Keywords

1 Introduction

The increasing competitive pressure combined with global crises is a tremendous challenge for manufacturing companies. The transition from the internal combustion engine to electric mobility, combined with governments’ environmental protection targets, poses major hurdles for the automotive industry in particular [1]. It is therefore important to optimize existing processes in order to produce as efficiently and cost-effectively as possible while maintaining quality, supply capability and punctuality. Maintenance plays a major role in this context, as it accounts for an average of 9% of manufacturing costs [1, p. 1]. Approximately 50% of these costs can be attributed to the repair of equipment during shutdowns.

Predictive maintenance makes it possible to detect upcoming failures in time and thus perform maintenance proactively, rather than just reactively [2, 3]. Advances in connectivity between machines are making this digital transformation possible. Nevertheless, predictive strategies, are only used by about 15% of the manufacturing companies [4, p. 5]. In addition, suitable methods for data evaluation and forecasting are not yet sufficiently established [4, p. 2f.].

The transformation towards the digital factory developed the virtual commissioning (VC). A complex interconnected model of the planned system is generated in order to shorten the real commissioning time and thereby save costs. However, the resulting model is currently not further used after the real commissioning of the plant [5, p. 9]. Continued use of the model would relativize the investment in its creation.

This paper investigates if and how the model from the virtual commissioning can be used in later stages. Furthermore, it is described how the model has to be improved in order to build a hybrid predictive maintenance system. The combination of these two elements of Industry 4.0 and Digital Factory represents a significant scientific contribution to automation in production.

2 State of the Art

In the following, the fundamentals of the life cycle of production plants will be covered, as well as the digitalisation concepts relevant for this paper.

2.1 Life Cycle of Production Plants

The life cycle of production plants can be divided into three phases. The planning and realization phase, the operating phase, and the redistribution. In the following, the first two of these phases will be discussed in more detail since the development of the virtual commissioning model is situated between them.

Planning and Realization Phase

In the beginning, sequences and processes are designed and, if necessary, simulated according to the customer’s requirements. Subsequently, mechanical, electrical, and pneumatic systems are developed and designed. The plant is then manufactured based on the construction plans. After the individual components have been completed, the final step is assembly and commissioning at the customer’s site [6, p. 11f.]

The elimination of software errors takes up to 90% of the entire commissioning. The software is created after the design phase, i.e., in parallel to the manufacturing and assembly phase. Therefore, there is usually not enough time for extensive function tests of the control software before commissioning. To meet this challenge, virtual commissioning has been established in the digitalisation of planning [7, p. 19].

Operating Phase

After the takeover of the production system by the customer, the ramp-up phase follows. During this stabilization process, technical defects and failures occur that were not detected during commissioning. One of the main reasons for this is the control software, due to the short test cycles. After the ramp-up has been completed, series production begins. In this phase, the system must be maintained in case of failures [6, p. 4].

2.2 Digitalisation of the Planning, Realization and Operation Phases

Dynamic and growing customer requirements demand more product variants with shorter time-to-market and reduced product life cycles. OEMs have to demonstrate flexibility and transformability due to the transformation towards electromobility. In the following, digitalisation and Industry 4.0 instruments are presented to meet these requirements [8, p. 103].

Virtual Commissioning (VC)

In virtual commissioning, a large part of the commissioning activities is simulated in advance. Based on the construction plans, a virtual model of the plant is created, with which processes and material flows can be modelled and thereby the control software can be tested. On average, the software deployment time is reduced by 75%. The quality of the control software increases from 37% to 84% [9]. Furthermore, critical and potentially dangerous situations can be simulated safely and without damage. By simulating complex interlinked processes, not only individual processes, the entire system behaviour can be defined and optimized in an early stage [10, p. 273f.].

Maintenance Strategies

In general, there are five different maintenance strategies. Reactive maintenance corrects faults only after they have occurred. Preventive maintenance avoids recurring faults through periodic servicing. The next strategy is monitoring the condition of a wearing part and maintenance is carried out when certain limits are reached. Predictive maintenance analyses patterns and trends of past failures and predicts the time of failure, including a confidence interval, based on current process data. The prescriptive strategy goes beyond and defines concrete solutions for what needs to happen at the predicted time of failure in order to maintain the plant [2].

A recent literature review shows that there are two basic approaches for building predictive maintenance strategies [2]. First, data-driven approaches that extract patterns and trends from process data of a plant and generate a prediction of the failure using machine learning methods [11]. On the other hand, physically driven approaches offer the possibility to generate prediction models based on the underlying physics [12]. A hybrid concept uses a combination of the two approaches above [13, 14].

Herein lies the novelty of the following concept. Building on existing research, this paper investigates how the virtual commissioning model can be developed into a digital twin to create a hybrid predictive maintenance system.

