AIoT in High-Volume Manufacturing Network

Introduction

Bosch Chassis Systems Control (CC) is a division of Bosch, which develops and manufactures components, systems and functions in the field of vehicle safety, vehicle dynamics and driver assistance. The products from Bosch CC are combining cameras, radar and ultrasonic sensors, electric power steering and active or passive safety systems to improve driver safety and comfort. Bosch CC is a global organization, with 20 factories around the world. Very high volumes, combined with a high product variety, characterize the Bosch CC production. Large numbers of specially designed and commissioned machines are deployed to ensure high levels of automation. Organized in a global production network, the plants can realize synergies at scale.

TBD: IMAGE

Naturally, IT plays an important role in product engineering, process development, and manufacturing. Bosch CC has more than 90 databases, 450 servers, and 9,000 machines connected to them. 3,900 users are accessing the central MES system (Manufacturing Execution System) to track and document the flow of materials and products through the production process.

Phase I: Data-centric Continuous Improvement

Like most modern manufacturing organizations, Bosch CC strives to continually improve product output and reduce costs, while maintaining the highest levels of quality. One of the key challenges of doing this in a global product network is to standardize and harmonize. The starting point for this is actually the machinery and equipment in the factories – and especially the data that can be accessed from it. Over the years, Bosch CC has significantly invested in such harmonization efforts. This was the prerequisite for the first data-centric optimization program, which focused on EAI (Enterprise Application Integration, with a focus on data integration), as well as BI (Business Intelligence, with a focus on data visualization). Being able to make data in an easily accessible and harmonized way accessible to the production staff resulted in a 13% output increase per year in the last 5 years. The data-centric continuous improvement program was only possible because of the efforts in standardizing processes, machines and equipment, and making the data for the 9,000 connected machines easily accessible to staff on the factory floor. The data-centric improvement initiative is mainly focusing on two areas:

• Descriptive Analytics (Visual Analytics): What happened?
• Diagnostic Analytics (Data Mining): Why did it happen?

The program is still ongoing, with increasingly advanced diagnostics analytics.

Phase II: AI-centric Continuous Improvement

Building on the data-centric improvement program is the next initiative, the AI-centric Continuous Improvement Program. While the first wave was predominantly focusing on gaining raw information from the integrated data layer, the second wave is applying AI and Machine Learning with a focus on:

• Predictive Analytics (ML): What will happen?
• Prescriptive Analytics (ML): What to do about it?

As of the time of writing, this new initiative has been going on for 2 years, with first results coming out of the initial use cases, supporting the assumption that this initiative will add at least a further 10% in annual production output increase.

TBD: IMAGE

Closed-Loop Optimization

The approach taken by the Bosch CC team fully supports the Bosch AIoT Cycle, which assumes that AI and IoT support the entire product life-cycle, from product design over production setup to manufacturing. The ability to gain insights into how products are performing in the field provides an invaluable advantage for the product design and engineering. Closing the loop with machine building and development departments via AI-gained insights enables the creation of new, more efficient machines, as well as product designs better suited to efficient production. Finally, applying AI to machine data gained via IoT enables root cause analysis for production inefficiencies, optimization of process conditions, as well as bad part detection and machine maintenance requirement predictions.

TBD: IMAGE

Program Setup

A key challenge of the AI-driven Continuous Improvement program is that the optimization potential cannot be found in a single place, but is rather hidden in many different places along the engineering and manufacturing value chain. In order to achieve the goal of 10% output increase, every year hundreds of AIoT use cases have to be identified and implemented. This means that this process has to be highly industrialized.

TBD: IMAGE

In order to set up such an industrialized approach, the program leadership identified a number of success factors, including:

• Establishment of a global team to coordinate the efforts across all factories in the different regions and provide centralized infrastructure and services
• Close collaboration with the experts in the different regions, working together with experts from the global team in so-called tandem teams

The global team started by defining a vision and execution plan for the central AIoT platform, which combines an AI pipeline with central cloud compute resources as well as edge compute capabilities close to the lines on the factory floors.

Next, the team started to work with the regional experts to identify the most relevant use cases. Together, the global team and the regional experts prioritize these use cases. The central platform is then gradually advanced in order to support the selected use cases. This ensures that the platform features always support the needs of the use case implementation teams. The tandem teams consist of central platform experts as well as regional process experts. Depending on the type of use case, they include Data Analysts, and potentially Data Scientists for development of more complex models. Data Engineers support the integration of the required systems, as well as potentially required customization of the AI pipeline. The teams strive to ensure that the regionally developed use cases are integrated back into the global use case portfolio, so that the can be made accessible to all other factories in the Bosch CC global manufacturing network.

