1 Introduction and motivation
Modern manufacturing systems face a growing demand for enhanced precision, heightened efficiency, stable performance, and rigorous safety levels. Additionally, these systems are designed to be flexible and adaptable to rapidly changing process-product-resource configurations or evolving materials. Such requirements necessitate high efficiency in collecting and utilizing process data for real-time monitoring, workpiece quality management, and condition-based maintenance, which entails integrating complex automation and sensor systems. The extensive utilization of sensors is an essential feature in industrial manufacturing. In addition to computer-aided simulation and numerical modelling, sensor-based systems offer advantages in managing high volumes of process parameters due to the extended tool and workpiece specifications, advanced process control, and high real-time requirements in the digitalization of manufacturing systems [1]. Furthermore, sensor-based systems capture process data that engineers can use to extract meaningful information for subsequent data analysis. Comprehensive data gathered from integrated sensors is beneficial to reduce machine downtime, protect the tools from failures, avoid scrap and, to maximize process efficiency by optimizing operating conditions.
In the context of Industry 4.0, data availability of the entire product life cycle necessitates implementing data-driven approaches to provide resilient and comprehensive process control. Furthermore, considering the increasing automation of processes and the reduction of qualified personnel, the quantity of data and quality are crucial to guarantee a reliable data-based model to characterize the state of a process or product. Combining data from high-performance sensors with machine learning (ML) algorithms can provide more robust models for classifying or predicting process conditions. Nevertheless, the performance of these models strongly relies on the quality of the data set, determined by the chosen data acquisition, preprocessing, and transformation techniques[1].
Anzeige
The appropriate selection and deployment of sensors are crucial for ensuring the quality of acquired sensor data in manufacturing systems, particularly in forming technology. Forming processes encompass a variety of operating principles and measurement conditions, e.g., high temperatures are characteristic of drop forging processes or high stroke rates of metal blanking. Consequently, measurement constraints and the type of sensors used differ from one process to another. Considering constraints derived from process and measurement conditions such as technical or data acquisition limitations is essential. Therefore, sensors in forming technologies possess domain-specific characteristics such as robustness to temperature, i.e., drop forging, high accelerations, i.e., metal blanking, and compatibility with formed materials. The selection and deployment of sensors in forming processes require knowledge of their type, operating conditions, integration in machines and tools, and domain-specific working principles [2]. For instance, the positioning of sensors in forming machines requires expertise from both experts and operators, directly influencing the accuracy of measurements and, therefore, the quality of the collected data [3].
To this end, this work aims to provide a more holistic and comprehensive overview of relevant interrelationships between sensor selection, deployment, and data processing to improve the implementation of data-driven methods. The approach emphasizes the influence of data acquisition methods and process characteristics on the quality and trustworthiness of the collected data. The main objective of the work is to assist engineers in choosing the appropriate data processing methods for efficient feature extraction for process monitoring of forming systems and tools by providing formalized sensor knowledge from different areas. To achieve this, sensor knowledge is modeled from two points of view: (i) engineering process design, i.e., technical considerations for sensor selection and deployment; (ii) operational data quality, i.e., obtaining and using sensor data efficiently for process monitoring. As a result, an extended sensor ontology is developed based on the W3C-published standard sensor ontology SSN/SOSA, which describes the considerations and constraints for selecting suitable measuring devices while illustrating the influence the physical environment has on the quality of the collected data.
The remainder of the paper is structured as follows: In Sect.2, a review of the literature on existing sensor ontologies and their applications in the manufacturing domain is introduced. Following, the state-of-the-art sensor data processing methods are presented to identify and highlight current challenges in data acquisition methods and effective feature extraction in forming processes. Subsequently, Sect.3 describes the main concepts of the proposed sensor ontology. The approach is then evaluated from a data analysis point of view on a flexible rolling machine. Finally, Sect. 5 concludes the paper and discusses potential future work.
2 Related work on data acquisition and pre-processing methods and sensor ontologies applied in manufacturing systems
This section begins with an overview of the current state-of-the-art sensor ontology development and applications, followed by a detailed review of related work in data acquisition and preprocessing methods in different manufacturing processes.
Anzeige
2.1 Sensor ontologies and their applications in the manufacturing domain
Ontologies, understood as “a formal, explicit specialization of a shared conceptualization,” have long been used to define sensor metadata in various application contexts to facilitate searching, sharing, and reusing sensor-related knowledge. The Semantic Sensor Network (SSN) ontology is a widely used sensor ontology [4]. The SSN ontology, formulated in the Web Ontology Language (OWL) and developed by the W3C Semantic Sensor Network Incubator group, encompasses four primary perspectives: sensor, observation, system, and feature and property. These perspectives collectively address sensory data collection, observation metadata, sensor network deployment, and property-specific, sensing, and observation capabilities. The ontology is further organized into ten conceptual modules, including Deployment and Operating Restriction, which handle concepts such as deployment lifetime, survival ranges, and the interconnection of sensors, observations, and properties. By leveraging the Stimulus-Sensor-Observation (SSO) design pattern, three of the four primary perspectives (sensor, sensed data, and observations) are interconnected, serving as the foundation for robust ontologies in the Semantic Sensor Web and Linked Data environments [3].
