Tool management optimisation through traceability and tool wear prediction in the aviation industry

In production, data acquisition is a key factor in improving availability and reliability of input data and to gaining data-based transparency. Hence, it provides the basis for optimising the coordination of production such as an improved tool management [5]. Traceability systems enable to increase transparency and provide optimisation potentials [6, 7]. Traditionally, traceability systems are used for proof of origin, product authentication, product liability and recalls. But there are more digitally advanced use cases that make use of the generated data of traceability systems actively during production in real-time. Among these are digitally documented work, real-time process monitoring, process interlocking, dynamic process control, process analysis and inventory management [8]. The potential of traceability systems is exploited when it is deployed as data source system enabling continuous data generation through entire production processes. The generated data is the basis for comprehensive process analyses, such as the identification of bottlenecks [9]. Current research is still lacking to show concepts on how to consider the most relevant aspects when implementing traceability systems such as technology alternatives, data acquisition and usage as well as the needed marking strategy covering the entire process that needs to be tracked.

Monitoring tool wear is an essential information to be considered in the maintenance strategy of tools. Conventionally, tools are replaced time-dependent or due to a specific failure. As a result, the remaining wear reserve of a tool is usually not fully utilised [10]. Therefore, research suggests data-based concepts to be a contemporary solution for achieving a reliable tool maintenance (TM) strategy [11‐14]. Data-based TM functions as extension of condition-based maintenance and offers the potential to continuously include real-time machining data into a company’s maintenance strategy [15]. Considering digitisation on the one hand and the need for sufficient wear indicated TM on the other hand, in those concepts, data processing technologies such as machine learning are applied to production processes [16]. They enable the prediction of the remaining tool life, based on wear-correlated machine signals [13, 17]. Although different concepts for tool wear prediction exist, a systematic guideline for practitioners is still missing in recent literature.

The SME’s objectives to improve the tool management with the abilities of a traceability system mentioned above, results in the identified requirements outlined in Fig. 2. To fulfil the objectives in the underlying use case by realising a demand-based traceability system, two major steps are conducted:

The individual configuration options can answer important questions about data generation such as what is tracked, where it is tracked and when it is tracked [19]. According to a comparable industrial use case where a traceability system is implemented [20], the configuration process to be conducted is divided into two steps (see Fig. 3).

The first step deals with the selection of the traceability objects that need to be tracked by the traceability system. Figure 3 gives an overview of possibly relevant traceability objects. The grey boxes represent specific objects that generally can be tracked in any production process, such as equipment, product, tool, and worker. Due to the use case requirements in Fig. 2 “tracking of all tool movements—(TR 1)” and “implementation of individual marking of tool and tool holder—(TR 2)”, the machining tool and the tool holder need to be tracked as illustrated in Fig. 3. In the second step the acquisition granularity is selected. The acquisition granularity deals with the data handling of the previously selected traceability objects. It defines the data points needed to be captured for the selected objects. In step 2 of Fig. 3 there are four dimensions of the acquisition granularity that were developed in comparable project [20]: acquisition level, object granularity, acquisition scope and acquisition density.

The acquisition level as a first part of the acquisition granularity determines where the traceability objects are being tracked. Tracking production areas such as machining, quality and assembly areas represents a coarse data acquisition compared to individual work station level recording of object movements within the areas. Considering (TR 1), the SME’s management aims to implement an acquisition level of the traceability system that tracks every tool and tool holder at least at every single work station marked with AG (see Fig. 3) of the TC process (see Fig. 1: Identified problems in the SME’s tool cycle in accordance to [1, 4]). The object granularity specifies whether every part is tracked as a single part, in a certain batch size or in bulks consisting of several batches. Requirements (TR 2–5) have in common that they require the individual tracking of every machining tool and tool holder so that single part is selected as object granularity (see Fig. 3). The acquisition scope defines the type of processes being tracked. Including transformation processes that change the appearance of the object, storage processes where the objects are planned to be stored, transport processes that are important from an intralogistics perspective and handling/ preparation processes that give more detailed information about necessary supporting activities [20]. The TC contains all types of processes on work station level, so that according to (TR 1) all process types are considered to be captured by the traceability system (see Fig. 3). The acquisition density determines how many capturing points are needed. Within the TC, any work station is captured when a tool arrives (start point) and when it leaves (end point) to the next process step in the TC (see Fig. 3). Summarised, as the configuration options marked with TO and AG show, the traceability configuration intends to track machining tools and tool holder in a single part granularity (what), considers all transformation, storage, transport and handling/preparation processes (where), at the work station level with start and end point (when) [19].

