control primarily requires a deep understanding of the process. In addition, knowledge of the use of machine learning methods for the regression of time series is required. The procedure for model-based predictive control can be applied to similar problems. An approach with machine learning methods is always promising if many measurable variables influence the process in an unknown way. As in the other application cases, a sufficient amount of data is absolutely necessary for adaptive control. 

Intelligent proposal management  

In the last decade, the trend towards product individualisation has continued. Within the framework of on-demand production or batch size 1 production, the variety of product variants is increasing disproportionately and thus the complexity. This complexity is also reflected in the available machine configurations. 

A large number of machine models with options as well as dependencies between the options quickly leads to a confusing variety of possibilities for manufacturers and customers.  

This diversity in machine configuration places high demands on the preparation of quotations. In the case of highly complex machines, the quotation process with product characteristics and price determination can take several weeks. This often leads to delays in the preparation of quotations and may be detrimental to the purchase process. An intelligent quotation system can automate the preparation of quotations in parts and thus considerably speed up and reduce the price. 

Information and extensive data about already created quotations, machine configurations and prices can be used for the semi-automatic creation of future quotations. Assuming that a similarly configured product leads to a similar cost structure, ML algorithms are used to train a model that learns the relationship between machine configuration and cost. This model is then used to estimate the cost of a machine configuration and to create an initial quote based on this estimate. Quickly generating quotes can increase the likelihood of sales and thus increase sales. Further advantages: The quotation process is simplified and the potential for errors is reduced. 

 Automation in the preparation of quotations 

As with the other use cases, intelligent quotation requires a sufficient set of historical data on machine configuration and costs. The quotations must be available in a structured form that can be read in by an algorithm. The biggest effort is to clean up the data and normalize the historical databases. For the development of the prediction model one should trust in the abilities of a Data Scientist. 

In addition, other application areas are suitable for ML in the bidding process, e.g: 

• Search for similar configurations: ML algorithms can be used to determine the machine configuration that leads to similar costs/results.
• Guided configuration: Based on historical data, permissible configuration variants are determined and successively suggested to the bidder.
• Augmented configuration: Based on the current machine configuration, the most popular additional machine options are suggested.
• Avoid or reduce over-specification in machine configuration: MLA algorithms can be used to determine the options relevant to the cost/performance ratio.  

 Data-driven innovation  

At the beginning of an Industrial Analytics project to improve the Overall Equipment Efficiency (availability, performance, quality), the optimization potential of the plant or machine is determined. From this, the commercial value of the data can be derived as part of the general business understanding. In the collection phase data from automation components and field devices is collected. In the analysis phase, the data is processed, modelled and analyzed using ML algorithms. If a problem is analyzed and understood, it is solved - if possible by a modification within the production. If the problem cannot be solved permanently, it should at least be recognized with foresight; for this purpose, a solution is sought by means of a streaming analytics implementation, e.g. an anomaly detection, and made available in the implementation phase.  

Experience shows that machine builders and plant operators typically first make an existing plant or machine available or improve quality with a reactive IndustrialAnalytics project. Only when this approach has been successful, do they consider transferring it to the next generation of performance improvement development. The approach is basically the same, but not the goals.  

In the classical order "Algorithms -> Data -> Decisions", the overall system effectiveness cannot be better than the human being who programs it. ML algorithms, applied to large quantities of production data, detect causalities that were previously hidden from the plant operator or machine builder, but which improve overall plant effectiveness. This not only reactively improves quality and availability, but also proactively improves the performance of future machines. 

A fault is detected. (Source: Softing GmbH) 

Today, the Industrial Analytics solution can only suggest possible correlations in the data provided. The domain expert of the machine builder or plant operator evaluates the correlations identified by the ML algorithms as "random" or as actual causality. On the basis of these correlations in the data, which the domain expert has identified as causality, decisions are made today to optimize production. In future "guided" analytics, the data analysis will be initiated by a domain expert, but will otherwise run automatically; in autonomous analytics, the entire data analysis process will be automated. 

To be continued… with follow-up articles on a weekly basis…