Development of a method for the deployment of Deep Learning algorithms with applications in Autonomous Systems and Industry 4.0
The utilization of Machine Learning (ML) techniques in production processes is so far inherently case-specific due to the complexity of every individual setting and thus they cannot be transferred as a solution approach to related problems. This PhD is on devising a method and a tool that are scientifically generalized from specific settings with the aim to enable systematic transfer by fast situative specialization to other concrete problems that could not be solved so far. The applied scientific paradigms are Design Science Research (DSR) in conjunction with Method Engineering. The resulting artifacts are a method with a tool, including applicable ML models. The impact to business is that now ML techniques can be adapted more efficiently to new problems. The resulting method and tool will be also eligible to be marketed as a stand-alone product / service offering.
The dissertation will bring together different areas: decades of experience in manufacturing technology (domain knowledge) and decades of knowledge in data science, in most cases two worlds so distant from each other, shall come closer as a result of this PhD work. The dissertation shall also promote the integration of those two areas, enabling the effective cooperation of those fields, the formation of a way of thinking about data, products and production through the presented, developed examples. The initial setting of the PhD work will be MEMS-based (micro electromechanical systems) production plant data for inertial sensors (gyroscope + accelerometer). MEMS sensors are based on a micro-mechanical structure in some 10..100μm sizes, which are basically vibrating or oscillating structures produced by a series of micro-machining steps (lithographic processes, milling, surface separation, polishing). Even the best production process for MEMS-sensors on a global scale yield a considerable faulty output. Due to the deviations in manufacturing technologies, the numerous production steps and the long supply chains, several thousands of technological measurement or product-specific parameters are generated during the manufacturing of the product, which can nowadays be merged and efficiently processed in a database. With the help of ML techniques this PhD work should first identify the group of parameters that are correlated with faulty output. This production optimization problem is so far unsolved because the amount of data is too big to be assessed with conventional approaches, including human sensemaking.
This IDSL position will profit from the unique advantage of the joint availability a world class MEMS-production plant, a huge data lake of related production data, and university know-how in Machine Learning.
The resulting method and tool (including models) take data analysis, forecasting and understanding of production behavior in industry 4.0 to a new level. The models used in the integrated database or their simplified versions, embedded in the manufacturer's SW, make the production of the product more efficient and more accurate. The knowledge gained here can also be fed back into the development of the next product and other settings.
As mentioned before, with the evolution of the industry there is a natural market demand to produce more and more accurate products, like e.g. MEMS-sensors, that can be manufactured only with a relatively wide scatter of technological parameters. In the whole value stream and supply chain several thousand different time-sequential, static, temperature, pressure, etc. dependent dynamic parameters are measured and set. It has only recently been possible to link these parameters from the different plants.
By means of AI it will be possible to understand, predict, and implement countless feedbacks and feedforwards in the manufacturing process that improve the production in a given factory or even across factories. AI models can be used to optimize processes or to avoid suspicious effects in the production. Fast feedback can reduce intervention times and help reduce technological standard deviations. Through the feedforward of information of earlier process steps, production is able to improve the product and its manufacturing processes (data driven production). Thus, intelligent controlling of the data helps to reduce standard technology deviations and increase product accuracy and improve quality at the end.
The created/resulting AI models provide a better understanding of individual or integrated sensor functions. The implementation of the models in a manufacturing environment opens up a new perspective in fulfill product specifications.
Often it is not the product but the equipment used for production itself that causes the deviation of the product parameters e.g. temperature-dependent vibration. Such and similar effects are difficult to separate from the real sensor functions. In such cases, a large number of behavioral analyses and modeling using AI tools can ensure that the product is within the specifications. The challenge is to understand and model the behavior of both the product and production. For this purpose the artificial intelligence tools like Machine Learning provide the features for understanding the main correlations.