Schlenoff et al. expanded the SSN ontology based on requirements created by reviewing existing sensor ontologies [5]. In addition to the SSN ontology’s capabilities to describe sensors, measurement capacities, and attributes, the extended ontology additions include action (a relationship between entities where one interacts with the other) and domain (a collection of objects belonging to a specific manufacturing application). Moreover, the extension includes a library allowing the assignment of quantitative attributes comprising a numerical value and a unit. Further, Haller et al. added a modular structure to the SSN ontology[6], splitting the SSN into ontology modules, including a lightweight core, the Sensor, Observation, Sample, and Actuator (SOSA) ontology. SSN then extends SOSA classes and relations. Consequently, SOSA replaces existing classes with new ones that contain a broader description, allowing them to be used for multiple acts.
The CREMA Data Model Core module (CDM-Core) is an ontology previously used in manufacturing. Developed by Mazzola et al., it comprises a manufacturing-related, a domain-specific, and a use-case-specialized layer [7]. Since CDM-Core needs to effectively represent domain knowledge for its use cases, requirements were derived from user input and use case descriptions, involving annotating process models, services, and sensor data while ensuring logical consistency. CDM-Core then allows for annotation of process models in the Business Process Model Notation, semantic service description, or data stream annotation. Klein et al. deemed some of the CDM-Core ontology’s concepts, e.g., humans and geo-locations, too exhaustive for a use case in a simulation environment [8]. In their approach to describing a simulation factory, they subsequently developed FTOnto, an ontology based on the Manufacturing’s Semantics Ontology (MASON) and the SOSA ontology, using MASON’s three core concepts: Entity, Operation, and Resource. Key use cases of FTOnto are the enablement of Flexible Production Processes, i.e., approaches to optimize processes by identifying similar machines, and Condition Monitoring Analysis, i.e., predicting a future failure of system components.
In addition to developing sensor ontologies to formalize sensor-related knowledge, these ontologies have wide applications in various areas within the manufacturing domain. Sensor ontologies are used for diverse purposes, including sensor selection, system maintenance, and research data management. Riedel et al. suggested a knowledge-based selection of sensors and actuators within the planning of industrial plants [9]. Their approach enhances pre-existing, standardized plant description models, stored in formats such as AutomationML or CAEX, with data and semantics to use Knowledge Based Systems’ two-part structure. Instead of selecting during planning, Roh et al. proposed a method to associate sensor data with process variables in real-time to identify the optimal sensor combination for monitoring and controlling part quality during Metal Additive Manufacturing (AM) processes [10]. In addition, they developed an ontology-based network that establishes connections between sensing features, physical phenomena, process parameters, and part quality throughout operations, aiming to facilitate real-time defect identification and increase the quality of manufactured parts. Maleki et al. offered an additional proposal for sensor selection to assist with the maintenance of industrial machinery [11]. Their developed ontology is built as a Web Ontology Language (OWL) file and includes sensor information and other concepts directly influencing the sensor. Intending to analyze and forecast energy consumption in manufacturing, Wenzel et al. used OWL to define models of production systems and data [12]. This approach identified three models defined in separate ontologies to describe the processes: equipment, product, and operations. In the context of research data management, Bodenbenner et al. described six critical challenges associated with collecting, processing, and storing industrial sensor data according to the FAIR (Findable, Accessible, Interoperable, and Reusable) principles. Consequently, they introduced a three-layer architecture for the software responsible for sensor control and communication interfaces. An additional FAIRification layer was incorporated between the sensor network interface and sensor implementation layers, aiming to provide a comprehensive description of sensor measurement context and metadata, thereby optimizing adherence to the FAIR principles.
2.2 Data acquisition and pre-processing methods applied in production systems
Machining processes involve multiple parameters derived from tool and workpiece specifications. Furthermore, contact between the tool and the workpiece is limited to short timespans and small areas, leading to the characteristic complex behavior of such systems. Kuntoglu et al. reviewed the state-of-the-art sensor systems and signal processing methods in mechanical machining processes. In a first-of-its-kind overview, the authors comprehensively compiled standard sensors, signal acquisition, and signal processing applications in turning, milling, drilling, and grinding operations [3]. Sensor integration necessitates domain expertise for a comprehensive understanding of the mechanical system and working principles of the sensors to provide reliable and robust decision-making systems. Liewald et al. state the importance of considering relevant technology-related challenges for data acquisition in the context of forming technologies. Due to high accelerations and short tool engagement, expert knowledge is essential for designing the measurement chain, and aspects such as sensor type and position must be accounted for [2. Sensor signals are exposed to noise and signal contamination due to the physical, i.e., static, and dynamic behavior of manufacturing processes.
