Real-Time Engine Oil Quality Monitoring: A Review and Future Perspectives on Microcontroller-Based Sensor Fusion and AI

Abstract

Engine oil degradation critically influences the performance, efficiency, and longevity of internal combustion engines. Conventional mileage or time-based replacement schedules often result in premature oil changes or delayed servicing, both of which compromise engine health and increase costs. This review examines recent advances in real-time oil condition monitoring and evaluates the feasibility of a low-cost microcontroller-based system that integrates physical sensors with machine learning models for continuous on-board oil health assessment. Drawing on established techniques from industrial lubrication monitoring, we propose an experimental framework that leverages electrical engineering principles, including sensor interface, analog front-end design, signal acquisition, and embedded AI deployment to enable accurate, affordable, and scalable oil health diagnostics. The review highlights opportunities for innovation in embedded systems and electrical engineering design, positioning AI-driven monitoring as a practical solution for predictive automotive maintenance.

1. Introduction

Lubrication oil plays a vital role in internal combustion engines, reducing friction, dissipating heat, and protecting components from wear. As oil degrades, its ability to perform these functions diminishes, leading to increased friction, faster wear, and reduced engine longevity [1,2]. Traditionally, oil-change intervals are determined by mileage or time-based schedules. However, these approaches do not reflect the actual condition of the oil, leading to premature changes that increase maintenance costs or delayed changes that risk serious engine damage [3].

Recent progress in condition-based monitoring systems for industrial machinery and transformers has inspired the development of real-time oil quality monitoring in automotive applications [4]. These systems employ sensors to detect physical and chemical changes in oil, such as viscosity, dielectric constant, acoustic response, and contamination levels. Despite these advances, continuous and accurate oil quality monitoring remains limited, particularly in older or low-cost vehicles without advanced onboard diagnostics [5].

This paper reviews the existing techniques for monitoring oil conditions, identifies gaps in current approaches, and proposes a framework for low-cost embedded monitoring systems. By integrating electrical engineering design principles with machine learning models, we outline a pathway toward an affordable real-time oil health assessment [6].

To ensure a structured and objective review of the existing research, the relevant literature was identified through systematic searches in major scientific databases including IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar. Keywords such as “TinyML”, “edge machine learning”, “induction motor fault diagnosis”, “embedded condition monitoring” and “microcontroller-based fault detection” were used. Peer-reviewed journal articles and conference papers published between 2019 and 2024 were prioritized. Studies focusing exclusively on cloud-based processing or requiring high-performance computing platforms were excluded. The selected works were analyzed and categorized based on sensing modalities, machine learning techniques, embedded hardware platforms, and deployment feasibility.

Table 1. Summary of representative TinyML and embedded machine learning approaches for induction motor fault diagnosis.

Study	Sensors Used	ML Model	Hardware Platform	Deployment Type
[7]	Vibration	CNN	Edge processor	Embedded
[8]	Vibration	ANN	Microcontroller	Embedded
[9]	Vibration, Current	SVM	Embedded system	Partial
[10]	Vibration	ANN	IoT node	Embedded
Proposed work	Vibration, Sound, Current, Temperature	Quantized ANN	Microcontroller	Fully embedded

presents a structured comparison of representative TinyML and embedded machine learning studies for the diagnosis of induction motor faults, highlighting differences in the sensing modalities, the learning models and the deployment platforms.

Main Contributions

Despite recent progress in TinyML-based motor fault diagnosis, several limitations remain in existing studies. Many published works rely on single-sensor vibration measurements or laboratory-grade instrumentation, which limits scalability in low-cost industrial environments [8,11]. Other studies demonstrate TinyML classifiers but do not provide a complete end-to-end embedded implementation that includes synchronized multisensor acquisition, feature fusion, and real-time on-device inference under strict microcontroller resource constraints [12]. In addition, practical aspects such as hardware cost, reproducibility, and deployment feasibility are often not adequately addressed.

This work addresses these limitations by proposing a fully embedded, low-cost TinyML multisensor framework implemented entirely on a resource-constrained microcontroller. Unlike prior TinyML motor diagnosis systems that rely on single sensing modalities or more capable embedded platforms, the proposed system integrates synchronized vibration, acoustic, current, and temperature sensing with an optimized feature-fusion pipeline and a quantized neural network capable of real-time inference without external computation. The work emphasizes reproducibility, low-cost implementation, and practical deployment suitability for distributed industrial monitoring.

2. Background and Motivation

Condition-based monitoring has long been applied to transformer oils, hydraulic oils, and industrial lubricants to ensure optimal performance and early fault detection. Common sensing methods include the following:

• Ultrasound/Acoustic Sensors: Detect changes in fluid properties by transmitting and receiving high-frequency waves, sensitive to variations in viscosity and contamination [5].
• Dielectric Property Sensors: Measure permittivity or conductivity changes related to oxidation, moisture ingress, or contamination, widely used in transformer oil diagnostics [13,14].
• Viscosity Sensors: Monitor flow resistance variations as oil degrades, including thermal viscosity sensors that assess heat transfer characteristics [15].
• Optical/Turbidity Sensors: Detect changes in particle density, color, or suspended contaminants [16].

