The Future Is Organic: A Deep Dive into Techniques and Applications for Real-Time Condition Monitoring in SASO Systems—A Systematic Review

Abstract

Condition Monitoring (CM) is a key component of Self-Adaptive and Self-Organizing (SASO) systems. By analyzing sensor data, CM enables systems to react to dynamic conditions, supporting the core principles of Organic Computing (OC): robustness, adaptability, and autonomy. This survey presents a structured overview of CM techniques, application areas, and input data. It also assesses the extent to which current approaches support self-* properties, real-time operation, and predictive functionality. Out of 284 retrieved publications, 110 were selected for detailed analysis. About 38.71% focus on manufacturing, 65.45% on system-level monitoring, and 6.36% on static structures. Most approaches (69.09%) use Machine Learning (ML), while only 18.42% apply Deep Learning (DL). Predictive techniques are used in 16.63% of the studies, with 38.89% combining prediction and anomaly detection. Although 58.18% implement some self-* features, only 42.19% present explicitly self-adaptive or self-organizing methods. A mere 6.25% incorporate feedback mechanisms. No study fully combines self-adaptation and self-organization. Only 5.45% report processing times; however, 1000 Hz can be considered a reasonable threshold for high-frequency, real-time CM. These results highlight a significant research gap and the need for integrated SASO capabilities in future CM systems—especially in real-time, autonomous contexts.

1. Introduction

Self-adaptive and self-organizing systems (SASO) refer to systems that perceive their environment and its changes through receptive components, evaluate the situation, plan an appropriate response, and execute it autonomously [1]. Two of the most significant initiatives for realizing such systems are Organic Computing (OC) [2] and Autonomic Computing (AC) [3], which propose concrete techniques, control mechanisms, and concepts based on Machine Learning (ML) for autonomous decision-making [4]. In particular, OC combines self-adaptation and self-organization to enhance system robustness, flexibility, and resilience [5].

Beyond these, there exists a wide range of additional self-* terms that refer to system capabilities enabling runtime adaptation, optimization, and resilience without human intervention. While the list of possible self-* properties is theoretically endless, the core set typically emphasized in research includes self-adaptive (adjusting behavior in response to changes), self-organizing (restructuring internal organization), self-healing (recovering from faults), self-protecting (defending against threats), and self-explaining (providing understandable reasons for actions) [6,7,8,9].

The most widely recognized reference model for SASO, particularly for self-adaptive systems, is the Monitor-Analyze-Plan-Execute over a shared Knowledge (MAPE-K) feedback loop [10]. This concept structures such systems into distinct components: monitoring the environment using sensors, analyzing the collected data, planning an appropriate response to environmental changes, and executing the response via actuators [11]. Additionally, the knowledge component helps refine future analysis and decision-making processes. A schematic representation of the MAPE-K loop is shown in Figure 1. It can be emphasized that MAPE-K is generally applicable to the approach of SASO systems and can be implemented in various patterns [12]. Specifically, it can be directly mapped to the Observer/Controller pattern in OC [13].

Figure 1. Schematic illustration of the MAPE-K loop, comprising the cognitive layer (Monitoring, Analysis, Planning, Execution and the overarching Knowledge), the system–environment interaction layer (Sensors and Effectors), and the monitored environment.

While numerous approaches and diverse use cases exist for implementing SASO, real-world systems remain scarce. Given that planning and execution components have already been extensively tested using deep reinforcement learning techniques [14], one of the key challenges appears to be the effective description of system states.

Achieving self-adaptation and self-organization in a system relies heavily on the ability to monitor its states and parameters, as well as those of its environment. Effective monitoring enables the detection, analysis, and response to changes, ensuring that the system remains efficient and adaptable.

Condition Monitoring (CM) refers to the process of continuously assessing the state of a system or process through one or more sensor elements, often enabling automated responses based on the collected data [15]. CM can range from detecting anomalies or faults to predicting future failures by analyzing the system’s current condition. In industrial machinery, for instance, it can be used to estimate remaining useful life and schedule maintenance proactively. A practical example is its application in the Industrial Internet of Things (IIoT), where vibration sensors are used to monitor and predict bearing wear in production equipment, helping to prevent unexpected breakdowns and reduce downtime [16]. As a result, CM systems can incorporate both anomaly detection and predictive modules. Predictive maintenance, in particular, stands out as one of the most innovative approaches to proactively preventing failures [17].

Beyond anomaly detection, CM can encompass the analysis, recognition, and classification of anomalies, as well as the diagnosis of their root causes. Therefore, CM can be regarded as a critical component of OC, contributing to the system’s ability to self-adapt and self-organize.

Certain processes and systems require real-time state monitoring due to their rapid operation and the need to adhere to strict safety constraints. Real-time monitoring allows the system to adapt its behavior or structure within a timeframe appropriate to the specific process or system. This aligns closely with the requirements of OC systems, as described by Tomforde et al., which emphasize the need for runtime modifications to behavior and/or structure.

A fundamental aspect of CM systems is the collection of data about the current state of the system or process using sensors. These sensors can be categorized into heterogeneous or homogeneous types. Heterogeneous sensors combine multiple types, such as vibration, temperature, force, and acoustic sensors, whereas homogeneous sensors rely on a single type of sensor.

This article aims to provide researchers with a comprehensive overview of real-time CM techniques that exhibit SASO system properties, along with their associated application domains. It also facilitates the comparison of these methods based on input data and predictive capabilities, while identifying research gaps and outlining potential directions for future investigation.

In particular, this article contributes to the field through an extensive review of the literature by:

• Overview of real-time CM techniques with SASO properties.
• Taxonomy and categorization of existing studies based on the implemented techniques, application domains, input data, and predictive approaches for CM in SASO systems.
• Analysis of self-* properties and the extent to which SASO system requirements are fulfilled.
• Exploration of high-frequency monitoring and CM system processing time
• Deriving research gaps and proposing a structured research agenda for future investigation

The review is structured as follows: Section 2 outlines the related works and distinguishes them from the current study. Section 3 provides a detailed description of the implemented methodology, including the definition of research questions, search strategy, selection criteria, and selection process. Section 4 presents the results derived from the included papers. In Section 5, an analytical discussion of these findings is conducted. Building on this discussion, Section 6 introduces the research agenda, which summarizes the identified research gaps and proposes potential future research directions. Finally, Section 7 concludes the survey with a summary of the key insights.

2. Related Works

In recent years, several review and survey publications have covered CM; however, they have not specifically focused on self-adaptive (SA) or self-organizing (SO) systems, and particularly not on SASO systems. These studies often focus on specific application areas, techniques, or a combination of both. Many authors concentrate their research on specific subdomains of CM, particularly anomaly detection. For instance, Hodge et al. presented approaches for outlier detection in 2004, without considering their applicability to SA or SO systems and covering only research up to that time [18].

Other reviews have primarily focused on fault diagnosis. For example, Sun et al. in 2012 explored fault diagnosis in oil-immersed power transformers by analyzing dissolved gases [19]. Their work provided an overview of ML techniques—including fuzzy logic, neural networks, and evolutionary optimization approaches—but did not address SASO systems. Moreover, their study not only emphasized a specific application but also limited its scope to ML-based methods.

Similarly, Yan et al. in 2014 investigated fault diagnosis in rotary machines, concentrating on wavelet techniques and showcasing various applications of these methods [20]. However, they only briefly mentioned an SA approach for diagnosis and covered research only up to 2014.