After the first major step of specifying the configuration of the traceability system, a seamless marking strategy throughout the complete TC needs to be developed to fulfil the requirements in Fig. 2. Typically, an auto identification concept is used to develop and visualise a seamless marking strategy [21]. The auto identification concept shown in Fig. 4 is an extended version to show not only the auto identification flow and marking type, but also fitting technology alternatives, the step-by-step traceability data acquisition and finally the data usage. These information are explained in detail below Fig. 4:

The continuous tracking via RTLS provides a more detailed movement monitoring than the discrete tracking. Within this project, this benefit does not outweigh the high costs of integrating RTLS tags into the tool holders. The high costs are caused by an expensive technical development process as this is not a standard solution on the market. For cost reasons option 2 is selected.

Besides developing a demand-based traceability solution, the utilisation of machining data in combination with traceability data such as time stamps, tool ID and station ID (see Fig. 4, step 6 and 7) enhances the TC transparency to determine the individual tool wear. The combined use of those sources enables companies to improve their individual tool usage (see Fig. 1) by determining the individual RUL of tools. Together with capturing the individual tool’s history this raises several requirements (TMR 1–4) for implementing data-based TM in manufacturing processes (see Fig. 2).

In this context, the traceability solution functions as data supplier for individual tools enabling the matching with recorded machine data. The introduction of data-based TM complements these characteristics by enabling the processing of large amounts of data from use-case specific process conditions. In literature, the implementation of data-based solutions in manufacturing environments follows nine stages. Among those, the stages technical understanding, technical realisation and technical implementation are important success factors [23]. Generally, these stages have a universal meaning and are not specified for TM. Therefore, from a practical perspective and under consideration of the identified use case objectives in Fig. 2, six fields of importance were identified (FOI 1-FOI 6). They refine the technical understanding, realisation and implementation and address possibly occurring problems by using a data-based TM. Every single FOI is provided with a guiding question (Q 1-Q 6), covering important aspects to be considered for the successful implementation in manufacturing companies (see Table 2).

In the following, a literature analysis is used to gather potential answers to the guiding questions addressing companies to create a use case specific strategic roadmap including important hurdles to overcome. Afterwards the results are validated based on the underlying use case under consideration of the TMRs from Fig. 2.

To start the data-based TM implementation, the company’s overarching objective must be defined. The solution to be developed strongly depends on the followed strategic path (Q 1) to consider benefits of certain methods as well as use case specific process requirements. Two possibilities are conceivable to be implemented. First, on tool level, extracting information from machine data allowing to draw conclusions about the individual tool RUL [24]. Second, on process level, the use of extracted features from machine data to monitor process behaviour by data-based anomaly detection (AD), or statistical approaches such statistical process control (SPC) [25, 26]. Such data-based approaches require a sufficient foundation of reliable data and an adequate degree of digitisation to fulfil the desired impact (Q 2) [27, 28]. Generally, the integration of sensors, edge devices and cloud solutions increase the process-digitisation and therefore is beneficial. These technologies ensure the reliable data collection, -storage and -processing. In case the digital maturity level is too low, or the required degree cannot be reached, a less data-dependent approach such as SPC must be selected.

Since proceeding tool wear is the crucial criterion for TM [3], it must be defined how tool wear can be determined in the first place. Literature suggests flank wear to be a suitable indicator to represent the current wear, caused by different wear mechanisms (chemically, physically, thermally) [16, 17, 24, 29, 30]. Afterwards it must be decided, how the wear is measured (Q 3), based on available knowledge about tool degradation and process characteristics such as available sensors and predefined wear threshold levels [26, 31]. Direct wear measurements via visual inspection techniques and indirect measurements of wear indicators from machine signals can be distinguished [16, 26, 32]. Direct tool wear measurements require process stops for measuring due to the continuous interaction between workpiece and tool as well as the continuous flow of coolant fluids during production. Those measurements are difficult to include in a manufacturing process. But they offer information on the actual tool wear, measured by geometric changes of the tool. Indirect measurements only estimate the wear, based on wear indicators extracted from machine data but at the same time offer an easy way for real-time in-process monitoring [2, 24, 32].

When wear is measured reliably and relevant process and machining mechanisms are understood, a suitable model for representing the proceeding wear can be developed (Q 4). Three categories of tool wear prediction can be distinguished. Physics-based approaches identify the wear degradation function using physical laws (e.g., mechanically, chemically, and thermally) as well as empirical expressions. Sensor-based approaches use sensors to measure tool wear indicators indirectly. Hybrid approaches combine prediction model results and inference data from sensor-based approaches [16]. All forms of representation require the use of domain knowledge and data science while providing advantages and disadvantages. There exist different physical models and a solid understanding on tool wear mechanisms in action, but the physical models oftentimes are associated with modelling uncertainties. Furthermore, sensor-based approaches allow real-time monitoring, while their monitoring quality strongly depends on measurement- and sensor noise. Typically, hybrid approaches are more complex and require equal competencies in terms of domain knowledge and data science. Hence, indirect wear measurements favour sensor-based and hybrid approaches, while direct measurements are useful for physical approaches [16, 33].

