Adaptable and Component-Protecting Disassembly in the Regeneration Path

Abstract

The disassembly initiates a product’s maintenance and regeneration. In order not to cause any additional damage to the components during disassembly, which would lead to higher repair costs or, in the worst case, to destruction and loss of the components, disassembly must be as gentle as possible on the components. Due to ambiguous causes, such as thermal or mechanical product loads during operation, the unknown product condition is a characteristic uncertainty factor in disassembly. This paper presents approaches and methods on how the disassembly of complex capital goods, which is usually carried out manually, can be automated while still being protective on components and adaptable to varying product conditions. Manual disassembly procedures are substituted using micro impacts induced by a piezo actuator. A learning model predicts optimized process parameters based on varying operational usage scenarios.

1 Introduction

Disassembly is a crucial step toward sustainable life cycle engineering, restoring a product’s functionality in subsequent regeneration processes or recovering product components (Seliger 2007). At regular maintenance, repair and overhaul (MRO), components are regenerated, their properties restored and renewed or, if necessary, replaced with new parts. After an initial assessment, the components are disassembled gently, to ensure proper preservation. Contrary to assembly, the product’s condition at the end of its life cycle changes to an unknown extent, making it difficult to plan an automated disassembly. Considering aircraft engines, due to environmental conditions, e.g., direct contact to hot exhaust gas mixture and high tensile loads during operation, its components like high-pressure turbine (HPT) blades are exposed to three major influences: fatigue, corrosion and creep (Bräunling 2015). As a result, the HPT blades inside the turbine disks change their condition. That change is characterized in particular by a solidification¹ of the joint, as it essentially hinders and impedes the separation of the joining partner.

Due to the unknown state of the assembly connections, process planning parameters, like tool dimension, needed forces or disassembly time only become apparent during the ongoing blade dismantling process. Depending on operating parameters such as operating hours, landing and take-off (LTO) cycles, or environmental influences, like flight routes over the sea or desert, the component’s joints solidify in an undefined manner and might be further damaged without a suitable disassembly strategy. That would destroy the components before the beginning of any maintenance and regeneration processes.

Especially the turbine blade’s special manufacturing for highest stresses makes this component very costly with approximately $ 10,000 each (IAE V2500 engine, according to the turbine’s manufacturer).

Besides the uncertain product condition, the shape of the joint differs. Depending on the engine and the engine stage, numerous shapes and dimensions exist. The HPT blade roots, for example, are assembled in fir tree slots into the turbine disks, as shown in Fig. 1 (Benad 2019). The disassembly tasks of any engine thus differ in their unknown condition due to their differing use and operation scenario and their geometric properties. Although the blades rest loosely in the fir tree slot during assembly, the solidification of the joint prevents them from being easily pushed or pulled out during disassembly. Therefore, the disassembly is mainly carried out by manual labor (Schmücker et al. 2021). HPT blades, for instance, are hammered out with hammer strokes. By being carried out manually, the disassembly is adaptable to changing product conditions while being protective to highly valuable parts. Nevertheless, damage to components can still occur during manual disassembly.

Fig. 1

a Sketch of blades inside a turbine disk, b Image of a turbine blade with fir tree profile

In this article, we present our research on developing an automated disassembly by overcoming the uncertain product condition as its obstacle. In the section below, we outline the challenges and the approach to implement an adaptable and component-protecting disassembly.

2 Objective

The CRC 871 focuses on the optimization of maintenance tasks of complex capital goods. By developing a novel process chain characterized by automated processes and integration of a virtual layer into the real layer, an efficient and resource-saving process is achieved (Aschenbruck et al. 2014). Our objective in the sub-project A5 of the CRC 871 is to establish a technologically plannable disassembly despite characteristic uncertainties. Our approach is to substitute the manual hammer blows with reproducible micro impacts, induced by a piezo stack actuator. The reproducible adaptation of the disassembly process parameters to the operationally varying condition of the joints while adhering to maximum force limits of the blade material allows component-protecting disassembly and its automation. In order to set up the investigation, we developed an experimental environment and a simplified model of a solidified joint. In Sect. 3, the setup of the investigation, including the development of an experimental environment and a simplified model of a solidified joint is illustrated.

For our research on a component-protecting disassembly, we could not use real solidified joints between HPT blades and turbine disks of a used aircraft turbine. Therefore, our first sub-objective is to investigate methods to provide valid and reproducible samples. Thus, a substitute model of the operationally changed and solidified joints provides a basis for further investigation on the disassembly process. Consequently, the second sub-objective focuses on the development of a component-protecting disassembly. For this purpose, we are investigating how the disassembly process causes damage to the components, and how damage to components can be prevented by choosing optimized process parameters. Section 4 outlines the results of the development of an adaptable and component-protecting disassembly.