3 Concept, Industrial Use Case and Challenges

In order to understand how the potential of digital twins can be maximized, this chapter presents a method to identify beneficial use cases and further develop the VC model into a digital twin. Finally, the concept is implemented in an industrial use case.

3.1 Use Cases for Digital Twins

The developed concept (Fig. 1) suggests a way to further expand the virtual commissioning model into a digital twin. This is used to build a hybrid predictive maintenance system. The advantage lies in the fact that the virtual model from the planning phase has a practical use in the operating phase of production plants.

According to [15] a virtual model can be developed into a digital twin by including an automatic data exchange between real and virtual machine. In case of a semi-automatic information exchange this is called a digital shadow.

Fig. 1.
figure 1

Architecture for developing a hybrid predictive maintenance system

Full size image

However, a predictive maintenance system should not be set up during the ramp-up phase, instead it should be built when the production plant is in stable series production. Further necessary information for setting up a hybrid predictive maintenance system is provided by experience of similar wear patterns, as well as literature. Predictive maintenance is particularly well suited for predicting wear patterns of mechanical components.

This paper presents the questions that can identify an appropriate use case for a given machine as shown in Fig. 2. By diligently answering those questions, a use case for predictive maintenance based on digital systems can be identified and specified.

After the basic requirements for creating a hybrid predictive maintenance system have been fulfilled, it is examined whether the benefit exceeds the cost. This turns a simple use case into a business case. Exemplary benefits of predictive maintenance are listed below [16, p. 5]:

  • Reduction of plant downtime and therefore increase of the OEE.

  • Optimal utilization of the operating lifetime of wearing parts.

  • Reduction of maintenance costs.

  • Assurance of the quality of products.

As soon as the economic viability has been proven, the mechanism underlying the wear is inspected. First, it is assessed if the existing sensor technology is capable of detecting the fault or if additional external sensors are required. Then it must be checked whether the data integration is sufficient to process the sensor information and to record process data. After a positive assessment, data collection can be continued.

For data acquisition, it is important that the data can be recorded not only from the real machine, but also from the extended virtual commissioning model via the PLC of the HiL setup - the Digital Twin. By combining both real process data and virtually generated failure data, a hybrid predictive maintenance system is realized. The advantage of using the digital twin is that more failure data can be generated. In practice, there often is not enough data available that contains the failure that is supposed to be predicted. This is where the synthetically generated failure data acts as an enabler by providing essential training data with as many failures as necessary.

Fig. 2.
figure 2

Identification of a use case for a hybrid predictive maintenance system

Full size image

3.2 Development of the Virtual Commissioning Model Towards a Digital Twin

After the use case has been identified, the task is to extend the virtual commissioning model to generate synthetic fault data. There are several approaches for this.

The first approach reproduces the wear in the form of mathematical equations. In addition to a deep understanding of the process, this also requires a high computing power, since the formulas are mostly multidimensional and are iteratively optimized.

The second approach is material removal simulation. Using the example of a cutting process, the cutting edge gradually reduces in sharpness with each cut until the cutting edges are so worn that the material can no longer be cut through. This is challenging, because such a detailed material simulation would require a tool that supports such a feature, thus limiting the choice. Secondly, and more importantly, the computational performance for simulating the degradation of the blades is substantial and as such would hinder a model or digital twin beyond feasibility and practicability.

The third and selected approach is synthetic failure data generation (Fig. 3) based on real patterns and trends. Real fault data are recorded and imported into the simulation program. These data are then randomized in the simulation within a statistically significant range. Hence, new synthetic fault data are generated, based on real patterns and trends. This could not be done if subprocesses were simulated separately. The connections between subsystems in the digital twin are necessary to understand the machine’s behaviour and to correctly predict failures.

In this paper, an architecture model is developed that provides the essential steps for building a hybrid predictive maintenance system (Fig. 1). On one side, real process data are collected. On the other side, virtually generated failure data are recorded to create a broader failure database and to avoid overfitting.

A crucial factor is the further development of the virtual commissioning model into a digital twin. First, all process parameters from the real machine are entered into the virtual model, like e.g. axis parameters or closing and opening times of cylinders. Furthermore, the current PLC program of the real plant is imported into the HiL configuration and then the virtual model is optimized until the automatic mode is in operation.

Ultimately, the developed failure data generation module is implemented in the corresponding process in order to simulate failure data based on real patterns and trends. In case of a constructional change of the real system, the model has to be modified accordingly.

After the digital twin behaves like the real machine in automatic mode, synthetic failure data can be generated. These are then incorporated into the prediction model from the real plant. This solves the problem that there often is not enough data from the time of failure to robustly train a prediction model.