Kubik et al. investigate how the type and position of integrated sensors in two forming processes affect the outcome signals. To investigate these influences, they examine the quality of a force signal of an unworn punch as a reference. Following, they quantify the deviations in the measured time series of different sensor positions and types and the influence of dynamic effects derived from increasing the processing speed. As a result, there is an apparent influence between sensor type and location in the quality of the measured time series and that the dynamic behavior of the process is also an influence [1].
The available literature shows an increasing trend in investigating the implementation of new ML algorithms to characterize process conditions. There is a widespread consensus on the importance of feature extraction and selection to enhance the performance of such models. However, very few investigations focus on the influence of prior stages, i.e., data acquisition, on the preprocessing and transformation methods. For instance, Mende et al. investigate the relevance of domain expertise in feature extraction and selection. To do this, the authors compare the explainability and performance of an ML model between features extracted from automatic tools (tsfel1) and features extracted using expert knowledge. They conclude on the importance of domain expertise to engineer and select features to use as input for data-driven approaches [13]. Feature extraction is crucial for enhancing data-driven approaches and can be categorized into frequency, time, and time-frequency domains. Time domain features can be directly extracted from raw sensor data, including mean values, extreme parameters, standard deviations, skewness, and kurtosis. Features in the frequency domain can be obtained by transforming the time series data through Fourier Analysis, with methods such as Discrete Fourier Transform (DFT), RMS frequency, spectral kurtosis, or spectral skewness. Finally, time-frequency domain features can be used to consider features from both domains simultaneously. Standard techniques include Discrete Wavelet Transform, Short-Time Fourier Transform, and local mean decomposition, amongst others [14].
Nevertheless, feature extraction and selection alone are not sufficient. Kuntoglu et al. argue the importance of signal preprocessing and processing methods to organize, prepare, and condition the collected data to achieve a clear understanding and analysis. Signal processing methods involve amplification, filtering, normalization, and denoising according to the type of signal and system requirements Warke et al. conduct a systematic literature review to identify trends, gaps, and research directions for professionals in machining processes involved in tool design, maintenance, and optimization and establish the essential steps data engineers must consider extracting meaningful information from raw sensor data[14]. Finally, Kubik et al. find that data acquisition and signal processing combined with expert knowledge, followed by data preprocessing and feature extraction assisted by domain knowledge, improves model performance [1]. Identifying interdependencies between tool wear, process parameters, and workpiece characteristics presents a challenge in forming processes due to complex and non-linear patterns in collected data sets [2]. Therefore, a trustworthy and holistic method is needed to provide a comprehensive overview of data acquisition and preparation assisted by expert knowledge for the successful application of AI projects. Figure 1 provides a graphical overview of sensors, data analysis methods, and feature extraction methods applied in manufacturing processes, as derived from the literature.
Fig. 1
Graphical representation, derived from literature review, of sensors, data analysis methods and feature extraction methods applied in manufacturing processes
×
![]()
Table 1
Most common data processing methods, derived from literature review, used in different manufacturing processes
Application | Signal processing | Descriptive analysis | Time domain | Frequency domain |
|---|---|---|---|---|
Milling process | ✓ | ✓ | ✓ | ✓ |
Ball screw drives | ✓ | ✓ | ✓ | |
Blanking process | ✓ | ✓ | ✓ | |
Sheet metal forming | ✓ | ✓ | ✓ | ✓ |
2.3 Derived research gaps and paper contributions
There are still several research gaps in sensor ontologies and their applications in industrial contexts. Firstly, many existing sensor ontologies focus solely on limited aspects of sensor information provided by vendors for specific purposes, hence, overlooking comprehensive modeling of sensor information to support general sensor selection use cases. In addition, sensors used in production systems, particularly in forming processes, encounter specific challenges such as robustness and adaptability to harsh industrial environments, including high temperatures, mechanical vibrations, and signal contamination. Moreover, these sensors should be compatible with various materials and adaptable to different forming processes and control systems. Therefore, to reduce uncertainties when integrating sensors in forming systems and tools, sensor ontologies should encompass design information from vendors and deployment details, including interfaces and properties like precise temperature range, installation positions, and integrated process steps.
Furthermore, current sensor ontologies also overlook perspectives related to signal and data processing techniques. Most approaches found in the literature regarding sensor data processing focus on implementing ML algorithms to predict or classify process conditions, states, or other aspects such as product quality. Nevertheless, very few approaches have focused on the influence of data acquisition and subsequent processing and how these affect the performance of the used ML models. Additionally, few efforts have been made to research the dependencies of data acquisition methods and signal processing in forming processes, mainly due to the complexity and variety of their behavior. Especial attention must be put into modeling the potential relationships between sensor selection and integration with the corresponding suitable signal processing, data aggregation and preparation, and feature extraction methods to enhance the quality of acquired sensor data and subsequent data outcomes. Modeling these relationships using domain knowledge can lead to more robust ML models, providing a more comprehensive understanding of process states and behaviors.