According to feedback from manufacturers and MRO service providers, no documentation is made of the extent to which the various influences affect the resulting solidification of the joint and disassembly effort. Subsequently, the third sub-objective comprises the summary of the results obtained. In order to adjust the disassembly process parameters adaptively to the joint’s condition, we designed and developed a learning model. Based on the varying engine’s operational usage scenario, it estimates, for example, required tool dimensions, the disassembly force and process times. With the knowledge of material-specific force limits, components like the HPT blades can be disassembled without damage by adhering to these limits. Hence, by using the learning model we obtain the correlation between the operational data, the resulting solidification and necessary disassembly forces and process parameters. Section 5 presents the development of a learning model to predict disassembly process parameters based on the condition of the solidified joint.

With regard to the automated regeneration process chain, capacity-critical requirements and subsequent processes, such as handling, are also considered. An exemplary partially automated disassembly workstation in a system demonstrator was developed and set up (Sect. 6).

3 Setup

Disassembly is a key step in the remanufacturing process. Regeneration for reuse in regular maintenance intervals or the recovery of valuable resources from end-of-life products becomes possible through disassembly. Airbus’ Process for Advanced Management of End of Life of Aircraft (PAMELA) program, for example, showed that up to 85% of aircraft components could be recycled or reused (Airbus 2008). The complex characteristics of disassembly, like the uncertainties and the lack of knowledge about the product’s condition, make disassembly planning more challenging than assembly planning (Bentaha et al. 2014) and have been investigated and improved over the last decades. Therefore, disassembly tasks in aircraft MRO are usually executed manually (Schmücker et al. 2021). However, productivity can be improved by overcoming the necessity of manual disassembly through automation.

3.1 Experimental Environment

As introduced, several effects lead into a varying solidification of the joints, resulting in laborious and challenging disassembly. While highly trained and experienced workers perform disassembly, little is documented about the condition of the joints. Intending to carry out disassembly in a reproducible and automated manner, we have developed an experimental environment, as shown in Fig. 2 (Bluemel and Raatz 2021).

Fig. 2

Experimental environment to perform disassembly tests (Bluemel and Raatz 2021)

It consists of a motor and a ball screw that converts the rotation to a linear motion. The pushing rod exerts the disassembly feed movement on the blade root. The resulting force to disassemble the blade can be measured and monitored with the integrated load cell. Centerpiece of the mechanism is a piezo stack actuator, which can induce and superimpose vibration on the disassembly movement. Researchers showed in simulation and experiments that superimposed ultrasonic vibrations can significantly reduce the coefficient of friction (Littmann et al. 2001; Popov et al.2009). Saffar and Abdullah (2021), for example, present the application of ultrasonic vibration to reduce process forces during drilling and milling. Experiments in which a bladed turbine disk was placed in an ultrasonic bath first showed a positive result. The blades slid out of the joint after a short time due to their own weight, indicating that the solidification of the joint was significantly reduced. However, further investigations showed a substantial change of the microstructure in the material’s surface zone which was severe enough to cause irreversible damage to the components (Lufthansa Technik 2014).

Consequently, vibrations are suitable for lowering the coefficient of friction and thus also reducing necessary disassembly forces. For a component-protecting disassembly process, however, the adjustment parameters must be carefully investigated and adjusted, in order to adhere load limits applicable on the blade root. According to our objectives, we intend to use the piezo stack actuator as a tool to support disassembly. Thereby, we use defined impacts instead of manual hammer strokes by superimposing vibration on the disassembly movement. As we have shown in previous work, it allows the required disassembly forces to be reduced (Wolff et al. 2016). The main characteristics of the piezo stack actuator are the oscillation amplitude, which is varied by adjusting the applied voltage, the period duration from which the frequency can be determined, and the waveform, such as a sinus or triangular wave. Secondary parameters such as the piezo actuator’s resonance frequency and the operating temperature determine the maximum set table frequency due to the high heat dissipation. According to the manufacturer, the piezo actuator has a total length of 139 mm and a diameter of 25 mm. The travel range is 120 μm, with a resolution of 3.6 nm. The actuator’s resonance frequency is specified at 4.5 kHz. The maximum push force capacity in motion direction is specified at 4,500 N, which represents the maximum force the disassembly experimental environment can apply. Therefore, the experiments in our study are limited to a maximum force of 4,500 N. In preparation for the experimental investigations carried out in the test environment, we created an analytical model of the solidified joint between the blade root and the turbine disk.

3.2 Simplified Model of a Solidified Joint

Within the research of the CRC 871, our focus is on the disassembly of the HPT blades. The blade root is the part connecting the blade to the turbine disk. Due to its appearance, the profile of the blade foot is referred to as the fir tree profile. Among other blade root profiles, the HPT blade’s fir tree profile has the highest bearing capacity appearing during the engine’s operation (Li et al. 2019). Varying operation scenarios, as described at the beginning, solidifies the connection, which makes disassembly more complicated.