AIoT Platform and AI Pipeline

The AIoT platform being built by Bosch CC is combining traditional data analytics capabilities with advanced AI/ML capabilities. The data ingest layer integrates data from all relevant data sources, including MES and ERP. Both, batch and real-time ingest are supported. Different storage services are available, to support the different input types. The data analytics layer is running on Microsoft Azure, utilizing Tableau and Power BI for visual analytics. For advanced analytics, a machine learning framework is provided, which can utilize dedicated ML compute infrastructure (GPUs and CPUs), depending on the task at hand. The trained models are stored in a central model repository. From there, they can be deployed to the different edge notes in the factories. Local model monitoring help to gain insights into model performance and support alerting. The AI pipeline supports an efficient CI/CD process and allows for automated model re-training and re-deployment.

TBD: IMAGE

Expert Perspective (Interview)

In the following interview, Sebastian Klüpfel from the Bosch CC central AI platform team shares some insights and lessons learned.

Dirk Slama: Sebastian, how did you get started with your AI initiative?

Sebastian Klüpfel: Since 2000, we have been working on the standardization and connection of our manufacturing lines worldwide. Our manufacturing stations provide a continuous flow of data. For each station, we can access comprehensive information about machine conditions, quality-relevant data and even individual sensor values. By linking the upstream and downstream production stages on a data side, we achieve a perfect vertical connection. For example, for traceability reasons we store 2,500 data points per ABS/ESP part (ABS: Anti Locking Brake, ESP: Electronic Stability Program). We use this data as the basis for our continuous improvement process. In the ABS/ESP manufacturing network, thus, we were able to increase the production rate by 13% annually over the last 18 years. And all this just by working on the lower part of the I4.0 pyramid, the data access / information layer.

Dirk Slama: Could you please describe the ABS/ESB use case in more detail?

Sebastian Klüpfel: This was our first use case. Our first use case for AI in manufacturing is at the end of the ABS / ESP final assembly, where all finished parts are checked for functionality. Up to now, during this test process, each part has been handled and tested separately. Thus, the test program did not recognize whether a cause for a bad test result was a real part defect or just a variation in the testing process. As a result, we started several repeated tests on bad tested parts to make sure that the part was faulty. A defective was thus tested up to 4 times until the final result was determined. This reduced the output of the line because the test bench was operating on its full capacity by repeated tests. By applying AI, we reduce the bottle neck at the test bench significantly. Based on the first bad test result, the AI can detect whether a repeated test is useful or not. In case of variations in the testing process, we start a second test and thus we save the part as a “good part”. On the other hand, defect parts are detected and are immediately rejected. Thus, unnecessary repeated checks are avoided. The decision whether repeated test makes sense or not, is made by a neuronal net, which is trained on a large database and is operated closely to the line. The decision is processed directly on the machine. This was the first implementation of a closed loop with AI in our production. Our next AI use case was also in the ABS/ESP final assembly. Here we have recognized a relation between bad testes parts at the test bench and the caulking process at the beginning of the line. With AI, we can detect and discharge these parts at the beginning of the line, before we install high-quality components such as the engine and the control unit.

Dirk Slama: And what about the change impact on your organization and systems?

Sebastian Klüpfel: With the support from our top management, new roles and cooperation models were established. The following roles were introduced and staffed to implement AI use cases:

• Leadership: management and domain leaders need to understand strategic relevance, advantages, limits, and an overview of tools.
• Citizen Data Scientist: work on an increasingly data-driven field and uses analytical tools applied to its domain. Therefore, a basic understanding and knowledge on Big Data and machine learning is necessary.
• Data Engineer: builds Big Data systems and knows how to connect to this systems and machine learning tools. Therefore, a deep knowledge in IT Systems and development is necessary.
• Data Scientist: develops new algorithm and methods with deep knowledge in existing methods. Therefore, the Data Scientist must be up-to-date, has know-how in CRISP-DM analytics project lead.

We want to use the acquired data in combination with AI with the maximum benefit for the profitability of our plant (data driven organization). Only through the interaction and change in the working methods of people, machines and processes / organization, we can create fundamentally new possibilities in order to initiate improvement processes and achieve productivity increases. We rely on cross-functional teams from different domains to guarantee a quick success.