Based on the identified research gaps in sensor ontology and data processing areas, the main contributions of this work and the proposed approach are summarized:
-
Extending the SSN/SOSA ontology with sensor components and types enables comprehensive sensor information modeling to support engineering process design.
-
Supplementing the conditions required for sensor survival and ensuring reliable performance, aiming to guide engineers in selecting and deploying sensors in forming systems and tools for effective data acquisition.
-
An extension of sensor properties and specifications, combined with domain knowledge, aims to support engineers in the data acquisition and processing stages in obtaining a trustworthy data set for a more efficient and robust process monitoring using data-driven approaches.
-
A holistic overview characterizing the interrelationships between data collection and processing methods, promoting efficient data analysis and aggregation, and contributing to the overall understanding of the data lifecycle.
3 The lightweight SSDP ontology
Based on the research gaps derived in Sect. 2, this work develops a lightweight sensor ontology based on the SSN/SOSA ontology. This approach focuses on supporting integrated sensors’ selection, deployment, and data processing of integrated sensors. Below is an overview of the main modules within the ontology, followed by in-depth descriptions of the key concepts and relationships.
3.1 Overview of the sensor ontology
The basic structure and the main modules in the SSDP ontology are displayed in Fig. 2. Details of each module and their relations are described individually in the following subsections. Building on the SSN/SOSA ontology, the SSDP ontology incorporates several concepts and relations from the SSN/SOSA modules, including property, condition, deployment, sampling, and observation. These modules are expanded with concepts crucial for sensor selection, deployment, and data processing in forming technologies. For instance, the SSDP ontology retains the class “SurvivalProperty” from the SSN/SOSA ontology, representing the extent of the system’s useful life under the specified conditions. On the other hand, the categories “GeneralPrperty” and “OperatingProperty” are extended in the SSDP ontology. Based on the collected datasheets of sensors used in the forming processes, subclasses such as “Size,” “OverloadCapacity,” and “MeasurementRange” are included in the “GeneralProperty.” At the same time, “Calibration” is incorporated into the “OperatingProperty” of the sensor. Additionally, the SSDP ontology introduces three main modules and one submodule. The “SensorComponent” and “SensorType” modules provide additional details on sensor elements and types, furnishing essential information for product and process developers and data analysts to fulfill their domain-specific tasks. For instance, certain sensors have extensions that execute specific signal-processing functions. These extensions can significantly impact the decision-making when selecting signal-processing steps for handling sensor data. The “Product-Process-Resource (PPR)” submodule is incorporated within the deployment section to establish connections between sensors, their mounting locations, and integrated process steps. These connections are crucial in determining parameters for data processing tasks, such as sampling rates or signal segmentation. Finally, the “DataProcessing” module compiles an overview of available feature extraction methods, signal processing steps, data preparation techniques for acquired sensor data, and their interrelations with other modules. The defined relations effectively bridge the gap between data analysis and inherent sensor knowledge.
Fig. 2
Overview of the SSDP ontology. Future expansions must investigate and include sensor calibration** methods such as manual/automatic, online/offline, dynamic/static, one-/two-/multi-point curve fitting and others. (Sensor Property submodule)
×
![]()
3.2 Concepts in sensor ontology related to sensor selection
This section presents a more detailed description of the main modules in the SSDP ontology, outlining the relevance of the newly introduced concepts and relations and discusses how to utilize this information.
3.2.1 Component and type
Sensors can be categorized based on various criteria, including their measurement principles, the measured physical properties, or application domains. Among these classification approaches, the measurement principle and the measured physical property are crucial considerations when selecting sensors for forming systems and tools. Therefore, introducing the “SensorType” module (see Fig. 3) in the SSDP ontology aims to support the initial filtering of available and suitable sensors based on process description and requirements. Two subclasses for the sensor type are defined: The “MeasurementVariable” class includes the typical physical properties the sensor can measure. In contrast, the “MeasurementPrinciple” class encompasses the underlying mechanisms by which the sensor detects these physical properties. Additionally, the sensor’s working principle is categorized into direct and indirect measurement, showing whether the sensor has direct contact with the measured physical property. This classification is essential, affecting factors such as calibration requirements and the selection of data processing techniques to ensure the accuracy and reliability of the results.
Fig. 3
Sensor type and component
×
![]()
Furthermore, a “SensorComponent” class has been incorporated into the SSDP ontology, detailing the potential elements present within a sensor. This addition provides supplementary information for sensor deployment and data processing from a functional perspective. In many cases, detecting the physical quantity requires mechanical conversion to activate the actual sensor element’s physical effect [15]. For example, a pressure sensor may utilize a diaphragm to transmit mechanical pressure to an attached sensor element strain gauge, which then converts its changed electrical resistance to its corresponding change in voltage. Moreover, many sensors also incorporate signal-processing elements, such as amplifiers and A/D converters for signal conditioning. Some sensors, known as intelligent sensors, are equipped with integrated microcontrollers, enabling extended functionalities like calibration [15]. Modeling the components within a sensor is advantageous not only for sensor deployments to determine if sensor calibration is necessary or if external bus system interfaces should be considered but also essential for identifying necessary subsequent signal-processing steps, such as amplification or filtering.