In order to describe the solidified joint connecting blade root and turbine disk, we use an analytical model in analogy to a friction model (Fig. 3) (Wolff et al. 2015).

Fig. 3

Simplified model of the solidified joint between blade root and turbine disk

We identify the solidification of the joint, caused by wear and tear due to the engine’s operation, as a layer formed between the contact surfaces. The solidification of the joint is modelled as a resulting auxiliary material between the blade root and the turbine disk with the resulting surface pressure ps(z) as the central operation-dependent solidifying characteristic and a material and surface dependent coefficient of friction μs. It results in a solidifying force Fs(z) opposing the disassembly movement. Therefore, the disassembly force FD (z) must be higher than the resulting Fs(z), so that the blade root can be pressed out. However, the disassembly force must not exceed a limit, in order to prevent damaging the blade. The contact surface AC(z) is known from CAD data and decreases during the disassembly process along the disassembly path z. It becomes equal to zero when the actual path reaches the total disassembly length lD(z) (Bluemel and Raatz 2021). The solidifying force can be described as follows:

$$F_{s} \left( z \right)\, = \,\mu_{s} \cdot p_{s} \left( z \right) \cdot A_{C} \left( z \right)$$

(1)

In our investigation, the causes and the manner of solidification are not to be characterized or modelled. For the development of the model and our investigation, the degree of solidification, the resulting disassembly force to be applied and disassembly force curves are decisive. In later phases of the investigation, a functional relationship between the degree of solidification, the operational usage scenarios, and the geometry will be developed. However, due to the lack of real engines with individual operational usage scenarios, we used a replacement model for a solidified joint between blade root and turbine disk. In the next section, we present the development of this replacement model.

3.3 Replication of Solidified Joints

In order to investigate the solidified blade-disk joint, we created the simplified solidification model (Sect. 3.2). That model can be used to extract properties that affect the solidification of the assembly joint. Different usage scenarios can be imitated by varying these properties, constituting the resulting solidifying force. Therefore, for the developed replacement model, it is necessary to allow reproducible adjustability of these properties. As in Eq. 1, the solidifying force Fs is influenced by the coefficient of friction μs, the resulting contact pressure as the solidifying characteristic ps and the contact surface AC. For the replacement model, we do not manipulate the contact surface, so AC is omitted, leaving μs and ps as adjusting parameters.

We had samples manufactured with re-designed blade root shapes based on the original HPT blade root to replicate solidified joints, as seen in Fig. 4. The samples consist of an inner part representing the blade root and the outer part representing the turbine disk with the blade’s negative contour. The samples were manufactured by wire cutting to achieve high surface accuracy and fitting precision. In contrast to the clearance fit of the real joint, we use a slightly tighter fit for uniform contact.

Fig. 4

Samples with the re-designed shape of blade roots: a fir tree “R”, b fir tree “K”, c dovetail “D”

In order to allow replicated variation of the parameters influencing the solidification, we considered several methods: We investigated adding auxiliary materials to the gap, artificial ageing, according to a technical standard for accelerated ageing, and exerting an external force, as presented below.

The first method we were investigating was the use of auxiliary materials, i.e., adding glue to the samples gap. We applied glue (Loctite 222) on the contact surface to prepare the samples. Using a clamping device, we applied a force for a defined period of time at a constant ambient temperature of 25 ℃, under which the adhesive hardened. With that, we achieved a uniform layer thickness of the adhesive. To carry out the experiments, the inner part representing the turbine blade was pushed out while measuring the required disassembly force. By varying the curing time and clamping force, we intended to substitute the deviating solidification of the joint. For the investigations, we considered two different shapes of the blade root, fir tree “K” and fir tree “R” (Fig. 4).

Figure 5 shows an extract of the results of the subsequent regression analysis to investigate the influence of the input factors on the disassembly force and, eventually, verify the method. The model obtained by the regression is significant, but with a coefficient of determination, representing the goodness of fit, of R² equal to 0.738, it has comparatively little meaningfulness. With the calculated standardized regression coefficients, we can compare the input variables despite different units of measurement (Siegel and Wagner 2022). As a result, a variation in the shape of the blade root and the curing time have a negligible impact. Since a variation in the shape also means a change in the contact surface, a dependency should also be recognizable in accordance with our friction model.

Fig. 5

Pareto chart of parameter’s standardized regression coefficients for auxiliary material

This method, therefore, does not allow reproducible replication of the joint’s solidification. We assume that this is caused by the manual application of the adhesive, the varying surface textures, and environmental conditions.

Along with using auxiliary material to replicate the solidified blade root—turbine disk joint, we investigate the method of artificially ageing. We used a salt mist environment to artificially age the manufactured samples according to a standardized procedure. To ensure rapid corrosion, the samples were made of mild steel. We set the salt mist chamber to 37 ℃ with a 5% saline solution. We then put samples into the salt mist chamber for 72 h and 168 h, respectively. Similar to the previous experiments, the samples’ inner blade root model was pushed out. In addition to the shapes fir tree “K” and “R”, we considered the third shape, dovetail “D” of the blade root, for the experiments using artificial ageing (Fig. 4).