Fig. 3.
figure 3

Structure of a hybrid prediction

Full size image

3.3 Example of an Industrial Use Case

Fig. 4.
figure 4

Illustration of a cutting process in the Hairpin production

Full size image

A new approach in the field of electric engines for automobiles is to use stiff, u-shaped hairpins in the stator instead of a copper winding made of round wire. In the production of the hairpins, there is a cutting process which cuts the raw material of the copper coil to the required length, which is then bent into the required shape.

The cutting tools of the process (Fig. 4) are worn down and lose their sharpness with every cut. Once the cutting edges are worn out, the machine stops. The blades have to be replaced. Changing them requires maintenance, which in turn requires production to be halted for some time. Since the wear depends especially on external influences, there is no exact regularity in the failures of the line. On the other hand, the costs for a preventive approach, e.g., replacing every week, are too high. Therefore, this example forms the ideal use case for a predictive maintenance application.

Prediction methods require a certain number of data sets to be able to predict with sufficient accuracy. The more irregular the patterns and trends that occur, the more data is needed to identify stable correlations. Conventional recording of process data therefore requires a significant time. Therefore, we developed a component in ISG-virtuos (Fig. 5) that incorporates the patterns and trends from real process data (LookUp Table) in the digital twin and then randomizes them (Randomization). The port in the top left corner (TurnOn_SyntheticData) allows the user to toggle this option.

By doing so, additional training data can be generated, in parallel to the real machine’s data plus any perturbations and failures the user chooses to include. Thus, the training data can be gathered at twice the speed or even faster, by using more setups simultaneously. Additionally, the generated data can yield more variation than the real data. Yet another advantage of the synthetic data generation with the digital twin, based on the further developed virtual commissioning model, is that the effect of aging on subsequent processes can be directly observed and investigated.

Fig. 5.
figure 5

Developed module for generating synthetic failure data with ISG-virtuos

Full size image

3.4 Challenges

To build a hybrid prediction model, a significantly more profound understanding of the process is required so that the virtual model can be further developed into a digital twin. On one side, it must be understood how the digital twin can be used to provide additional information about the occurring failures. Use and business cases must be identified and evaluated. On the other side, the virtual commissioning model has to be extended correspondingly to be able to generate synthetic failure data.

Data pre-processing is particularly challenging, since in practice a lot of data are recorded, but usually only a small part of it is relevant for predicting a failure. When using machine learning algorithms to predict failures, it is important to define an accurate benchmark to select the best model. It is also possible to average multiple predictions depending on how well the forecasts perform.

4 Conclusion and Outlook

With virtual commissioning models becoming increasingly available, this paper presents a method for evolving them towards a hybrid predictive maintenance system. Finally, the presented results are concluded and an outlook on future work is proposed.

4.1 Conclusion

The concept of a hybrid predictive maintenance system makes it possible to generate predictions more efficiently and accurately. The virtual commissioning model is further developed into a digital twin of the machine. With this, synthetic failure data are generated based on the patterns and trends of real process data. This helps training prediction algorithms faster and more accurately.

The core requirement for the creation of such a hybrid system is the existence of the virtual commissioning model. In addition, the plant has to be producing in a stable condition and has to offer the possibility to collect data. First experiences of experts regarding the failures of the plant as well as the underlying mechanisms are of significant advantage. If these conditions are fulfilled, a use case can be identified where the benefit can outweigh the effort.

From an economic point of view, this concept is advantageous in several respects. First, the cost-intensive virtual commissioning models can be further utilized. Second, the synthetically generated fault data can be used to implement a predictive maintenance system at an earlier stage compared to conventional procedures, improving plant availability and reducing maintenance costs.

It has to be noted that in order to follow the presented method and to actually set up such a maintenance system, expert knowledge is required. Without it, the use cases cannot be identified, and the hybrid system cannot be implemented to the necessary extent. It is of utmost importance to also identify business cases and ensure economic benefits. A digital twin always needs to serve a concrete purpose. Additionally, the approaches have to be adapted to existing structures and processes. If the maintenance is not equipped to train models and use their output, the benefits will diminish. In conclusion, in the right hands, this work presents a further step towards the digitalisation of production and maintenance.

4.2 Outlook

Digital twins in general and the hybrid system in particular hold a lot of potential, as this work shows. However, current virtual commissioning models usually do not have the necessary modelling depth to generate synthetic failure data. Therefore, in the further development to the digital twin, the sub-process to be predicted must generally be modelled in even more detail. The automation of this could be subject of future research. If use cases are recognized early enough, the models can be created with enough detail from the start, thus avoiding additional costs later on.

The digital twin is also well suited for cycle time simulations. In this case, aging in certain sub-processes can be simulated and the effect of this on following operations can be analysed. The next step could be the observation of the decrease of parts produced, due to wearing effects. Ultimately, further knowledge for maintenance and production can be generated.