3.2.2 Conditions
Sensors only have reliable properties when satisfying specified working and survival principles. Information provided by manufacturers in sensor datasheets is designed to guide users in selecting suitable sensors for their intended applications. For instance, the operating temperature range in the datasheet specifies the lower and upper-temperature limits within which a sensor can function effectively. On the other hand, the survival temperature range indicates the values of extreme temperatures the sensor can endure without sustaining permanent damage. Understanding these operational and survival principles is essential for selecting appropriate sensors and analyzing sensor performance degradation and failures. Moreover, these principles are also decisive for deploying sensors, such as adding additional protection housing when the sensor requires a high level of protection class to ensure a stable performance during operation.
To comprehensively capture the constraints essential for sensor selection and deployment, the concept of “Conditions” found in the SSN/SOSA ontology is expanded (see Fig. 4). This expansion divides conditions into three categories: “Environmental Conditions,” “Installation Conditions,” and “SafetyConditions.” The “EnvironmentalConditions” category contains the external elements impacting sensor properties, including survival and operating temperatures, vibrations, and humidity. On the other hand, the “InstallationConditions” category addresses guidelines for sensor installation, encompassing factors such as the surrounding medium and maximum allowable levels of vibrations. Lastly, the “SafetyConditions” category outlines general safety criteria for sensors, including their IP class and explosion protection levels determined by the corresponding process. Given that industrial standards or guidelines commonly guide these protection requirements, the introduction of the “SafetyStandard” subclass serves to include relevant information in this regard.
Fig. 4
Sensor conditions
×
![]()
3.3 Concepts in sensor ontology related to sensor deployment and sensor data processing
The following section describes the modules in the SSDP ontology related to data acquisition methods, such as sensor deployment and data processing techniques. The goal is to characterize the potential relationships between modules and investigate if those described in subsection 3.2 have an influence.
3.3.1 Deployment and PPR
Manufacturing systems consist of numerous sensors producing large amounts of process data. Data-driven approaches enable engineers to capture and derive knowledge from the information captured in sensor signals, which, in many cases, can lead to increased process efficiency. Nevertheless, several factors determine the accuracy and reliability of these approaches. For instance, data quality, comprised of multiple dimensions, is essential to obtain a meaningful and reliable output. However, in the context of data-driven approaches used in forming technologies, previous work often overlooks the influence of data acquisition on the validity and quality of the data sets.
Process design in forming technologies strongly depends on expert knowledge and expertise due to the complexity and variety of operating principles and measuring and environmental conditions, impeding knowledge transferability between processes [2]. To ensure data quality, as seen in Sect. 3.2, sensors must be suitable for the corresponding data acquisition constraints, i.e., high accelerations, short tool engagement, or elevated temperatures. Additionally, to enrich the quality of the data set for subsequent processing and ML implementations, other aspects of data acquisition - i.e., data volume, speed of acquisition, or sensor position - must be considered and addressed appropriately. Consequently, the proposed ontology expands upon the “Deployment” and “Platform” classes in the standard SSN/SOSA ontology to capture the most relevant influences and dependencies between data quality and process constraints (see Fig. 5). In the SSDP ontology, the more extensive module, “Deployment,” encompasses the sub-module “PPR.” The reason behind this structure derives from the necessity of including domain knowledge, i.e., PPR knowledge, for overall optimization when deploying sensors into plants or machines. Furthermore, the “Deployment” class is divided into two sub-classes: “DeploymentInterface” and “DeploymentProperty.” For instance, lubricant is commonly used at the tool-workpiece interface to slow down abrasive tool wear in sheet metal forming processes [16], therefore conditioning sensor selection.
Fig. 5
Deployment and PPR
×
![]()
3.3.2 Observation and sampling
Data can be collected from a single sensor or via multiple sensors, known as data fusion. Typically, process data is captured from many sources, i.e., different sensors and sensor types, to enhance robustness and decrease uncertainty. Sensors used in forming technologies may produce thousands of data during short periods. Therefore, in the extended version of the SSN/SOSA ontology, the class “ObservationCollection” was introduced, comprising a set of single data points and properties captured by a specific sensor using the same principle (see Fig. 6). Data acquired in forming processes originates from multiple sources, leading to types of heterogeneous character data and extending the concept by including “DataType.” As previously mentioned, another challenge derives from the speed of forming processes, leading to extensive amounts of captured data during operation. Sensors typically produce time-dependent series data, especially in forming technologies, resulting in vast data sets that require appropriate capabilities. Hence, extending said concept of “SamplingCollection,” which includes “DataVolume” and “SamplingFrequency,” addresses constraints including storage and the velocity of the acquisition and transmission of the sensor data, correspondingly. An adequate sampling frequency must be selected; this frequency determines the number of samples collected per unit of time and depends on the characteristics of the input signal and manufacturing process.