Due to the formation of corrosion, an increase in disassembly forces was observed. The measured disassembly force for the samples at 72 h was between 2,500 and 3,500 N. At 168 h, the disassembly force considerably increased over 5,000 N and thus beyond the load limit of the actuator. Figure 6 shows an extract of the results of the subsequent conducted regression analysis. With a coefficient of determination R² equals 0.778, the model shows a dissatisfactory fit. However, the standardized regression coefficients for different shapes of the blade root show a greater impact than using adhesives to substitute the joint’s solidification.

Fig. 6

Pareto chart of parameter’s standardized regression coefficients for artificially ageing

As a result, the artificial ageing showed restricted applicability to substitute the operational solidified joint. An increase in the required disassembly force was measured, and the force reduction by the piezo stack actuator was observed. However, the limiting factor of the artificial ageing method was the one-time examination of each sample. When disassembled, the sample became scrap and could not be re-prepared. Because of the lack of reproducibility and the costly preparation of the samples, we did not consider this method for further investigation.

Using auxiliary material and the artificial ageing showed unsatisfactory results, lack of reproducibility and reusability. Following the friction model, these methods did show an increase in the disassembly force by changing the coefficient of friction and the contact pressure. However, it is barely possible to adjust a determinable change. Therefore, we developed a clamping unit that exerts an external clamping force FCl on the blade root segment, inducing a surface pressure in the gap, as seen in Fig. 7. Instead of synthetically creating the solidified joint by manually adding glue or corrosion, we can repeat the experiments infinite times and can set a certain contact pressure by adjusting the clamping force, which represents the solidification of the connection (Bluemel and Raatz 2021). The clamping unit holds the samples fed into the disassembly experimental environment. It consists of a servo motor with an integrated gearbox connected to a machine vice. Both are placed on a linear guidance to position the sample holder under the disassembly pushing rod manually (see Fig. 2). The sample mount is placed between the vice’s jaws and includes a load cell to measure the normal force exerted on the samples. The samples are placed and clamped with repeatable accuracy using a locking pin and an adjusting screw. By using a load cell to measure the normal force exerted on the samples, we can describe and reproducibly adjust the substituted solidification of the joint, allowing us to reproduce the solidified blade root—turbine disk connection.

Fig. 7

Adapted model of a solidified joint using a clamping force

We presented in previous work that by varying and adjusting the clamping force, we can mathematically and statistically describe the induced contact pressure representing the solidification of the joint (Bluemel and Raatz 2021). The regression analysis revealed a good model fit with a coefficient of determination R² equals 0.981. To obtain results in the later process of our investigations that are more closely related to the IAE V2500 aircraft engine’s HPT blade root, used as an application case, we have designed new blade root geometries. Similarly, as in the previous experiments, we designed slightly modified contours to vary the blade root’s shape and the contact surface, as seen in Fig. 8 According to the design, we named the original shape “Original”, the sample with twice the size “Double”, and a contour having the same contact surface but a different contour as the original blade root contour “B-shape” (Bluemel and Raatz 2021). The new samples were made of stainless steel to prevent the occurrence of incidental corrosion. Also, the new samples have a cutting and thus consist of three parts. Similar to Fig. 4, the inner part represents the blade root. On the contrary, the outer part, representing the turbine disk, is separated by the cutting. Below, we present the exemplary results using the original HPT blade root’s shape.

Fig. 8

Overview of the different blade root samples similar to Fig. 4 with an additional cutting (Bluemel and Raatz 2021)

The subsequent regression analysis showed the influence of the input parameters and their interactions on the disassembly force. Figure 9 shows an extract of the results. The disassembly force shows the greatest influence. The “B-shape” contour has a negligible impact (0.031) since it has the same surface as the “Original” contour. Whereas the “Double” contour has a slightly larger influence (0.095). In comparison, the clamping force has the major (1.130) influence, as expected.

Fig. 9

Pareto chart of parameter’s standardized regression coefficients for clamping the samples

Consequently, we have defined a reproducible method to replicate the joint’s solidification. We can perform disassembly tests with varying joint conditions by replacing the a priori data with the clamping force, substituting the engine’s individual operational scenarios. The induced contact pressure generates the force that opposes the disassembly force. With this replacement model, we have achieved the first key objective by implementing the clamping force to validly and reproducibly model and substitute the joint’s solidification.

Thus, as in the simplified model in Sect. 3.2, the disassembly force must be above the resulting solidification force to allow the blade root to be disassembled. Likewise, a material-specific force limit must also be adhered in order to prevent damage to the blade root and thus ensure a component protecting disassembly. The following section presents the development and results of a component-friendly disassembly process.