Trained models can be integrated efficiently in the production as an improvement, or further conditioning of these Machine Learning or Deep Learning models can be also provided to production environment. This solution can be very attractive for the ramp-up of a new product. Artificial intelligence tools provide an opportunity to reduce sequential technological steps of an algorithm by providing the prediction of the working point. The latter makes it possible to reduce the production time, thus reducing the production cost of the product.
Generally as in most manufacturing process, the three most crucial factors are cost, quality and delivery time which contribute to the growth and success of any company in this Industry.
MEMS manufacturing is a highly complex process consisting of several hundred steps. The real-time data captured during those process control steps results in a huge data base. Analysis of that enormous amount of data in real-time with high sample rate during production for eventual fault detection and prediction is very challenging. A typical analysis can involve several million parts with several 1000 different parameters. In some cases, correlations remain hidden because either models are not available or methods are missing to understand the behaviour. In the past the infrastructure was not able to handle, process and understand the behaviour of a problem with <0.1% failure rate and handle the necessary amount of data with such complexity.
The state-of-the-art so far is tuning of a production process by experts. Even a world-class plant like the one of Bosch in Hatvan, Hungary, was dimensioned by human expert knowledge and not by Machine Learning.
The solution consists of a method and a tool (including models) to solve complex production optimization problems using machine learning on data bases of production parameters.
Using these artifacts, we can move ahead in describing or predicting feedback, complex relationships and behaviors that have not been possible before. That shall result in static or customizable / trainable tools, models or methods to help production. They can be then used to separate effects that were not separable before, or to reduce production time. With the support of the available Big Data tools, it is possible to manage a complexity that without the use of AI tools had not been possible before and thus ensure that the increasingly stringent requirements for the product are met.
A better understanding of product behavior allows us to create models that can be used not only in manufacturing but also later in development. The knowledge, models or correlations acquired on a large number of data in production can also be utilized in the development of a new product family. In a further step, artificial intelligence applications can be embedded in a product, thus describing a specific behaviour of it. That shall add a new feature to the product providing greater customer satisfaction. Such can be, for example, the description and recognition of mechanical stress, which can reduce the accuracy of the product not only after production, but also after its incorporation into a device or application. In order to understand such effects, the analysis of the behaviour of a large number of products is required, taking into account the response of the sensors to standard technological deviations.
Classical production processes rely on the local decisions of the stations in the line, i.e. decisions are based on the evaluation of the data generated at a given station. Data-driven manufacturing, which is the essence of this innovation work, allows the use of any data in the production chain on any workstation. That provides the opportunity to extend the classic I4.0 implications through publications and conferences not only within Bosch but also beyond. I4.0 applications, coupled with artificial intelligence, raise manufacturing technology to the next level, generating large amounts of data from the outset during product manufacturing.
In our specific case, the PhD dissertation demonstrates the applicability of the idea trough the example of MEMS manufacturing technology, but at the same time it goes far beyond showing the path to any manufacturing technology or service where large amounts of data are generated during the manufacturing or process.
Analyses of large amounts of data during the life cycles of products or services, as well as the experience gained here, show what connections, needs and specifications need to be taken into account in the early stages of product or service development in order to be more successful in the market. In this way, the activity both reshapes the development processes of products or services and creates the need for new approaches in development processes by taking into account the complex relationships, models, principles provided by artificial intelligence and by incorporating them into the product development process.
The expected outcome is to formulate a generalized and transferable method using Machine Learning and Deep Learning for Process optimization contributing to smart and efficient production systems; which will result in improved products or production processes. The work will target both scientific excellence and practical relevance.
This will lead to publications and conferences in different research domains such as in the field of Machine Learning, Deep Learning, and Information Systems (for Method Engineering), etc. The research area surely has strong base for PhD thesis.