Fig. 6
Observation and Sampling
×
![]()
3.3.3 Data processing
In recent years, the digitization of forming processes has gained more attention and effort to capture and analyze their behavior and characteristics during operation to enhance process robustness and understanding. The previous sections argued the importance of data acquisition and its combination with domain knowledge to ensure good data quality. Nevertheless, raw sensor data can only be fed as input to the ML algorithms with previous processing, i.e., signal processing, data preparation, and feature extraction, for further analysis [2]. Data preparation has been recognized as essential to achieving accurate, reliable, and robust models with the increasing implementation of data-driven approaches. Furthermore, data sets originate from multiple sources and may contain redundant or irrelevant information. To enhance support for subsequent data processing and knowledge discovery, the class “DataProcessingMethod” was introduced and linked to the observation data through the “usedMethod” relations. This addition aims to better support the manipulation and handling of raw sensor data by presenting available processing techniques. Furthermore, it establishes a connection between data collection and processing methods, promoting efficient data analysis and contributing to the overall understanding of the data lifecycle (see Fig. 7).
Fig. 7
Data processing methods
×
![]()
Moreover, the approach extends the concept of “DataProcessingMethods” with the required steps engineers must follow to effectively retrieve meaningful information from the collected data. “SignalProcessing” is an essential but often overlooked step for data preparation and conditioning. Sensor signals are exposed to unwanted noise, which must be eliminated (e.g., noise present in the time series data derived from the influence of vibrations due to the dynamic effects of the process). This class is divided into three categories: (i) “SignalProcessing,” (ii) “DataPreparation,” and (iii) “FeatureExtraction.” As previously mentioned, forming processes are characterized by the complexity of their behavior. Combining domain knowledge with subsequent signal processing is crucial, i.e., to differentiate between physical effects in the process from noise derived from other factors. Hence, further extending the ontology with the category “SignalProcessing” is decided. “DataPreparation” is more widely recognized as essential when implementing ML algorithms. It includes standard practices such as outlier detection or data segmentation. The latter is significant in forming processes since only some intervals of the time series data contain relevant process information. Finally, meaningful feature extraction is necessary for a robust and reliable ML model. However, forming processes sometimes present non-linear behaviors and many process parameters. Engineers must also consider the trade-off between model accuracy and computational effort. Therefore, special attention must be placed on data transformation and meaningful feature extraction from different domains (e.g., time domain, frequency domain, or time-frequency domain), with the help of expert knowledge, to prevent the extraction and selection of irrelevant or redundant features, resulting in reduced computational effort. The final category, “FeatureExtraction,” has been included to model the most suitable techniques for effective feature extraction in this context.
3.4 Applying sensor ontology for supporting effective usage of sensor data in forming processes
Implementing robust and accurate ML models depends on data preparation and the quality of the raw sensor data. High-quality data can be achieved by selecting the most suitable type of sensors and integrating them appropriately and supported by expert knowledge to understand process behavior and its influence on data acquisition. Therefore, the extended sensor ontology has been developed as the first step in capturing and modeling relationships between sensor selection and integration, process characteristics, and data processing. The goal is to support engineers in the early stages when selecting sensors to monitor forming processes by providing a comprehensive overview of all relevant considerations and constraints to obtain high-quality data and later process it most efficiently by understanding how it was acquired and from what source. This understanding will lead to higher-quality data sets (i.e., more accurate measurements) and a better understanding of the data sources and potential limitations in its acquisition. However, it is essential to note that working with standard sensors installed in systems and tools is preferable to minimize costs and maintenance work.
4 Case study – flexible rolling machine
The following section evaluates the proposed model and quantifies the influence of data preparation and feature extraction on the performance of ML models for predicting workpiece quality. Building on the knowledge presented in [17], a data set of process and quality parameters for 20 manufactured blanks, acquired on the testbed of a flexible rolling machine, is used. The depicted classes and relations captured in the SSDP ontology (see Fig. 2) are implemented in OWL using the Python package OWLready2 [18]. Furthermore, this section proposes implementating two new prediction models to demonstrate how the SSDP ontology can be utilized to support engineers in the data processing stage. To this end, a comparison of the performance between the regression model presented in[17] and the two models proposed in this work is provided. The ML algorithms used for the case study are commonly known as LASSO. This algorithm introduces a regularization technique (L1) in the loss function of a standard linear regression. This algorithm is often used in machine learning tasks to perform automatic feature selection. LASSO is beneficial when there is a vast number of features and a small number of observations. Once the sensor data is processed, it is divided into training (80%) and test data (20%). To evaluate the models’ performance, the coefficient of determination \(\left({R}^{2}\right)\) is used to compare the goodness of fit on the training set. On the other hand, the mean absolute error \(\left(MAE\right)\) is calculated for the test set and determines the predictive power of the models. Finally, the results and the benefits of using the SSDP ontology for data processing are discussed.