Fault prediction and maintenance interval estimation are also critical aspects of CM. Unlike the previous studies, El-Thalji et al. in 2015 examined predictive health monitoring methods and fault modeling specifically for rolling element bearings, yet they mentioned only one SO method [21]. In 2016, Rai et al. also focused on rolling element bearings, investigating fault diagnosis and prognosis using signal processing techniques [22]. However, their study provided only a few SA preprocessing and feature extraction techniques for CM in SASO systems.

In contrast, Zhou et al. in 2018 addressed CM methods with a specific focus on tool CM in the milling process [23]. Their review analyzed the literature in terms of CM models, sensor types, and feature extraction methods, but mentioned only one SA technique for preprocessing, without covering CM comprehensively. Similarly, Wang et al. in 2019 concentrated on CM techniques, restricting their scope to methods that utilize vibration sensor data for turbine planetary gearbox applications [24]. Their study emphasized fundamental analysis, signal processing, feature extraction, and fault detection, yet it addressed only one SA feature extraction method.

A more recent publication by Kumar et al. in 2020 explored gear defect diagnosis and prognosis methods for CM by categorizing different types of gear defects, although they mentioned only one SO method and one SA preprocessing technique [25]. Lastly, Chaupal et al. in 2023 examined damage identification methods for laminated composite structures using vibration-based data [26]. Their review highlighted optimization techniques, signal processing methods, and ML algorithms, but it referenced only one SA technique for CM.

The discussed review papers primarily focus on specific methods, data, application domains, or a combination of these aspects. Table 1 provides a comparative overview and differentiation of various related works in the field of CM, with respect to their limitations, techniques analyzed, time periods covered, and how they differ from the present study. However, many fail to comprehensively address all areas of CM. Moreover, there is either a lack of or an insufficient consideration of CM implementation in SASO systems, with no explicit focus on SASO systems. To bridge this gap, this work provides a comprehensive review of established and recent methods, applications, and various input data used in the field of CM for SASO systems. Furthermore, all key aspects of CM are thoroughly analyzed, including an investigation of the self-* properties implemented in the CM methods. In addition, the processing time of the applied approaches is carefully examined.

Table 1. Comparison of related surveys on CM with respect to their limitations, techniques analyzed, time periods covered, and distinction from this work.

Study	Years Covered	Focus Area	Techniques Analyzed	Limitations
[18]	<2004	Outlier detection	Statistical, ML	No SA/SO or SASO; early scope
[19]	<2012	Transformer fault diagnosis	ML: fuzzy logic, NNs, optimization	Application-specific; no SASO focus
[20]	<2014	Rotary machine diagnosis	Wavelet transforms	One SA approach; no SO or SASO
[21]	<2015	Bearing health prediction	Signal proc., ML, and dynamic models	One SO method; narrow scope
[22]	<2016	Bearing diagnosis and prognosis	Signal proc.	Few SA methods; no system-level view
[23]	2000–2016	Tool CM in milling	ML models, feature extraction techniques, signal-based sensors	One SA technique; limited coverage
[24]	2000–2019	Turbine gearbox CM	Signal processing, ML algorithms, feature extraction methods	Application-specific; no SO, one SA method
[25]	<2020	Gear defect diagnosis	Signal proc. ML, and feature extraction	One SA + one SO method
[26]	2010–2022	Composite structures	ML, optimization, signal proc.	One SA method; Only specific application
This work	<2024	CM across domains with SASO focus	ML, DL, non-ML; self-* properties*, real-time, input data	Covers SASO, real-time, self-* properties, data types/input data, and cross-domain comparison

3. Methodology

The primary aim of this survey is to provide a comprehensive review of the literature, focusing on techniques for CM in SASO systems, with particular emphasis on high-frequency systems. Furthermore, it distinguishes between non-ML methods and ML approaches, including Deep Learning (DL) techniques. The methodology of this study follows the approach outlined by Piliuk et al. for the review of quantitative-focused contributions and is structured as follows [27].

• Defined research questions to guide the direction of the review.
• Designed a search strategy, including the specific search terms and the contribution sources.
• Created selection criteria to identify studies to exclude from the review.
• Established a systematic procedure to apply the selection criteria consistently.

These steps are designed to ensure an unbiased and comprehensive study. Each step is explained in detail within this section.

3.1. Research Questions and Search Strategy

In this section, the research questions and the search strategy are described. The research questions serve as a foundation for developing an appropriate search strategy and selection process for contributions, ensuring a focused and systematic approach to addressing the research questions. The following research questions have been identified, and they will be answered through the analysis of the selected contributions.

• In which application domains is CM employed in SASO systems and how are the applied methods influenced by specific applications?
• To what extent do the presented CM approaches implement self-* properties, and how well do they meet the requirements of a SASO system?
• What technologies and methods are utilized for CM in SASO systems, and what is the role of ML and DL in this context?
• What types of sensor data and features are used for CM, and how frequently is the data processed?
• What challenges and open research questions exist in integrating CM into SASO systems, particularly with the use of DL techniques?

The basis of the search strategy is a clearly defined search term, which was used to identify relevant sources. To determine the most suitable and relevant databases, an initial exploration was conducted using Google Scholar [28]. This process identified four key databases recognized for their significance in the fields of ML and engineering sciences. For the specific source search, the selected databases included ACM Digital Library, IEEE Xplore, Springer Nature Link, and ScienceDirect [29,30,31,32].

The ACM Digital Library and IEEE Xplore were utilized for gathering primary studies due to their extensive collection of publications in computer science and engineering, with a specific focus on ML, real-time systems, and adaptive technologies. In contrast, Springer Nature Link and ScienceDirect were included for their broad range of subject-specific contributions relevant to SASO systems and CM. By selecting two method-specific and two general sources, we aim to provide a comprehensive overview while ensuring a balanced representation of both method-specific and subject-specific contributions.

To achieve a transparent selection of contributions that aligns closely with the main focus of the study while simultaneously providing a comprehensive literature overview, a precise search term was established. This search term is defined as follows:

(“condition monitoring”) AND (“self-adaptive” OR “self-organization” OR “self-aware” OR “self-correcting”) AND (“real-time”) AND NOT (“framework”)

The first part of the search term narrows the scope to CM, while the second part focuses on the self-* properties and characteristics of SASO systems. Together, these address the autonomous monitoring of system states. The third part specifically targets real-time systems, as the primary focus of the study is on high-frequency systems. Combined with the first two parts, this ensures alignment with the core objectives of the study. Lastly, the exclusion of frameworks in the final section of the search term ensures that only concrete techniques are considered, as frameworks generally describe a structural outline rather than a specific model or methodology. The contributions were identified and selected using the search strategy conducted on 15 November 2024. This systematic review adheres to the PRISMA guidelines [33], with the corresponding PRISMA checklists and flow diagram included in Appendix D.

3.2. Selection Criteria and Process

Initially, contributions categorized as reviews, abstracts, glossaries, datasets, keywords and indices, or retracted papers were excluded (EC 1). Furthermore, book chapters that do not constitute standalone publications and primarily provide summaries or overviews of current techniques and topics related to CM were excluded. Additionally, duplicate studies and older versions of contributions were manually removed in the next step (EC 2). Subsequently, studies that did not explicitly address CM or related concepts, such as predictive maintenance, were excluded. This included all contributions that did not describe a concrete CM process for system operations (EC 3), as well as those focusing on monitoring physical states or conditions. Furthermore, contributions lacking a detailed description of the methodology—i.e., studies that did not clearly explain how CM was conducted—were excluded (EC 4). This includes contributions that failed to specify concrete techniques. Finally, inaccessible paper were also removed (EC 5).