4 Component-Protecting Disassembly

The second key sub-objective focuses on the development of a component-protecting disassembly. As we have shown, the applied disassembly force is an essential factor when detaching the blade root. To separate the joint between blade root and turbine disk, a minimum force must be applied while at the same time adhering to a force limit. That allows the blades to be disassembled without damaging them, ensuring component protection. The following section presents our research on determining the material-specific force limit for HPT blades used in our experiments.

4.1 Determination of Force Limits

The goal of a component-protecting disassembly is to detach the operational solidified joint between blade root and turbine disk without damaging components. Therefore, we determined the force limits that can be exerted on the blade root. The manufacturer’s specifications in the engine manuals did not consider the application of a load as we plan to exert on the blade root.

In order to determine the force limit, we prepared already disassembled HPT blades which the manufacturer provided to us. We cleaned, labeled, and photographed the blade root to capture its condition before the investigation. Using a universal testing machine for tensile and compression tests, we performed static load tests. For this, we exerted a defined static force in 10 kN increments on the blade root using a tungsten carbide ram with a diameter of 12 mm and increased it successively until a significant plastic deformation was visible.

Up to a force of 20 kN, traces of the ram were observable by visual inspection. From 30 kN on, more visible imprints were observed until the imprint became clearly visible above an applied force of 60 kN. At higher forces, more and more distinct indentations were formed, which showed a palpable plastic material deformation. Above 70 kN, a significant deformation and thus damage to the blade material resulted.

After the visual inspection, we examined the blade roots using a confocal laser microscope for a more precise inspection. Figure 10 shows excerpts of the results of the tests for a static load of 20 kN and 70 kN: Up to a force of 50 kN, we could not observe any deformation within the measuring range. At forces up to 60 kN, the slightest deformation could be seen, besides indications of leveling of the surface’s roughnesses below 10 μm. From a force of 70 kN and above, a definite imprint was measurable.

Fig. 10

Examination of the surface using a confocal laser scanning microscope (rotated by 90°) after a force exertion of 20 kN (a) and 70 kN (b) (Middendorf et al. 2022)

The investigation into the component protection results can be concluded: For a component-protecting disassembly, the material of the blade root must not be affected in any way. The maximum force can be determined using the presented destructive test method. With that, damage during disassembly can be prevented. In the case of the HPT blades used in our research, the maximum force is less than 20 kN at 12 mm ram diameter or a static load of 44.2 MPa, since no damage was observed. However, an accurate assessment of the surface and component defects requires the manufacturer’s agreement. We have published further in-depth findings in our study (Middendorf et al. 2022).

4.2 Reduction of the Disassembly Force

With the knowledge of the maximum applicable load, the subsequent task to develop the component-protecting disassembly focused on reducing the disassembly force. The Response Surface Method (RSM) allows identifying and describing the influential disassembly process parameters affecting the force to dismantle the joint between the blade root and turbine disk. The experiments were based on investigations to substitute the operationally solidified joint.

The initial step of the RSM was the specification of the input and output variables (Witek-Krowiak et al. 2014). The disassembly force is the decisive factor for a component-protecting disassembly. Therefore, we determine it as the investigation’s output parameter. The selection of the influential parameters is shown in Fig. 11. Besides the disassembly force, we identified the disassembly speed as having a major influence, as it affects the capacity planning. With slow disassembly speed and, therefore, long times, not only the following regeneration tasks are delayed. It also delays following re-assembly tasks, leading to long slack times. We also include the vibration-determining adjustment variables of the piezo stack actuator-amplitude, frequency, waveform. The mentioned a priori data, such as operating hours, and LTO cycles, are replaced by the described clamping force, which induces the contact pressure replicating the joint’s solidification state. In order to further investigate the dependence on the blade root’s shape and to decouple geometric properties, we take the shape into consideration as well.

Fig. 11

Ishikawa-diagram for the influences on the disassembly force

The following step of the RSM is the design of an experimental plan. The design of experiments allows with a limited number of experimental runs adequate informative value of the results (Witek-Krowiak et al. 2014). Using a face-centered composite design of experiments (CCF), we executed 72 experiment runs for each shape of the blade root. Exemplarily, we present the results for the “Original” form, which corresponds to the original HPT blade root. The following regression analysis showed a good model fit with a coefficient of determination R² equals 0.961.

Figure 12 shows the standardized regression coefficients of the analysis. As in the previous studies, the clamping force as a substitute solidification model has the weightiest influence. The influence of the piezo’s amplitude and disassembly speed is similar in comparison. However, the influence of the frequency is noticeably lower. The comparison of the waveforms reveals the triangle waveform to have a greater influence than the sinus waveform.