4.1 Process description
The flexible rolling machine consists of three main components: two rolling tools on the left and right sides and a rotary table in the middle. The workpiece is affixed to the rotary table and held in place by a blank holder. Each rolling tool can move independently along the x-axis in positive and negative directions, removing material from the workpiece. Additionally, the rotary table moves vertically in the z-direction, controlling the thickness reduction of the workpiece. Six force washers are installed in the flexible rolling machine to measure and monitor the forces acting on the workpiece during the forming process (see Fig. 8). This is considered a discrete forming process consisting of four strokes, each reducing or smoothing different workpiece areas with varying thicknesses to achieve the desired geometry of the manufactured blanks. The final quality can be measured based on seven criteria regarding the geometry. However, in this case study, only the homogeneity of the inner rolling zone is considered for quality prediction.
Fig. 8
Description of the flexible rolling machine, force washers, and the forming process
×
![]()
4.2 Data set and data preparation
A resulting quality and process parameters data set of 20 manufactured blanks is utilized. Each process parameter (described in [17]) was recorded for each stroke with a time interval of 0.1 s (sampling rate of 10 Hz). The visualization of the left Force Washer (see Fig. 8) illustrates the data processing techniques an engineer should consider and which parameters from the sensor metadata are needed to perform a specific technique. Expert knowledge is used to isolate only data from relevant strokes (stroke 3), and the data concerning the actual rolling process, enabling the differentiation between set-up times and pullback of the tools (see Fig. 8 (1)). The separation of relevant data is insufficient and necessitates generating relevant features to describe the process parameters’ behavior. Since the data set only includes 20 samples of the quality parameter, a single value is calculated for each parameter and assigned to a set of time series [17].
4.3 Data processing and feature extraction
The extracted features must be computed for each process parameter within an aggregated value and assigned to the corresponding time series. Kirchen et al.[17] limit the feature extraction methods of their approach to the time domain and statistical analysis (see Fig. 9 (Linear Regression)). In contrast, this work uses the relationships captured in the ontology to prepare the data and generate additional features from the frequency and time-frequency domains. The extraction of features from different domains is used to evaluate and compare model performance (see Sect. 4.4). First, a LASSO regression with Manual Feature Extraction is proposed. This model includes features from the time domain, frequency domain, and statistical analysis (see Fig. 9 (LASSO Manual)). Feature extraction from the frequency domain requires prior signal processing to remove unwanted noise and perform signal normalization. Afterward, Fourier Analysis was implemented resulting in 25 new features from the frequency domain. The SSDP ontology extracts parameters needed for the data transformation from the time domain to the frequency domain. For instance, the sampling rate of the process is essential for said transformation (see Fig. 9 (SSDP Relations)). The time domain and statistical analysis features are extracted following the method presented in [17]. Manual feature extraction is time-consuming; therefore, an alternative feature extraction method is implemented with the aid of the SSDP ontology. Therefore, a second proposed model consists of a LASSO regression with Automatic Feature Extraction. The tsfel library is used to extract features from the three domains automatically. This method is applied to the data set without prior signal processing, extracting 3501 features from the time, frequency, and time-frequency domains (see Fig. 9) (LASSO Automatic). Finally, the following subsection evaluates the performance between the two resulting models proposed in this work and the linear regression presented in [17].
Fig. 9
Visualization of the relevant interdependencies between sensor metadata, process characteristics and data processing techniques
×
![]()
4.4 Model evaluation
The homogeneity of the inner rolling zone of a manufactured blank can be determined by measuring the standard deviation of its final thickness. This subsection evaluates the models’ performance using information extracted from the SSDP ontology for feature extraction and compares it to the results from the regression model presented in [17]. The prediction models discussed in this section are (i) regression analysis presented in [17], (ii) LASSO model using manually generated features from the time and frequency domains and (iii) LASSO model using automatically extracted features from all domains, with tsfel python library. Table 2 provides a summary of the relevant parameters to evaluate the performance of the models. It can be seen that the two LASSO models built in this investigation outperform the regression model presented in [17], with both showing higher coefficients of determination, \({R}^{2}\), and lower mean absolute errors, \(MAE\). More concretely, the LASSO model uses automatic feature extraction, and therefore, features from all domains show the best predictive power among the three models. Due to the small data set and the high number of features, the LASSO algorithm is an appropriate method for automatic feature selection and reducing input data used in the regression. With the increasing digitalization of manufacturing processes, the use of sensors is rising, generating a large number of measured variables and complicating the data processing and selection of relevant variables. This is of importance for both process development and monitoring. The approach presented helps not only to manage and process sensor data efficiently but also to decide which variables are essential.