The filtering process was conducted step-by-step as outlined in Table 2, ensuring that the criteria were not weighed against each other. No formal tools or automation software were employed to assess the risk of bias in the included studies. Additionally, no risk of bias assessment was performed for individual studies. Furthermore, neither an assessment of the certainty or confidence in the evidence nor sensitivity analyses were carried out, as the study relied on a qualitative synthesis. The selection process was conducted independently by both authors. Applying this process resulted in the inclusion of 110 papers. The total number of papers identified through the selection process and search strategy can be found in the Appendix A. Additionally, Table A9 provides a list of excluded papers, while Table 3 presents the distribution of included and excluded papers across the various databases used. Analyzing the number of publications over two-year intervals reveals a growing interest in intelligent self-regulating CM systems, as illustrated in Figure 2. This study was not registered, and no formal protocol was developed prior to its initiation. Consequently, no amendments to a protocol or registration were necessary.

Table 2. Overview of the applied exclusion criteria.

Criterion	Description	Studies Excluded
EC1	The contribution is a review, dataset, specific book chapters, etc.	79
EC2	Older version or duplicates of paper	2
EC3	Paper that does not address CM	76
EC4	Contributions without specifying a technique for CM	10
EC5	Papers that were inaccessible or unavailable	7

Table 3. Distribution of included and excluded papers across the various databases.

Digital Library	Contributions (Included)
ScienceDirect	186 (75)
Springer Nature Link	95 (35)
ACM Digital Library	2 (0)
IEEE Xplore	1 (0)
Total	284 (110)

Figure 2. Representation of the selected and queried contributions per time interval, where each time interval spans a period of two years.

4. Results

This section provides an explanation of the taxonomy used in this study. It was applied to the 110 selected studies that remained after the filtering process.

4.1. Taxonomy

The selected papers were categorized based on several criteria to ensure a comprehensive analysis and to effectively address the research questions outlined in the methodology of this study. The primary classification criteria include the application domain and the techniques employed, which are further divided into non-ML and ML methods. Within the ML category, an additional distinction is made between traditional ML methods and DL approaches. Furthermore, the papers were analyzed based on the type of input data used and the type of CM implemented. The contributions were further examined with respect to the self-* properties of the applied CM methods. Moreover, the analysis includes the presence of a control component for a feedback loop, the number of sensors used, the feature size, the nature of the data (e.g., real or simulated), and the processing time required by the CM approaches.The following sections provide a detailed overview of this classification.

4.1.1. Application Domain

The selected contributions are categorized based on their practical application described in the study or the specific application addressed by the method. The categorization is as follows:

• Manufacturing/Process: Papers that focus on CM of manufacturing processes or procedures, such as cutting, drilling, or milling operations, as well as the examination of the produced object, are classified in this category.
• Structure: Studies categorized here address CM of static objects, including buildings and other structures.
• System: This category includes the monitoring of system parameters, such as machines, circuits, or autonomous vehicles.

4.1.2. Technique

In this category, the classification is based on the techniques utilized.

• Machine Learning: This category includes papers that utilize ML techniques for CM, with a further distinction made between DL and traditional ML.
• Non-Machine Learning: Contributions in this category rely on traditional techniques such as signal processing or statistical methods for CM.

4.1.3. Input Data

This section categorizes the selected papers based on the types of input data utilized in the studies.

4.1.4. Type of Condition Monitoring

This category differentiates between the types of CM: the prediction of future states using predictive approaches and the detection of anomalies based on current states. Additionally, papers that address both aspects are classified in a separate category.

• Predictive Approach: This category includes papers that focus on predictive approaches to monitoring states, such as forecasting the lifespan of a machine or estimating maintenance intervals.
• Anomaly Detection: Contributions in this category examine the current state, specifically identifying existing faults or changes within the system or process.
• Combination of Both Approaches: This category encompasses papers that integrate both predictive methods and current state monitoring techniques in their approaches.

4.1.5. Self-* Properties

For the realization of SASO systems, the implementation of the CM component must incorporate self-* properties, particularly self-adaptivity and self-organization. This section examines whether the CM techniques implemented in the analyzed papers exhibit these properties and, accordingly, whether they are suitable for deployment in SASO systems. Furthermore, it is investigated whether existing publications have partially realized the MAPE-K feedback loop through CM.

4.2. Application Domain

CM is applied across various applications. Some applications focus on entire processes, such as manufacturing workflows, while others target the CM of individual machine elements, such as bearings or specific processes. Additionally, Table 4, Table 5 and Table 6 provide a classification of the papers by three CM application categories—Manufacturing/Process, System, and Structures—along with their specific application areas and the number of contributions in each category.

Table 4. Number of contributions in various application areas from the Manufacturing/Process category.

Application	Contributions
Additive manufacturing	1
Milling	12
Cutting	5
Drilling	4
FMS	1
Industry 4.0	1
Micro-fluidic chip	1
Pickling	1
Machine surface	1
Tapping	1
Welding	2
Semiconductor manufacturing	1
Total	31

Table 5. Summary of the distribution of contributions across the application areas of the System category.

Application	Contributions
Lithium battery	1
Fluidized bed boilers	1
Circuit	2
Network intrusion	1
Energy harvesting	1
Internet of things	1
Oil refining	1
Spinning	1
PEMFC	1
Satellite	1
Power plant	2
Bearing	35
Gear	6
Shaft	2
Aircraft engine	1
Industrial environment in general	1
Wear particle	1
Inverter-fed machine	1
Induction machine	1
Autonomous transfer vehicles (ATV)	1
Autonomous underwater vehicle (AUV)	1
Pressure system	6
Complete machinery system	1
Oil plant	1
Nuclear plant	1
Total	72

Table 6. Overview of the distribution of contributions across the application areas of the Structure category.

Application	Contributions
Nonparametric damage detection (civil structure)	1
Damage detection (steel structure)	1
Crack detection (steel beams)	1
Wheel polygonization detection (heavy haul locomotive)	1
Leak detection (long transportation pipeline)	1
Road pavement diagnostics	1
Foreign object debris detection (rocket tank final assembly process)	1
Total	7

4.2.1. Manufacturing and Process Monitoring

A significant proportion of the analyzed studies, approximately 38.71%, addressing CM in manufacturing processes focus on the milling process, particularly in CNC machines [34]. Tool wear monitoring and the detection of faults and anomalies occurring during the process play a crucial role in these applications. Similar processes, such as drilling and cutting, are also frequently studied in CM research, accounting for approximately 16.13% and 12.9% of the contributions, respectively.

Beyond tool monitoring during manufacturing, the literature also explores the monitoring of materials or products throughout the entire production process [35]. The primary focus remains on the detection of defects and irregularities. Some studies further address the classification and analysis of such faults [36]. Additionally, a number of publications examine the development of CM diagnosis systems [37]. A detailed distribution of manufacturing applications is presented in Table 4.

4.2.2. System-Level Monitoring

Many publications shift their focus from monitoring process flows and manufacturing outcomes to analyzing individual system components, standalone systems, or systems composed of multiple subsystems. The distribution of these applications is provided in Table 5. A key area of interest in CM research is the assessment of critical machine components, particularly bearings, which account for 48.61% of the studied cases [38,39]. A significant portion of the research is dedicated to CM of rolling element bearings, commonly found in rotating machinery [40], including motors such as those used in Autonomous Underwater Vehicles (AUV) [41], aero engines [42], and generators like those in whirlpool turbine generators [43]. Other bearing types include planetary bearings used in wind turbines [44] and ball bearings found in induction motors [45].

Beyond bearings, CM studies also investigate other components, such as gears (approximately 8.33%) in wind turbine gearboxes [46], and shafts (about 2.78%), such as rotor shafts [47].