Fig. 12

Pareto chart of parameter’s standardized regression coefficients for RSM analysis

In order to evaluate which influencing variables reduce the maximum disassembly force, the signs must be observed. The results showed that increasing the piezo’s frequency (−0.093), amplitude (−0.129) and their interaction (−0.203) decreased the disassembly force, as expected. Although an increase of the disassembly speed (−0.183) decreased the disassembly force, its interaction with, for example, the piezo’s frequency (0.165) and amplitude (0.218) had a negative effect on the reduction of the disassembly force. The interaction between the disassembly speed and the waveform has a negligible influence (−0.029 for sinus and −0.070 for triangle). When comparing the waveform, the sinus (−0.012) decreases, and the triangle (0.222) increases the maximum disassembly force compared to the sawtooth waveform. However, both the sawtooth and triangular wave-forms have an uncomfortable acoustic effect. More complex relationships existed with the interactions of the clamping force. The experiments and the statistical evaluation confirmed the physical assumption of the clamping force having the most considerable influence, which is supported by the standardized regression coefficient of 0.975.

Since the aim of the experimental investigation is the minimalization of the disassembly force, we calculated its maximum reduction in the next step. According to the results of the RSM, we obtained optimized setting parameters, and we were able to calculate the maximum reduction of the disassembly force that can be realized. The RSM analysis showed that the amplitude and frequency must be maximum and the disassembly speed minimum to reduce the disassembly force. In order to show the disassembly force’s complex dependencies and interactions on the disassembly speed, we want to present the results for the levels chosen in our study. Also, we show the dependency on varying the vibration’s waveform using the sinus and triangle vibration. Table 1 shows the parameters we set up for this investigation.

Table 1 Optimized parameters for maximum reduction of the disassembly force (Blümel et al. 2023b)

To investigate the maximum reduction of the disassembly force, we randomized and repeated a series of experiments, 45 runs total. Table 2 shows the mean values of the results. The tests were performed at a clamping force of 4,000N. Therefore, the calculated values are only valid for this clamping force. At other clamping forces, the values for the reduction also change and can be more or less.

Table 2 Result of the maximum reduction of the disassembly force (Blümel et al. 2023b)

As seen in Table 2, the maximum reduction of the disassembly force was achieved using a disassembly speed of 1 mm/s and sinus waveform for the superimposed vibration. By increasing the disassembly speed, the achievable reduction of the disassembly force decreases when superimposing triangle vibration. When superimposing sinus vibration, the reduction of the force is more complex. Here, the interactions of the individual inputs became noticeable. Thus, in this validation test, the required maximum force was reduced by 20.4% with a sinus wave and 16.5% with a triangular wave compared to disassembly without using the piezo actuator. As seen in the varying reduction for increased speeds, the interactions of the influencing variables have a non-negligible effect on the reduction.

In addition, the disassembly force curve over disassembly time for vD of 1 mm/s was exemplary, plotted in Fig. 13. The test rig’s pushing rod was moved to contact with the sample. After this, the disassembly force stagnates briefly and then decreases continuously over time.

Fig. 13

Maximum reduction of the disassembly force for varying waveforms with vD of 1 mm/s

Occasionally when running the tests, some blade root samples tilted slightly in the disk segment just before the end of the disassembly process. That resulted in a slight increase in the disassembly force until the samples fell out of the disk segment. Overall, the graph also shows that the use of a piezo actuator reduces the disassembly force. According to our second sub-objective, we demonstrated a method for component-protecting disassembly by reducing the maximum required disassembly force. As long as material-specific force limits are adhered, the disassembly speed can be increased to reduce cycle times, despite increasing the disassembly force. Thereby still component protection is achieved while optimizing cycle times.

Based on the results of the RSM on the reduction of the disassembly force, we developed a learning model. Using the condition of the joint, affected and altered by the engine’s operation, it estimates disassembly forces and times, as presented in the section below. Thus, the optimized parameters are calculated individually for each occurring joint’s condition.

5 Learning Model to Predict Blade Specific Process Parameters

As we have shown, optimized process parameters for reducing the maximum disassembly force and thus component-protecting disassembly can be calculated as a function of the clamping force. The clamping force is a replacement model of the real solidification between the blade root and turbine disk caused by the engine’s operation. Therefore, our third sub-objective is the interconnection between the operational characteristics and the resulting optimized process parameters to dismantle the joints. However, since the joint’s condition varies depending on the individual engine’s operations, an estimation of the respective a priori data is necessary. Therefore, the aim is to use that data to estimate the necessary disassembly force and its resulting reduction by optimized selection of process parameters. In order to achieve the aim, we have developed a learning model, as seen in Fig. 14. The a priori data, substituted by the clamping force as the replacement model for solidified joints, serves as input. Based on the data, the learning model predicts and determines the optimized process parameters, adaptive to the joint’s condition. In a later industrial application, the replacement model can be re-replaced with the real operational data (a priori data), which will then be correlated with documented disassembly forces of real aircraft engines’ disassembly tasks.