Table 2
Summary of the performance and feature domains found in the different regression models. The results from the linear regression are extracted from [17]
Application | Feature extraction | Time domain | Frequency domain | Time-frequency domain | Statistical analysis | 𝑅2 | 𝑀𝐴𝐸 |
|---|---|---|---|---|---|---|---|
Linear regression | Manual | ✓ | ✓ | 0.9676 | 0.0009 | ||
LASSO | Manual | ✓ | ✓ | ✓ | 0.9919 | 0.0007 | |
LASSO | Automatic | ✓ | ✓ | ✓ | ✓; | 0.9998 | 0.0006 |
5 Conclusion and outlook
Data acquisition, sensor selection, and sensor data can enable engineers to improve and enhance the performance and robustness of manufacturing processes. This work provides a more holistic overview and a comprehensive characterization of relevant interrelationships between data acquisition and sensor data. The SSDP supports engineers in choosing the appropriate data processing methods for efficient data processing in manufacturing systems and tools. To this end, the SSDP ontology, based on the W3C-published standard sensor ontology SSN/SOSA, is developed. Sensor knowledge is modeled from two points of view, i.e., process design and operational data for process monitoring. The SSDP ontology is built by combining expert knowledge, sensor data sheet information, and data aggregation methods used in different manufacturing processes, focusing on forming technologies.
A data set of 20 manufactured blanks from a flexible rolling machine is used to evaluate the benefits of the proposed method during the data processing stage. This work builds two quality prediction models using the SSDP ontology as a support tool and implementing different processing techniques. The difference between the models lies in the processing techniques implemented before the ML implementation. To evaluate which data processing techniques are more appropriate the coefficient of determination and the mean absolute error of the prediction models are compared. As a result, it is demonstrated that the two prediction models containing features from different domains provide a higher goodness of fit (\({R}^{2}\)), and a lower \(MAE\) compared to those of the linear regression model provided in [17]. Figure 8 illustrates the relationship between necessary sensor metadata and the corresponding data processing techniques. This ontology can provide engineers with essential information to implement more robust and reliable ML models.
The following conclusions are derived from the use case: (i) Data processing methods must be supported with domain knowledge. The SSDP ontology captures and formalizes such knowledge, enabling engineers with more comprehensive information for choosing the appropriate data processing methods; (ii) the SSDP ontology further extends sensor properties and considerations and captures underlying dependencies between the different phases, providing a holistic overview. The SDDP ontology is a first step towards a systematic approach to characterizing interrelationships and emphasizing the influences of data acquisition methods and process characteristics on the trustworthiness of data sets for the implementation of AI projects. The approach demonstrates how information captured throughout the different stages of process design and data acquisition, with the aid of domain knowledge, in one information model can benefit different aspects of the data processing stage, e.g., effective feature extraction or reducing time-consuming efforts such as appropriate data segmentation.
So far, the approach proposed in this paper has been used to model one type of sensor from a specific forming process. However, the SSDP ontology was developed by considering different manufacturing processes, i.e., milling or turning and types of sensors, e.g., accelerometers or force sensors, and is built to enable the modeling of different sensors and manufacturing processes with various characteristics. In future work, the transferability of the approach needs to be validated. Furthermore, from a data point of view, a natural next step would consist of evaluating and quantifying the benefits of using the knowledge captured in the SDDP ontology in the data processing stage of different manufacturing processes.
Future updates to the SSDP ontology can be driven by various factors, including integrating new sensor types, new sensor data processing methods, or changes in system configurations. Depending on the type of change, which typically includes “add,” “delete,” and “modify” operations, updates to the SSDP ontology in the owl format can be implemented using SPARQL. As a query language for ontologies, SPARQL offers functionalities like INSERT, DELETE, and CONSTRUCT for ontology updates. These SPARQL queries can be executed using tools like OWLready2 or web-based platforms like Neo4j. For instance, when adding a radar sensor to measure the position of moving workpieces in forming processes, concepts, and attributes in the “Sensor Component,” “Sensor Type,” “Sensor Property,” and “Sensor Condition” modules should be updated based on the sensor datasheet. Using an INSERT statement in SPARQL, “Measurement Principle” attributes can be, e.g., extended to include the radar sensor’s measuring principle of “radio wave.” In practical applications, predefined query templates can also be created for reuse, aiding engineers lacking ontology expertise during updates. The SSDP ontology should differentiate between updates in the core classes and relations (TBox) and project-specific application levels (ABox). For instance, when integrating a new radar type sensor, updating the “Data Processing” module can involve adding TBox-level signal processing steps for handling radar data in the point cloud format. Additionally, project-specific sensor information at the ABox level, such as defining the sampling frequency in the “Sampling” module based on system production speed in the “PPR” module, should be considered. A new version of the SSDP ontology can be created when there are updates to the TBox-level information or significant changes in project-specific data. These updated procedures can be standardized within a company by integrating them into enterprise workflow management systems. To track changes between ontology versions, each element in the ontology can be assigned a Uniform Resource Identifier (URI) with additional version information. The modular structure of the SSDP further restricts modification access to user groups, such as data engineers, to the respective “Data Analysis” sections of the ontology. In addition, various ontology versioning methods and tools can be utilized to check inconsistency and ensure long-term reusability while managing updates to the SSDP ontology.
Declarations
Competing interests
The authors declare no competing interests.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.