Standalone system monitoring encompasses diverse applications, ranging from components in pressure systems, such as reciprocating compressors in refrigerators [48], to entire machines, such as induction machines [49], aircraft engines [50], and turbomachinery in oil plants [51]. Applications concerning pressure systems constitute 8.33% of all system-related studies, while oil plants and induction machines account for only 1.39% each. Overall, 81.94% of system-level applications focus on machinery, with approximately 72.88% concentrating on individual key components such as bearings, gears, or shafts. Additionally, 8.33% of system-related studies examine the monitoring of multiple interconnected systems. A detailed breakdown of these applications is presented in Table 5.

4.2.3. Full-System Monitoring

Comprehensive CM approaches also address entire systems composed of multiple interconnected subsystems. Examples include CM of entire wind farms [52], AUVs [53], electrical circuits [54], and nuclear facilities [55].

As with manufacturing process monitoring, the primary objective in full-system monitoring is the detection of faults and anomalies, as well as system-wide analysis and diagnostics. Some studies focus on predicting the remaining useful life of a system or the occurrence of potential failures [56]. Additionally, research has been conducted on monitoring the deterioration of system conditions and degradation processes [57].

4.2.4. Structural Monitoring

Another important focus area is the monitoring of static structures. This includes civil and steel structures [58,59], as well as tanks, such as rocket fuel tanks [60], and pipelines [61]. CM is also applied to smaller structural components, such as steel beams [62] and heavy haul locomotive components, particularly wheels [63]. The primary objective in structural monitoring is damage detection.

Overall, only 6.36% of the analyzed studies focus on CM of structures and static objects. In contrast, 28.18% of contributions address manufacturing processes, while the majority (approximately 65.45%) focus on system-level CM. A detailed distribution of CM applications for structural and static objects is presented in Table 6.

4.3. Techniques

To conduct CM, various methods are employed, including both ML and non-ML approaches. ML techniques can be further divided into DL models and conventional ML methods.

4.3.1. Machine Learning

A variety of models are used for system CM. The distribution of employed models is illustrated in Table 7. As depicted in Figure 3, there has been a steady increase in the use of ML techniques over time. A particularly notable rise in DL applications begins around the 2013–2014 period. Before that, only conventional ML methods were used. This rapid growth in DL adoption reflects a clear shift toward more complex models and a gradual decline in the reliance on simpler techniques. Among the analyzed studies, 56.36% apply conventional ML models (excluding DL techniques). Neural networks (NN) are the most frequently implemented, appearing in 23 studies (approximately 37.1% of ML techniques). These include standard Artificial Neural Networks (ANN) and Multilayer Perceptrons (MLP) [64], Fuzzy-Neural Networks (FNN) [53], Self-Adaptive Growing Neural Networks (SAGNN) [65], and Self-Organizing Mapping Neural Networks (SOM) [66]. Other commonly used ML models include clustering techniques, such as mean shift clustering [67], Principal Component Analysis (PCA) [68], and k-means-like Adaptive Sequential K-means (ASK) [69]. Classical approaches like Decision Trees (DT) [70], regression models [71], and Support Vector Machines (SVM) combined with optimization algorithms, such as genetic algorithms (GA) [72] and Grey Wolf Optimizations (GWO) [73], are also employed. Additionally, CM implementation utilizes Gaussian mixture models (GMM) [61] and Markov models [74].

Table 7. List of ML models and the number of articles utilizing these techniques (excluding DL methods).

Method	Contributions
Comparison of different methods	5
LVQ	1
ANN	16
Fuzzy-NN	4
SAGNN	2
SNN	1
SVM	6
Decision tree (DT)	3
Regression	3
Clustering	9
PCA	2
PLS	1
SOM	2
GMM	2
SR	1
Markov model	4
Total	62

Figure 3. Annual distribution of machine learning methods used for condition monitoring, with a particular focus on DL. Background bars represent the total number of studies applying machine learning techniques in each two-year interval, while the foreground line illustrates the subset specifically employing DL approaches.

To enhance CM accuracy and quality, feature extraction, optimization, and preprocessing techniques are frequently applied. Feature extraction and preprocessing often involve decomposition algorithms such as Empirical Mode Decomposition (EMD) [75], Variational Mode Decomposition (VMD) [76], Singular Value Decomposition (SVD) [40], Intrinsic Time-scale Decomposition (ITD) [77], Local Characteristic-scale Decomposition (LCD) [78], and Local Mean Decomposition (LMD) [58], along with their extensions like Ensemble Empirical Mode Decomposition (EEMD) [79]. Wavelet transformations, such as Discrete Wavelet Transform (DWT) [80], PCA [78], and Linear Discriminant Analysis (LDA) [81], are also common. Signal processing techniques include Fast Fourier Transforms (FFT) [62] and Adaptive Wiener Filters like Energy Variation-Driven Adaptive Wiener Filter (EDAWF) [61]. Optimization algorithms include Fruit Fly of Algorithm (FOA) [46], Particle Swarm Optimization (PSO) [54], and GA [82].

4.3.2. Deep Learning

DL models utilize deeper layers to facilitate advanced pattern and feature learning. The distribution of DL models in CM studies is detailed in Table 8. Convolutional Neural Network (CNN)-based approaches, often combined with preprocessing and optimization techniques similar to those discussed in Section 4.3.1, are prevalent. CNN models are frequently integrated with other ML or DL techniques, such as SVMs [83], short-time Fourier transform [84], particle filters (PF) [42], and decomposition methods like EEMD [85]. CNN implementations utilize both one-dimensional and two-dimensional approaches [86]. Long Short-Term Memory (LSTM) [87] are also frequently used, often in combination with preprocessing techniques [88] and CNNs [63]. Additional DL models include Gated Recurrent Units (GRU) [57] and Safe Semi-Supervised Support Vector Machine (S4VMs), often paired with autoencoders such as Stacked Denoising Autoencoder (SDAE) [89]. CNNs constitute 57.14% of DL models, followed by LSTMs at 21.43%.

Table 8. Distribution of contributions across various DL techniques.

Method	Contributions
CNN	8
LSTM	3
Comparison of different methods	1
SVM-AE	1
GRU	1
Total	14

4.3.3. Non-Machine Learning

CM is also conducted using non-ML approaches, which can be categorized into mathematical models, signal processing techniques, and hybrid adaptive optimization methods. The distribution of these methods is presented in Table 9. Signal processing techniques account for the largest portion (67.6%), followed by mathematical models (11.8%) and hybrid adaptive optimization techniques (20.6%). Mathematical models include approaches such as boundary models and the 3$\sigma$ threshold principle [90]. Signal processing techniques encompass various filtering algorithms, such as second-generation wavelet transforms [91], frequency-domain transformations such as FFT [49] and Continuous Wavelet Transform (CWT) [59], as well as decomposition methods including EMD [34], LMD [92], and VMD [93].

Table 9. Overview and count of papers classified as non-ML methods, organized into three sections.

Method	Contributions
Statistical/mathematical methods	4
Signal processing	23
Hybrid adaptive and optimization techniques	7
Total	34

Hybrid adaptive optimization methods integrate optimization algorithms such as Nelder-Mead Salp Swarm Algorithm (SSA-NM) [94] and Clonal Selection Algorithm Differential Evolution Algorithm (CSA-DEA) [47], along with adaptive filters combined with signal processing techniques [95]. Additionally, advanced approaches like Synchronized Switching Harvesting on an Inductor with Self-Adaptive Mechanical Switches (SSHI-SAMS), which utilizes self-adaptive mechanical switches, have been employed [96].