Fig. 14

Disassembly process using a learning model to predict process parameters, based on operational data (a priori data), substituted by a replacement model

The basis for the development of the learning model are the results of the previously conducted research and RSM analysis, as we presented in (Blümel et al. 2023b). We trained a learning model that, similarly to the experimental investigation described in Sect. 4.2, uses a multiple linear regression algorithm with the results of the RSM analysis. In addition, we performed further disassembly runs with randomized values for the input values to obtain a test subset for the trained model. That gives us an approx. 75 to 25% split between the training and test data set. In order to evaluate the trained learning model, we calculated the coefficient of determination R² and the symmetric mean absolute percentage error (sMAPE), being a coefficient to evaluate machine learning and regression studies (Chicco et al. 2021). An R² equals 0.9248, and a sMAPE equals 8.594% indicate good predictive performance. Using the same data set, we have also developed a feed forward neuronal network with two hidden layers. With a calculated R² equals 0.9251 and a sMAPE equals 8.635%, it indicates a slightly better fit. However, the neural network’s root mean square error (RSME) being higher than the regression’s indicate a more inaccurate result. We have concentrated on the regression being more descriptive. Yet, a more extensive database might improve the neural network’s results. The design and operation of the learning model is shown in Fig. 14. When the maintenance task is planned, the disassembly planning is provided with the operational characteristics of the engine (a priori data), in our case, substituted by the clamping force of the replacement model. Using the IAE V2500 high-pressure turbine’s first stage as an example, the material-specific maximum applicable disassembly force is known and also provided to the learning model as a priori data. The disassembly parameters of the previously executed disassembly tasks, which were used to train the learning model’s algorithm, represent the database. With the trained algorithm, the learning model calculates the optimized parameters to reduce the disassembly force. The disassembly speed determines the time of each dismantling process and is the key factor affecting the disassembly process and following tasks planning. The learning model attempts to keep the disassembly speed at maximum, while adhering to the force limit to achieve shortest process times. By optimized selection of the piezo stack actuator’s parameters, amplitude, frequency, and waveform, the disassembly force is minimized. After each successful disassembly task, the learning model calculates the difference between the measured and material-specific maximum disassembly force. If the difference is in between a specified safety margin, the disassembly speed will be increased while adapting the vibration’s parameters to continue minimizing the disassembly force for the following disassembly runs (Blümel et al. 2023b).

To illustrate the functionality of the learning model, we present an exemplary application on the “Double” HPT blade root contour in a disassembly environment. Figure 15 shows ten individual disassembly executions, for example, ten blades that are disassembled one after another. At the beginning, the disassembly is executed at a previous defined low speed, which is gradually increased by the learning model in following disassembly executions, as illustrated with the black squares. It can be seen that the green line as the material-specified maximum disassembly force, is not exceeded by the measured maximum disassembly force, indicated by the green bars. By adjusting the vibration’s parameters of the piezo stack actuator, even the disassembly speed can be increased. That allows increasing the speed to a previously defined maximum (6 mm/s in the shown example). All subsequent blades can then be disassembled with these parameters.

Fig. 15

Diagram of an exemplary disassembly process including ten runs

In addition, the learning model can recalculate the degree of solidification, in this example, the clamping force after each successful disassembly run. The blue bars show the measured clamping force, whereas the blue stars show the recalculated clamping force, replacing the joint’s solidification. The comparison of the measured and recalculated solidification allows traceability and an evaluation of the learning model’s performance. The learning model thus has information about the degree of solidification, which the model can calculate. From the variation of the measured clamping force, the learning model’s adaptability to varying joint conditions is recognized. The parameters of each disassembly run are stored in the learning model’s database and can be used to improve its performance (Wolff et al.2019).

Based on the operational data of the engine, the learning model can, therefore, estimate the disassembly force and calculate the optimized process parameters. If an engine is to be disassembled with an identical operational usage scenario, the setting parameters learned can be reused. However, for a disassembly task of an engine with a different usage scenario, the learning model will estimate the disassembly force and calculate optimized process parameters using an initially low speed, as described previously. Thus, if the flight is unknown to the learning model, within a short approximation interval of, in our case, ten out of 64 blades, the learning model adapts the disassembly speed as well as the vibration’s parameters amplitude, frequency and waveform, to ensure optimal disassembly times.