Based on the representation of tool wear, different RUL prediction options are conceivable (Q 5). First, similarity models or transfer models use the knowledge of similar processes to apply learned knowledge to unknown problems [34]. They are favourable, if historic data from similar processes, components or machines is available and the company has the capability to extract information from this data [34]. Second, degradation models utilise previous degradation behaviour (regression). They consider several degradation processes at the same time or use statistical fitting to represent complex degradation processes [31]. Therefore, they are useful, if knowledge on physical approaches for wear representation has already been gained or is gained in parallel by executing experiments. Third, survival models assume that the failure rate correlates with time and its covariates and therefore induce the reliability of a tool under certain circumstances [35]. They require knowledge of so-called survival and hazard functions, providing the probability of tool-survival at certain timestamps and inducing the risk of experiencing a failure event at a given time [36]. Besides RUL prediction, TM can be realised by the detection of abnormal process states, based on machine signal analysis. AD systems utilise the differentiation between regular machining and abnormal conditions [37, 38]. They are trained solely on regular machining data and identify abnormal states according to the learned characteristic features and patterns [39, 40]. Although AD may not be able to predict the tool RUL, even the in-time detection of tool indicated abnormal process condition is beneficial. Especially due to the typical behaviour of tool failures that oftentimes occur spontaneously [38].

Executing the presented prediction and detection approaches leads to complex problem-solution-environments including multiple input sources as well as merging machine signals and sensor data [41]. Oftentimes those environments only can be handled by utilising advanced computer techniques such as Artificial Intelligence [14, 42]. These techniques play a decisive role against the background of the increased relevance of predictive analytics in industry [26, 43]. In this context, research mainly focuses on comparing the performance of different algorithms (FOI 6) to address the tool wear prediction problem [24, 28, 44]. The SME’s use case shows that selecting the appropriate algorithm reaches too far ahead at the beginning of implementation. Practitioners rather benefit from a structured and practical guideline, covering the identified FOI one to five. Such a guideline for practical implementation of data-based TM does not exist in research, yet. Therefore, a literature-based morphology, including the guiding FOI and different inherent characteristics, is derived as shown below, guiding practitioners through important fields on their way to data-based TM implementation (Table 3).

For the validation of the developed morphology, the underlying use case in an aviation company is used. Following the proposed morphology, the first step is to select the overarching objective (Q 1). Derived from the SME’s list of requirements in Fig. 2, “predicting remaining useful tool life” is chosen, addressing the problem of uncontrolled tool failure and unplanned downtimes. Next, the hurdle of digital process maturity (Q 2) must be overcome. Through the developed traceability solution in combination with retrofitting an edge device, the process digitisation is enhanced, so that data is gathered and combined in an individual tool history, including tool name, location and number of jobs executed. This multi-source data builds the foundation for answering guiding questions Q 3 to Q 6. Considering Q 3 the SME’s individual process conditions must be considered. A real production process with active coolant fluid flow is observed. This indicates using indirect tool wear measurements. But several different products are manufactured on the considered machine. Therefore, the retrofit of additional sensors, required for indirect measurements, is prevented due to complex and diverse axis traverse paths. So, direct wear measurements are executed for tool wear labelling at defined timestamps according to a predefined experimental plan. By monitoring several geometric features, such as flank wear, diameter, radial angle, rake angle and cutting edge rounding, wear is quantised as preparation for developing a hybrid wear representation (Q4) including two steps. First, the physical, wear-induced tool degradation is investigated. Based on several test-series and by utilizing the systematic direct tool wear measurements, the physical wear curve is approximated. Besides the SME’s choice to focus on RUL prediction instead of anomaly detection, the physics-based approach is chosen for wear representation, due to the lack of transferable information from similar processes or survival functions. Second, contemporary data processing technologies are used within a sensor-based approach, to identify the physical wear patterns in the machining data and match them with the gained knowledge about the physical degradation.

As knowledge about the tool degradation behaviour is gained by the executed experiments, a degradation model is used as prediction approach (Q 5). In the future, this procedure enables the SME to replace direct wear measurements via visual inspection by indirect wear measurements through a data-based TM solution utilising real-time machining data. The investigation, which algorithm to use for use case specific prediction is the object of ongoing research and still considers several tests with machine learning and statistical prediction approaches.