Preprocessing techniques vary depending on the applied method and include strategies such as Bi-Spectrum based EMD (BSEMD) [97], Spline-Based LMD (SBLMD) [98], and FFT-based feature extraction in LMD-FFT [99].

4.4. Input Data

Despite variations in techniques, both ML and non-ML approaches utilize a limited range of sensor types, leading to overlaps in input data types. Vibration data from accelerometers [100], acoustic data from microphones [101], and force data [102] are commonly used. Additional input data include temperature [103], current [35], and pressure [104]. Sampling frequencies range from 0.003 kHz to 6250 kHz, with vibration data typically between 0.1 kHz and 100 kHz. Almost all reviewed papers validate their proposed methods using real-world data and case studies, with 18.18 % additionally incorporating simulated data. Only one of the reviewed papers exclusively used simulation data, specifically from a 320 W power plant [105], accounting for 0.91% of all papers. For the generation of training, testing, and validation datasets, between one and 63 different sensors (or process parameters) were monitored [106]. Figure 4 illustrates the heterogeneity of sensors used in the reviewed publications, grouped in two-year intervals. Over the past ten years, the number of studies relying on a single type of sensor has remained relatively constant. However, there has been a noticeable increase in publications combining two different types of sensors, with an even more pronounced rise in the use of more than two sensor types. This highlights a clear trend toward multi-sensor data acquisition. From these data, between one and 2000 distinct features were extracted—or up to 2000 feature values used as input—primarily through spectral and/or statistical feature extraction techniques [63]. The data characteristics of the reviewed papers can be found in Table 10. Additionally, a comprehensive overview of all selected publications, including the number of sensors used, the type of data, and the feature size, is provided in Table A6, Table A7 and Table A8.

Figure 4. Distribution of sensor diversity in the analyzed publications over time. The stacked bar chart illustrates the number of studies in each two-year interval using only one sensor type (light green), exactly two sensor types (medium green), and more than two sensor types (dark green). The visualization highlights the trend toward increasing sensor heterogeneity in recent years.

Table 10. Overview of data characteristics.

Characteristic	Value
Sampling frequencies [kHz]	0.003 to 6250
Vibration data [kHz]	0.1 to 100
Number of features/input feature value size	1 to 2000
Number of sensors/parameters	1 to 63
Share of contributors providing additional simulation [%]	18.18

4.5. Types of Condition Monitoring

CM techniques extend beyond real-time fault detection and anomaly diagnosis to include predictive maintenance strategies. Some studies implement algorithms to assess the current health status and generate predictions using models such as Support Vector Regression (SVR) [107] and Markov models [74]. Other approaches combine real-time fault detection with predictive maintenance, often utilizing CNNs [85,108].

4.6. Self-* Properties

In the domain of process and manufacturing applications, only 20 out of 31 papers (64.52%) demonstrate self-* properties. Among these, two papers do not implement true self-adaptivity, but rather introduce only a self-adaptive learning rate algorithm [64] and a self-adaptive scaling factor [73]. Furthermore, six of the remaining 18 papers employ solely self-adaptive feature extraction methods, such as EMD [100], EEMD [79], a self-adaptive Kalman wave filter [37], and Second Generation Wavelet (SGW) [91]. As a result, only 38.71% of the papers in this category exhibit concrete self-* capabilities. Six papers implement self-adaptive CM methods, for instance, using an MLP [109], a self-adaptive learning algorithm [90], self-adaptive models [106], a model combining self-learning and self-adaptation [88], or even self-adaptive control [108]. Additionally, six papers implement CM techniques based on self-organization, such as Adaptive Resonance Theory 2 (ART2) used as a self-learning and self-organizing algorithm [110] and SOM [111].

In applications addressing system-level monitoring, only 41 out of 72 papers (56.94%) utilize techniques with self-* properties. Among them, five publications implement only self-adaptive techniques for noise removal, such as the self-adaptive noise cancellation algorithm (SANC) [39]. Nineteen papers apply self-* properties solely for feature extraction methods, including EMD [40], EEMD [85], LCD [78], LMD [99], Local Oscillatory-Characteristic Decomposition (LOD) [112], SVD [77], VMD [113], and Statistic Filter (SF) [43], as well as two filtering algorithms based on Least Mean Square (LMS) [56] and a self-adapting sample representation method Dependent Feature Vector (DFV) [114]. Only ten of the 72 papers specifically address the implementation of self-adaptive methods for CM. One study utilizes a self-adaptive Differential Evolution Algorithm (DEA) for fault identification [47], while others propose adaptive thresholds [92] or self-adaptive baselines [57] for anomaly detection. Additional approaches involve self-adaptive CM models, such as SAGNNs [65] and Online Sequential Extreme Learning Machine (OS-ELM) [115]. Other techniques include self-adaptive tuning of model parameters [116] and the self-adaptive adjustment of clustering algorithms to input data [69]. Three papers focus on parameter correction and control based on updated sensor data: two publications implement self-adaptive control algorithms [96,117], and one proposes a self-correcting algorithm [118]. In addition to self-adaptivity, one paper addresses self-awareness and self-regulation for online fault monitoring [119], while five publications implement self-organizing approaches, including SOM [120], Self-Organising Feature Maps (SOFM) [48], and self-organizing models such as ART for classification and fault identification [55].

In the field of structural monitoring, three out of seven papers (42.86%) employ CM approaches with self-* properties. However, one paper only utilizes a self-adaptive Adam optimizer [63], and another applies a self-adaptive feature extraction method based on LMD [58]. Moreover, one contribution implements a self-correcting neural network for road pavement diagnostics [121].

Table A6, Table A7 and Table A8 provide an overview of the self-* properties, the presence of additional control components, processing time, the number of sensors used, and the number of extracted features or input features categorized by application type. Furthermore, Table 11 provides an overview of the number of self-* properties applied in the reviewed publications for implementing CM approaches across the various application categories.

Table 11. Overview of the self-* properties used in the different application areas for CM.

Application	Self-* Properties	Category	Approach	Contributions
Manufacturing and process monitoring	SA			14
		Learning rate	Learning rate algorithm	2
		Scaling	Scaling factor	1
		Feature extraction	EMD	3
			EEMD	1
			Kalman wave filter	1
			SGW	1
	(+SL)	Concrete techniques	CNN-based	1
	(+SL)		Hybrid approach	1
			PLS forecast models	1
			MLP-based	1
		Learning	Boundary model-based	1
	SO			6
			ART2-based	2
			SOM-based	2
			SOM-NN-based	1
			NN-based	1
System-level monitoring	SA			34
		Noise canceling	SANC	5
		Feature extraction	EMD	5
			EEMD	3
			LMD	1
			LCD	1
			SF	1
			LOD	1
			VMD	2
			SVD	1
			Envelope extraction	1
		Filtering algorithms	LMS	2
			DFV	1
		Threshold	Baseline-based	1
			Threshold-based	1
		Control	ML-based	1
		Concrete techniques	DEA	1
			SAGNN	2
			OS-ELM	1
			SSHI-SAMS	1
			QTA-based	1
			ASK-based	1
	SO			4
			ART-based	1
			SOM-based	1
			SOFM-based	1
			FNN-based	1
	SS		ESRIR-based	1
	SAW (+SR)		SAE-CNN-based	1
	SC		CMI-based	1
Structural monitoring	SA			2
		Concrete technique	LGC-based	1
		Optimization	Adam optimization	1
	SC		SCNN-based	1
Total				64

5. Discussion

5.1. Machine Learning and Application Domains

For the implementation of CM, various methods have been utilized. In applications categorized under manufacturing and process, the literature predominantly employs DL and conventional ML techniques. Based on the analysis of 110 examined papers, 76 contributions (69.09%) utilize ML techniques. Among these, 14 papers implement DL (18.42% of all ML models and 12.73% of all examined contributions). Conventional ML techniques can be categorized into 16 different model types, whereas DL methods encompass four model categories. Additionally, 34 (30.91%) contributions address non-ML approaches, which can be further classified into three broad categories.