For the adaptability of the learning model to different engine types and, therefore, blade root contours, we investigated the transferability of geometric dependencies. In extensive FEM simulation studies, we investigated a parameter with geometric information (geometric parameter, Wolff et al. 2018). Figure 16 shows the FEA for two different contours. Using the coefficient of friction and an offset, i.e., an interference, of the contacting surfaces, we substituted the solidification of the joint. The geometric parameter was identified by varying the simulated solidification of the joint and analyzing geometric dependencies. That geometric parameter allows the decoupling of the joint’s solidification from geometric properties as well as estimating the disassembly force of a yet unknown contour based on an already disassembled contour. The parameter, independent of the condition of the solidified joint, indicates how much the required disassembly force differs from a contour “A” to a contour “B”. To illustrate, Eq. 2 shows an example regression equation for estimating the disassembly force with x₁ and x₂ being arbitrary inputs. The standardized coefficients β_i are the estimators to indicate each input’s weight. The blade root is integrated as a binary variable: When disassembling contour “A”, d_A will become equal to one, and d_B equal to zero. That leads to omitting β_B · d_B and adding β_A to β₀. Contour “B” is vice versa (Bluemel and Raatz 2021).

$$F_{D} \, = \,\beta_{0} + \beta_{1} \cdot x_{1} + \beta_{2} \cdot x_{2} + \cdots + \beta_{A} \cdot d_{A} + \beta_{B} \cdot d_{B}$$

(2)

Fig. 16

Exemplary result of the FEM simulation for two different contours

Aided by the FEA, depending on geometric properties, we can estimate the contour-specific regression coefficient β_J of yet not disassembled contours. Thus, the coefficient β_J is decoupled from and independent of operational data. The estimate in advance increases process reliability, as the learning model is not required to calculate the yet not disassembled contours regression coefficient β_J through progressive learning.

For a vivid presentation of the developed adaptable and component-protecting disassembly, we engineered and built a disassembly workstation, as shown in the next section.

6 Development of an Automated Disassembly Workstation

We engineered and built a disassembly workstation integrated into the CRC 871’s system demonstrator to summarize the results. In addition to the disassembly environment, the workstation includes a robot-assisted handling system (Fig. 17a).

Fig. 17

a Automated disassembly workstation, b Component-protecting gripper based on soft robotics (Blümel et al. 2023a)

The disassembly workstation consists of a section of a turbine disk in which a blade can be prepared as an example. The disassembly environment is identical in structure and function to the test environment. However, a robot gripper supports the disassembly process: During the disassembly, the blade is held securely so that it does not drop when it is completely detached. The blade held by the gripper is then transferred to the handover station in the workpiece carrier developed in the CRC 871 and is available for the subsequent steps of the regeneration process.

In order to ensure component protection also during handling tasks, the robot gripper is developed from soft robotics, as shown in Fig. 17b (Blümel et al. 2023a). The soft surface of the gripper jaws allow a flat and adaptable contact with the complex shape of the blade’s surface.

An actuator inspired by an origami structure controls the opening and closing of the gripper. In addition, the gripping force can be adaptively adjusted. Compared to a rigid gripper, this can compensate for slight tilting that can occur when holding the blade during the disassembly process. The flexibility also ensures that the blade is held evenly and can be held more securely overall. The soft gripper can also insert the blade into the workpiece carrier. However, current work and research is still focused on improving the service life by investigating and selecting suitable gripper materials and optimized gripper design.

7 Conclusions

The presented research on component-protecting and adaptable disassembly demonstrates possibilities to integrate the disassembly characterized by uncertainties into the regeneration chain. The varying product histories and usage scenarios lead to significant differences in disassembly processes. Therefore, disassembly is usually carried out by manual labor. Considering aircraft engine high pressure turbines, we established a technologically plannable disassembly despite characteristic uncertainties. Since the engine type and scenario data are known a priori, our objective was to interlink this a priori data with disassembly planning parameters.

Due to the lack of operational altered and solidified joints between blade root and high-pressure turbine disks, we developed a reproducible method to replicate the joint’s solidification. A clamping force was used to replace operational solidified joints between blade root and turbine disk. Thus, we were able to perform experiments on component protecting disassembly with varying joint conditions.

For component-protecting disassembly, i.e., to prevent damage on components, it is necessary to keep the required process forces to a minimum. As known from the literature, vibrations, depending on their parameters, reduce the coefficient of friction of objects in contact. We were able to successfully demonstrate transferring that effect to the disassembly of operationally solidified joints. By using a piezo stack actuator, which superimposes vibrations on the disassembly movement, a reduction of the maximum force required to disassemble the joining partners was achieved. Reducing the disassembly force ensures that material-specific force limits are adhered, which prevents damage. Since the joint’s condition varies depending on the individual usage scenario, we developed a learning model to adapt process parameters to varying a priori data.

The learning model was trained using experimental data and integrated into an automated disassembly environment. Using a regression model, the learning model estimates the disassembly force based on the joint’s condition, known from the geometric data and engine’s operational characteristics. With that data, the learning model chooses optimized parameter settings to adhere to material-specific force limitations. That allows adaptive adjustment to the uncertain and varying product conditions. Ultimately, the learning model predicts needed process times for each disassembly execution, enabling process planning to be carried out before initial disassembly steps.

Eventually, we illustrated the application and suitability of the method on a system demonstrator. With this system demonstrator, the transferability of the developed learning model to an exemplary automated disassembly workstation including subsequent handling processes, integrated into the process chain developed in CRC 871, could be successfully demonstrated.