Dirk Slama: What were the major challenges you were facing?

Sebastian Klüpfel: For the implementation of these and future AI use cases, we rely on a uniform architecture. Without this architecture, the standardized and industrialized implementation of AI use cases is not possible. The basis is the detailed data for every manufacturing process, which are acquired from our standardized and connected manufacturing stations. Since the year 2000 we have implemented a MES (Manufacturing Execution System) that enables the holistic data acquisition. The data is stored in our cloud (Data Lake). As a link between our machines and the cloud, we use proven web standards. After an intensive review of existing cloud solutions internally and externally, we decided to use the external Azure Cloud from Microsoft.Here we can use as many resources as we need for data storage, training of AI models and pre-processing of data (Data Mart). We also scale financially, and we only create costs where we have a benefit. Thus, we can also offer the possibility to analyze the prepared data of our Data Mart via individually created evaluations and diagrams (Tableau, PowerBI). We run our trained models in a edge application close to our production lines. By using this edge application, we bring the decisions of the AI back to the line. For the connection of the AI to the line, only minimal adjustments to the line are necessary and we guarantee a fast transfer of new use cases to other areas.

Dirk Slama: How does your ROI look like for the first use cases?

Sebastian Klüpfel: Since an AI decides on repeated tests on the testing stations of the ABS / ESP final assembly, we can detect 40% of the bad parts already after the first test cycle. Before introducing the AI solution, the bad parts went always through 4 test cycles. Since the test cells are the bottleneck stations on the line, the saved test cycles can increase the output of the line, reduce cycle time, increase quality and reduce error costs. This was proven on a pilot line. The rollout of the AI solution offers the potential for an increase in output of around 70 000€ per year and an invest reduction of nearly 1 million (since no additional test stations are needed to be purchased for more complex testing). This was only the beginning. With our standardized architecture, we have the foundation for a quick and easy implementation of further AI use cases. By the implementation of AI into the manufacturing, we expect an increase in productivity of 10% in the next 5 years.

Dirk Slama: What are the next steps for you?

Sebastian Klüpfel: Our vision is that in the future we will fully understand all cause-effect relations between our product, machine and processes and create a new way of learning with the help of artificial intelligence to assist our people to increase the productivity of our lines. In order to achieve this, the following steps are planned and already in progress:

• Pioneering edge computing: First, we are working on a faster edge-application. We have to bring the decisions from the AI even faster to the manufacturing station. For the first use cases, our edge-application is still sufficient. However, for AI use cases for short-cycle assembly lines (approximately 1 second cycle time) the actual edge solution is no longer sufficient. Here we are already working on solutions to deliver the predictions back to the line within fractions of a second even for such use cases.
• Automated machine learning: 80% of our data is already pre-processed automatically. Our target is to further increase the automation rate. In addition, we have ideas how to automate the selection of the right ML model with an appropriate hyper parameter search. Of course we are also working on an automated analysis of the ML decisions to monitor the health status of our already productively working models
• Implement more use cases: Our architecture is designed for thousands of AI use cases. We have to identify and implement these. By doing so, we make sure that we do not implement show cases. We want to implement real use cases for our Digital Factory, including:
  • Predict process parameters: learn optimal process parameters from prior processes (e.g. prior to final assembly)
  • Adaptive tolerances
  • Bayesian network: We want to train a Bayesian network on all parameters of HU9 final assembly. This means that influences and relations can be read from the graph. Relations are much deeper than pairwise correlations.

Dirk Slama: What are the key lessons learned thus far?

Sebastian Klüpfel: For a continuous improvement program like ours, KI must be industrialized: we have been proven right with our approach to make AI applicable on a large scale instead of individual “lighthouse projects” which are not easily adaptable to other use cases. Another key success factors is the standardized architecture: by storing all the data in a cloud, we can ensure a holistic view on the data across all factories in our global manufacturing network, and use this to train the models centrally with all the required data at hand. Next, we can then take the trained models and deploy them on the edge layer, as close to the actual production lines as possible. In order to work quickly and efficiently on the implementation of AI use cases, new competence and work models have to be established. The most important thing is that the digital transformation must be a part of the corporate strategy. After all, digital transformation can only succeed if all employees follow together the path toward a data driven organization.

Dirk Slama: Thank you!