Analyzing the contributions focused on CM in manufacturing processes, 21 out of 31 applications (67.74%) employ ML models, accounting for approximately two-thirds of the total. A detailed list of applications, the employed techniques, input data, preprocessing techniques, and sensor data sampling frequencies is provided in Table A1. It is evident that despite belonging to the same application domain, no standardized methods or techniques are employed. Even when identical applications and input data are used, different methods and procedures for CM are implemented. For example, in milling applications using vibration data, various methods are applied. However, in some cases, specific applications consistently use the same input data and similar techniques, such as neural networks combined with force data. Among the 21 ML models implemented in the manufacturing category, four (19.05%) are DL models, with 50% being LSTM-based and 50% CNN-based. These models are distributed across different application categories rather than being concentrated in a single category. The sampling frequencies used in these studies vary not only concerning the application and input data but also within the manufacturing category itself.

In the application domain categorized as system-level monitoring, 46 out of 72 publications (63.89%) implement ML. A list of the examined contributions within this category, including their application areas, employed techniques, input data, preprocessing techniques, and sensor data sampling frequencies, is provided in Table A2 and Table A3. These tables distinguish between machine-related and non-machine-related applications. Among the identified ML techniques, nine (12.5%) belong to DL methods. Bearings applications, with 35 contributions, represent the most extensively studied application in this category; however, only two contributions (5.71%) implement DL methods. The remaining DL methods are distributed across various applications without a clear concentration. Notably, vibration data is used in 49 studies across different applications, with 39 exclusively in the system category. Among these, 29 out of 35 publications (82.86%) focused on bearings employ vibration data as input. This suggests a correlation between input data and application domain. Despite the frequent use of identical input data, no standardized techniques are applied within specific applications. Additionally, there is no uniform approach regarding sampling frequencies. However, for similar applications such as bearings, the sampling frequencies generally range between 4 kHz and 20 kHz, with occasional exceptions.

The smallest number of examined papers falls under the structure category. Out of seven studies, four (57.14%) employ ML approaches. A detailed breakdown of applications and other critical parameters is presented in Table A4. Among the four ML methods, only one is a DL approach. Due to the limited number of studies and their diverse nature, neither the use of similar models nor standardized input data can be identified. Furthermore, the choice of sampling frequency varies significantly across the applied methods.

Overall, the examined papers employ a diverse range of methods, including both ML and non-ML techniques, across various application domains. The applied methods differ not only between applications but also within the same application category. This can be attributed to the highly specific focus of each study, even within identical application categories, as well as the extensive availability of different methods that have evolved independently. Due to the large number of analyzed papers, certain correlations can be identified. Broad connections exist between input data and applications, as similar sensors are used across multiple studies. However, significant variations in sampling frequencies within the same application and for identical input data are observed. Different methods may require different frequency ranges. Vibration data is the most commonly used input for CM, appearing in 53 out of 110 (48.18%) analyzed publications, while system-level applications dominate the examined studies. It can be observed that applications requiring rapid response times, such as those in manufacturing, predominantly utilize conventional and simple ML techniques. This is mainly because complex models, particularly DL methods, introduce latency that can hinder system responsiveness. In summary, 76 contributions predominantly employ ML techniques for CM, with DL methods accounting for only 12.72% of all analyzed contributions. The lowest proportion (0.91%) is found in structural applications. Additionally, most studies rely on homogeneous data, meaning that data from a single sensor type is predominantly used. Only about 24.55% of all examined publications utilize more than one sensor type, and approximately 10% employ more than two different sensor types.

5.2. Type of Condition Monitoring and Applications

Predictive maintenance and remaining life estimation exclusively utilize ML models due to their high capability. Among the 110 examined papers, only 18 (16.36%) implement predictive approaches. Of these, four (22.22%) employ DL methods. DL techniques are primarily used in studies combining fault detection with remaining life or fault prediction. Half of the analyzed papers implementing predictive approaches (50%) focus on manufacturing processes, while one-third address machine components. Overall, only a few contributions (38.89%) simultaneously address remaining life prediction, predictive maintenance, and anomaly detection. This may be due to the need for more complex models, which could reduce system responsiveness. A detailed overview of the implemented condition monitoring predictive approaches—highlighting the type of CM, employed models and techniques, data types, and prediction targets—is provided in Table A5.

5.3. Distinction Between Real-Time Categories

The term real-time categorizes the examined papers based on the implementation of techniques, models, sensor data acquisition, and preprocessing techniques within a single processing cycle over a specific timeframe. This implies that the implemented systems can respond to environmental changes within an acceptable time frame to detect and mitigate errors or problems promptly. However, system response times can vary significantly, necessitating a distinction between low-frequency and high-frequency real-time systems. A key question arises: at what frequency or processing time does a system qualify as high-frequency? To address this, the examined papers were analyzed. While some specify the sampling frequency of sensor systems, only few provide concrete data regarding processing speed or response time.

Out of the 110 reviewed papers, only six (5.45%) report execution times for processing, training, or detection tasks as shown in Table A6, Table A7 and Table A8. However, due to the varying operational contexts and hardware requirements or limitations of each application, different hardware components are used, making direct comparisons difficult. Nevertheless, the reported execution times can serve as reference points for achievable processing speeds and performance expectations across different types of applications.

In manufacturing and process-related applications—particularly milling—processing times ranged between 0.25 s and 1.3 s. These include two signal processing techniques, SGW with a processing time of 0.25 s [91], and Katz’s fractal dimension (FD) method at 0.84 s [34]. Two basic ML approaches were also reported: mean shift clustering [67] with a processing time of 1.3 s, and an artificial neural network (ANN) [64] with a processing time of 0.5 s. In contrast, more complex models such as LSTM-based architectures used for cutting monitoring demonstrated significantly longer processing times, reaching 138 s [88].

In the context of system-level applications, specifically for wear particle detection, the WP-DRnet model achieved a total runtime of 23.4 ms for both detection and recognition [83].

Furthermore, rough estimates can be made based on the implemented techniques and sampling frequencies. However, due to the differing hardware components and combinations of techniques used, accurately estimating the processing time is difficult. Additionally, the required frequency varies significantly depending on the application. For example, high response speeds are necessary in production environments and processes, as well as in individual systems like AUVs. In contrast, lower frequencies suffice for defect detection in structures, as these issues can be addressed over an extended period. Even within the same process category, such as pressure systems and bearings, different response time requirements exist. While bearing degradation may not be immediately critical, pressure fluctuations in a system require immediate intervention.

Overall, applications related to manufacturing processes and system categories typically demand higher-frequency CM systems. In contrast, structural applications generally require lower response speeds. However, the monitoring frequencies required are highly dependent on the specific application.