Concrete innovations expected as the outcome
The concrete artifacts will be a method and a tool (including models) to improve complex industrial production processes. Its application to the MEMS-production domain will yield in a more efficient and effective production process. Also, new product planning will be improved. The method and the tool themselves will be further productized leveraging the Bosch Innovation Management department. The innovative artifacts will be tested on MEMS sensor production plant settings of Bosch. The innovation will be tested in a prototype related to this Bosch plant and will result in improved production. It is foreseen that the resulting process (as part of the method) will be patented and leveraged for use outside this setting, e.g. as an own service offering that could also result in a spin-off, if the business case is viable.
Bosch has more than 10 years of experience in MEMS sensor development and manufacturing at the company's sites in Germany and Hungary. The need for continuous improvement as well as advances in technology have made it possible for the huge database generated in the field of product development and manufacturing to become available for further analysis in a new common integrated database. Combining Big Data technology with sensor knowledge and experience allows us to take product development and production to a new level, ensuring that Bosch sensors remain competitive on the market.
The innovations are manifested in the reduction of technological deviations, the reduction of production time and so at the end the reduction of the product price. The developed models make both the product and the production of the product more accurate, and it is possible to understand new product behaviors. The description of this behavior can be fed back into either the production of running products or their development, which will increase the performance of the next generation of MEMS-based sensors. By increasing the accuracy, it is also possible to ensure the spread of the next generation MEMS-based sensors into the autonomous driving function.
Thus, in addition to classical behavior prediction, there is a need for methods that not only predict a given behavior, but also contribute to its understanding of its physical description where possible.
The machine learning and deep learning methods or algorithms will be used on the sensor manufacturing dataset which is extracted during the manufacturing process control.
The testing phase will be conducted at the company premises as the data is confidential. In the testing phase it will be verified whether the machine learning and/or deep learning algorithms is/are capable of achieving the target. Here the target can be described as different pain areas in the sensor manufacturing process. If the target is achieved, the final solution will be tested generally on other set of problems.
The resulting method artefact will be transferred to an innovation with commercial impact by Bosch’s Innovation Management department with established and proven processes of further commercialization.
The PhD work will for the first time analyse the huge amount of production data that is available from the Bosch production process and find so far unknown ways of process improvement. The work will include generalizing the insights to a method and a tool that are eligible to transfer and apply them to new use cases. As a result, companies using the method and tool will have a competitive advantage over other companies not using this approach. In addition, a business can be created by offering the method and tool as a service to paying customers (other producing companies).
For example, in the concrete use case of MEMS-sensor production: The PhD activity will ultimately facilitate the further development of organizations working in the field of MEMS, enabling the wider dissemination and application of artificial intelligence technology within the development and manufacturing organization. The application of Big Data technology opens up new perspectives in terms of development-side production support. In this way, the organization is able to react to challenges and changes with a faster reaction time, which ultimately manifests itself in better quality and customer satisfaction.
From now on, as mentioned above, by increasing the accuracy of MEMS-based products, the use of exclusively MEMS-based sensors in new areas such as autonomous driving functions can be realized more widely. This will result in a more affordable commodity end product, facilitating the wider application of the technology in the vehicles of the future. At the same time, in addition to product development, there is an opportunity for artificial intelligence devices in the field to be integrated into a product or its manufacturing / development process as part of a product or service.
In addition, the application of AI technology to products already in production can provide a competitive advantage over products from other manufacturers on the market by improving the price of the product through production time and accuracy.
Time-line and milestones
Expanding knowledge about Method Engineering, Design Research and Sensor production processes.
identifying different use cases:
Researching in depth about the existing approaches and evaluating their implementation within the aforementioned use cases. Formulating new or improved methods using the correlation between the use cases and existing approaches as a guideline:
Applying the formulated method/s to other use cases for validation, examining the scalability and combinability of solutions for other areas of MEMS product development and manufacturing:
Make a formalization of the derived method/s based on the obtained results. Draw a conclusion. document and finalise the PhD thesis
If you are interested in applying to this position, please complete the EIT Digital Expression of Interest form here.
Deadline: 2 December 2020