To effectively detect anomalies and quickly respond to potential failures in production processes or system conditions, it would be beneficial to develop systems with very short response times. For such applications, a frequency of 1000 Hz (1 ms response time) would be ideal to immediately correct any deviation from the normal state. In other cases, response times of 10–100 ms (10–100 Hz) may suffice, as already implemented in some of the reviewed concepts. Although the reaction times achieved by the reviewed systems may be sufficient for monitoring purposes, they may be too slow when immediate system adjustments or corrective actions are required. Based on this observation, we propose a general classification of real-time frequency ranges:

CM systems with closed feedback loops and parameter adaptation—such as concrete SASO systems—can be considered high-frequency real-time systems when operating at processing frequencies above 1000 Hz (reaction time < 1 ms). Systems with processing frequencies between 1000 Hz and 100 Hz (1–10 ms reaction time) fall into the mid-frequency range. Those operating below 100 Hz (>10 ms reaction time) are considered low-frequency systems, while systems within the 10–1 Hz range (100–1000 ms reaction time) may be categorized as very low-frequency, which are generally unsuitable for applications requiring fast feedback. A summary of this classification is provided in Table 12.

Table 12. Proposed classification of real-time CM systems based on processing frequency, response time, and typical application domains.

Frequency Range	Processing Frequency	Reaction Time	Applications
High-frequency	>1000 Hz	<1 ms	Manufacturing processes, autonomous vehicles
Mid-frequency	100–1000 Hz	1–10 ms	Bearings, drive systems, underwater vehicles
Low-frequency	<100 Hz	>10 ms	Static structures, rotating machines
Very low-frequency	<10 Hz	>100 ms	Offline diagnostics, long-term condition tracking

Notably, none of the reviewed papers explicitly define required or minimal response times for detection and corrective actions—indicating a significant gap and the need for further research in this area.

Based on this classification, CM systems operating above 1000 Hz can be classified as high-frequency systems. However, as mentioned earlier, since the reported frequencies are based on rough estimates, a definitive classification into different real-time categories remains unresolved.

5.4. Self-* Properties

The results of the analyzed papers regarding their self-* properties reveal that only approximately 58.18% of all reviewed publications implement such features. Among these, 59.38% base their self-* properties on self-adaptive feature extraction methods, while 7.81% use self-adaptive noise canceling algorithms. Additionally, 4.69% solely implement self-adaptive learning rate algorithms, self-adaptive scaling factors, or self-adaptive optimizers. Only 25% of the papers that implement self-* properties present a concrete or partially self-adaptive technique specifically designed for CM. Furthermore, 17.19% of the publications employ methods aimed at achieving self-organization or partial self-organization within the system.

Self-awareness is implemented in only 1.56% of the cases, and self-correcting features appear in just 3.13% of the papers. None of the reviewed publications combine approaches that simultaneously incorporate self-adaptivity, self-organization, self-awareness, and self-correction. However, combining self-adaptivity and self-organization is essential for realizing a SASO (self-adaptive, self-organizing) system.

While most contributions address either the self-adaptivity or the self-organization of the model parameters, only four papers (6.25% of those using self-* properties) implement a controlling component. For instance, Gouarir et al. propose an approach where a CNN learns the behavior of the tool and the workpiece and predicts their states [108]. Based on this prediction, a feedback loop adjusts the feed rate and spindle speed to achieve the desired force. Another paper monitors solder layer aging and adapts the parameters of an initial thermal network model in response to aging effects [118].

Additionally, Liu et al. developed a synchronous switching circuit with a self-adaptive peak-detection system that automatically tracks the displacement amplitude and adjusts the synchronous switching operations accordingly [96]. Gao et al. described an approach for intelligent process optimization and self-adaptive control of the spinning process using predictive modeling of the process behavior [117].

These papers implement system-level self-adaptivity—that is, adaptation to changes in the external environment—moving closer to the concept of OC and the realization of SASO systems. In contrast, most of the other examined publications merely adapt the CM model itself without addressing broader system adaptation.

However, several critical aspects required for a true SASO system are missing across these approaches. Notably, there is no monitoring of sensor health, nor diagnosis and recovery mechanisms (i.e., self-healing capabilities). Furthermore, the self-organizing part—essential for dynamic reconfiguration of model architectures and thus system stability—is absent. Furthermore, none of the examined papers explicitly addresses the aspect of self-explainability.

Moreover, the feedback loop proposed by Gouarir et al. remains purely conceptual [108], and Gao et al.’s work lacks practical experimentation [117]. Although Liu et al. achieve adaptation to amplitude changes, their system remains static and does not continue learning based on new data [96]. None of the systems are decentralized, nor do sensor subsystems communicate with one another. Nevertheless, runtime adaptation of the systems is possible.

Regarding the implementation of a full MAPE-K loop, several key elements are missing. For example, while some approaches implement prediction and reactive adjustments based on current input data, they lack planning components that evaluate and calculate alternative actions. Only Liu et al.’s approach includes a concrete actuator component [96], while the others merely adapt model parameters without the involvement of a direct actuator. Furthermore, there is no dedicated component for storing and utilizing learned knowledge about specific environmental or system states—although past data are sometimes retained for model retraining.

In conclusion, none of the reviewed approaches for CM can be classified as true SASO systems. Although some aspects are addressed, significant efforts are still required to fully realize SASO capabilities in this domain.

To achieve a fully SASO-compliant CM system, particularly in industrial environments, it is recommended to implement the self-properties outlined in Table 13. A critical requirement is the integration of self-adaptivity, which enables the CM system to dynamically adjust to changing environmental conditions—such as varying sensor readings—without missing current or future faults. For example, this may include self-adaptive thresholds for anomaly detection, dynamic feature selection to handle novel data patterns, and continuous adaptation of the system’s internal knowledge base to cope with previously unseen anomalies and measurement values. Such adaptability improves the overall resilience and performance of the system and allows it to react effectively to evolving input conditions.

Table 13. Recommended self-* properties and their corresponding functions for realizing a fully SASO-compliant CM system.

Self-* Property	System Function
Self-adaptive	Dynamically adjusts monitoring strategies (e.g., thresholds, feature selection, network structure, knowledge, sensor parameter).
Self-organizing	Reorganizes sensor networks, internal network structure data fusion, or internal modules based on failure, or environment.
Self-healing	Detects sensor or module failures and automatically compensates or replaces them.
Self-protecting	Identifies tampering attempts or attack-induced anomalies and protects critical components.
Self-explaining	Provides understandable justifications for anomaly detection/prediction, decisions, or adaptations.

In addition to adapting to external environmental changes, the system must also be capable of adapting to changes within its own components. This necessitates the implementation of self-healing capabilities. Over time, sensors and actuators are subject to degradation, which can result in altered readings that do not necessarily indicate a fault. Self-healing mechanisms can detect, diagnose, and compensate for such gradual shifts, recalibrating sensors as necessary and identifying internal failures independently of external conditions.

Another essential property for a SASO CM system is self-organization. Components such as sensors and actuators may degrade at different rates or even fail entirely. Self-organizing capabilities allow the system to autonomously compensate for such failures by restructuring its internal architecture and data pathways. This includes dynamic fusion of heterogeneous sensor data and reorganizing the system’s internal network in response to environmental or structural changes, thereby enhancing the overall system resilience.

In industrial settings, the self-protecting property is particularly vital due to the risk of external disturbances such as cyberattacks. The system must be able to identify malicious interferences or falsified sensor readings and prevent inappropriate self-adaptive or self-healing responses that could compromise safety or system integrity. Protecting critical components from such threats is essential for maintaining secure and reliable operations.

Lastly, self-explaining capabilities are crucial, especially in industrial applications, to ensure transparency and trust in the system. The CM system should be able to provide comprehensible justifications for its actions and decisions, such as anomaly detections, system adaptations, or predictive maintenance measures. This is key to avoiding costly misunderstandings or misinterpretations of system behavior.

While the core self-properties discussed here form a strong foundation, they may be extended by additional self-* capabilities as required by the specific application context.

6. Research Agenda

Following a thorough analysis of the reviewed publications, this section highlights the existing research gaps and explores potential approaches to address them. The examined studies target different application areas and scenarios, employing a variety of models, input data types, feature sizes, sensor types, and numbers of sensors for CM, even within similar or identical use cases. This raises the question:

• Is it possible to develop a concept that performs equally well across different application domains and scenarios?

To explore this, a benchmark study could be conducted to evaluate the performance of a unified concept across multiple use cases. Currently, only a few publications employ more than two distinct sensor types, suggesting considerable research potential in this direction. A key question that emerges is:

• Can the fusion of more than three heterogeneous sensor types improve the performance of CM systems compared to systems with fewer sensor types?

To investigate this, a real-world scenario could be selected in which the number and type (homogeneous vs. heterogeneous) of sensors are varied. The resulting system performance can then be evaluated accordingly.

Moreover, only a limited number of DL models have been proposed for CM of SASO systems, revealing a need for deeper exploration into their advantages over traditional ML methods. In particular, there is a need for systems that integrate prediction with real-time anomaly detection while considering computational efficiency. Given their deeper architectures, DL models are theoretically more capable of learning complex fault features, which may lead to more accurate predictions and anomaly detection. This prompts the following research question:

• Can more efficient and faster DL models be developed to address all aspects of CM in real time?

The literature points to the superiority of Temporal Convolutional Networks (TCNs) over LSTM models in terms of stability, performance, and training speed [122]. Therefore, future research should investigate the integration of TCNs with other model architectures.

However, a major limitation in current studies is the lack of reported processing times, compounded by the use of varying hardware platforms, making it nearly impossible to compare the proposed methods in terms of processing efficiency. Furthermore, none of the existing contributions address the required processing time necessary to effectively carry out CM tasks, especially when integrated into a complete SASO system with a runtime feedback loop for error correction. This leads to several key questions:

• Is a comparison of processing time and performance feasible across proposed methods despite varying application domains and hardware components?
• What processing time requirements must CM systems meet to be sufficiently fast for different application areas and real-time scenarios?
• What constitutes “high-frequency” processing in the context of real-time, and how can real-time be meaningfully categorized based on processing time thresholds?

To address these questions, a benchmark study could be conducted, initially comparing existing methods within similar application areas and input data types. Subsequently, a broader evaluation of processing time requirements across various real-time scenarios should be carried out. This would involve identifying the maximum permissible time by which CM and corresponding system adjustments—potentially via a feedback loop—must be completed. The process speed of the respective systems should be analyzed to support this effort. Moreover, an investigation into real-time categories based on the fastest and slowest required processing speeds can offer a structured classification.

An additional and highly relevant research direction involves the implementation of SASO principles within CM systems and the realization of a complete SASO system. The reviewed literature indicates that only a few studies present systems with true self-* properties—beyond merely self-adaptive feature extraction or self-organizing model parameter adjustments. Crucially, none of the current approaches implement a comprehensive feedback loop that enables real-time SA and SO, nor have they explored the combination of SA and SO capabilities in a unified system. This leads to the following research questions:

• How can a CM system be designed to meet the requirements of a full SASO system?
• Can such a system be integrated into a complete SASO framework with a feedback mechanism or MAPE-K loop, enabling real-time, autonomous adaptation and reconfiguration?

To realize such a system, models based on self-unsupervised learning and continuous self-unsupervised learning should be explored. Additionally, the system should include components capable of autonomously adjusting sensor parameters to the environment, correcting deviations, and self-reorganizing in case of component failures. A control unit that self-adaptively tunes system parameters via feedback and responds to problems through self-organization is also essential.

Initial approaches already exist that address self-healing in manufacturing processes through anomaly detection and pattern recognition, offering valuable starting points [123]. Similarly, the use of TCNs for self-adaptive anomaly detection has been explored and can serve as a foundational method [124]. In addition, unsupervised continuous learning techniques for anomaly detection in industrial environments provide further inspiration [125]. Another publication introduces a combined approach that integrates self-organizing and self-protecting capabilities, which may serve as an early framework [126].

An additional research gap builds upon this: the aspect of self-explainability, which so far has not been addressed in any of the reviewed papers. This raises the following question:

• Can a SASO system be designed in such a way that its actions and decisions are inherently self-explainable?

To address this, one possible approach could involve leveraging XCS (eXtended Classifier Systems) or similar methods to enhance the interpretability of the system’s behavior, providing transparent and understandable explanations for its autonomous decisions [127,128]. Additionally, Gated Recurrent Units (GRUs), as proposed in [129], could be employed to implement mechanisms for self-explainability, providing insights into the system’s internal decision-making processes. In addition, a TCN-based model has already demonstrated self-explaining capabilities and may serve as a foundational approach for further development in this area [130].

A previously unaddressed issue, not covered by the examined papers, is the lack of labeled datasets, which is crucial for comparing the performance of techniques. Furthermore, there is a lack of diversity in the types of anomalies in real-world scenarios, especially those that are unpredictable or rare. These infrequent errors often result in poor performance of CM techniques. This raises two critical questions:

• How can unlabeled real-world datasets be accurately labeled in an unsupervised manner?
• How can the lack of rare anomalies, which are crucial for robust model performance, be addressed?

To tackle the long-tail distribution problem, the dataset shaping approach can be considered [131]. This involves using generative AI to synthesize data-label pairs for underrepresented anomalies—specifically those for which the model shows poor performance due to a lack of training data. In parallel, few-shot or zero-shot learning techniques may help address the challenge of unlabeled datasets by enabling models to generalize to unseen anomaly types with minimal or no labeled examples [132].

To enable community-wide evaluation of SASO-capable CM approaches, a realistic real-time scenario could be implemented using current simulation platforms such as NVIDIA Isaac Sim [133]. This would involve a simulated yet physically plausible environment featuring multiple high-fidelity virtual sensors, a realistic system setup, and diverse operating conditions. The scenario could incorporate realistic challenges such as system and sensor degradation, cyberattacks, and various types of anomalies, thus providing a comprehensive and reproducible testbed for benchmarking self-* capabilities.

7. Conclusions

A comprehensive review of existing contributions on CM of SA and SO systems in the literature has been conducted. For this purpose, an appropriate search strategy was implemented, which identified 284 publications, of which 110 relevant papers were analyzed. To minimize potential selection bias, sources from various publishers were used. The goal of this extensive review was to identify established and emerging techniques and applications, as well as to pinpoint current challenges and open questions for further research. The examined techniques include both ML methods—specifically differentiating DL approaches—and non-ML techniques. The analyzed papers were categorized based on their application and the techniques used for implementing CM. Additionally, the contributions were analyzed with respect to sampling frequency, the types of input data, and predictive approaches.

The majority of the reviewed papers employed ML techniques; however, only a small subset specifically utilized DL. Most studies focused on monitoring individual critical system components, standalone systems, or systems comprising multiple subsystems. A significant number of authors addressed monitoring in manufacturing processes, while static objects and structures received minimal attention. Furthermore, only a limited number of publications explored predictive approaches, often without integrating anomaly detection and system lifespan prediction.

The potential for more precise categorization of real-time systems based on processing speed was discussed. While some authors provided details about sampling frequencies, explicit information on system processing speeds was rarely included. Although a rough estimation for classification was feasible, a definitive definition of high-frequency systems could not be established.

Furthermore, the proposed approaches presented in the selected papers were examined with regard to the implementation of self-* properties and their alignment with the requirements for realizing a SASO system. In addition, the identified research gaps were summarized in a research agenda, outlining potential directions for future investigation.