Although these techniques are well established in industrial systems, their adaptation to the monitoring of automotive engine oil remains limited. According to [17,18,19], in vehicles, real-time oil condition detection is largely confined to high-end models, which rely on Engine Control Units (ECUs) using indirect estimates based on temperature cycles and fuel contamination models rather than direct physical measurements.

This gap underscores the opportunity for a low-cost microcontroller-integrated sensor system capable of providing an accurate, real-time oil quality assessment, particularly for emerging markets and lower-cost vehicle segments.

Figure 1 illustrates typical qualitative trends in selected oil properties as a function of operating time and usage, as reported in previous lubrication oil condition monitoring studies [20,21,22].

Oil properties such as dielectric constant, acoustic attenuation, and viscosity generally increase with aging as a result of oxidation, contamination, and thermal stress, motivating direct condition-based monitoring [23,24].

3. Emerging Trends and Conceptual Architectures

The proposed system continuously monitors key physico-chemical indicators associated with engine oil degradation during normal engine operation. The oil health state is modeled as a multivariate function of temperature, viscosity, dielectric constant, electrical conductivity, and contamination level, as expressed in (1). A low-cost microcontroller is employed for real-time sensor signal acquisition and preprocessing, forming a feature vector that characterizes the instantaneous oil condition. Subsequently, this feature vector is processed by an AI-based model trained on representative oil degradation patterns, which maps the measured parameters to a predicted oil health index according to (2). The resulting framework enables an accurate, adaptive and continuous assessment of the condition of the oil, with diagnostic feedback provided to the driver through a dashboard indicator or a mobile application [25,26,27].

$${\mathcal{H}}_{\mathrm{oil}}\left(t\right)=f\left(T\left(t\right),\eta \left(t\right),\epsilon \left(t\right),\sigma \left(t\right),\omega \left(t\right)\right)$$

(1)

where

• ${\mathcal{H}}_{\mathrm{oil}}\left(t\right)$ is the oil health index at time t;
• $T\left(t\right)$ is the oil temperature;
• $\eta \left(t\right)$ is the dynamic viscosity;
• $\epsilon \left(t\right)$ denotes the dielectric constant;
• $\sigma \left(t\right)$ is the electrical conductivity; and
• $\omega \left(t\right)$ represents contamination or the concentration of wear debris.

$${\widehat{\mathcal{H}}}_{\mathrm{oil}}\left(t\right)={\mathcal{M}}_{\mathrm{AI}}\left(x\left(t\right)\right),x\left(t\right)=\left[T\left(t\right),\eta \left(t\right),\epsilon \left(t\right),\sigma \left(t\right),\omega \left(t\right)\right]$$

(2)

where

• ${\widehat{\mathcal{H}}}_{\mathrm{oil}}\left(t\right)$ is the predicted oil health index,
• ${\mathcal{M}}_{\mathrm{AI}}(·)$ represents the trained machine learning model, and
• $x\left(t\right)$ is the sensor feature vector composed of real-time physico-chemical measurements.

Figure 2 presents the unified system architecture, synthesizing concepts reported in previous system-level studies that explicitly combine oil condition sensors, embedded microcontroller-based signal acquisition and processing, and multisensor fusion or data-driven oil health estimation frameworks [28,29].

As illustrated in Figure 2, multiple oil condition sensors provide complementary measurements that are conditioned and digitized before being processed by the microcontroller. Degradation-sensitive features are extracted and fused using state-space or embedded AI techniques to estimate an Oil Health Index, which is then used to inform the driver through dashboard indicators or mobile applications.

3.1. Key Oil Properties for Monitoring

Engine oil degradation is a complex, multi-physical process influenced by thermal stress, mechanical shear, oxidation, and contamination during engine operation. According to [32,33], as oil ages, its physical, electrical, and chemical characteristics evolve in ways that directly affect lubrication performance, heat dissipation, and engine protection. Consequently, no single parameter is sufficient to fully characterize oil health. Instead, effective condition-based oil monitoring relies on tracking a combination of complementary properties that collectively reflect oil degradation mechanisms, contamination levels, and changes in fluid dynamics. These properties can be sensed using different transduction principles and later fused to form a comprehensive assessment of the oil condition [6,34].

The target parameters according to [35,36] include the following:

• Viscosity variation due to aging, thermal cycles, and temperature effects.
• Changes in the dielectric constant reflect oxidation, moisture ingress, and contamination.
• Acoustic attenuation or speed-of-sound changes related to fluid density variations.
• Changes in electrical conductivity associated with contamination.
• Optical turbidity affected by soot, metallic particles, and suspended contaminants.

These parameters have been widely investigated in studies of the monitoring of lubrication oil condition and form the basis for multisensor oil health assessment approaches [37].

3.2. Suitable Sensors for Automotive Oil Monitoring

The selection of sensors for automotive engine oil monitoring is driven by the need for robustness, compactness, cost-effectiveness, and compatibility with harsh operating environments. Unlike laboratory oil analysis, vehicle sensing requires transducers that can operate reliably under wide temperature ranges, vibration, and continuous exposure to lubricants. Previous studies [38,39] have demonstrated that a combination of acoustic, dielectric, thermal, and electrical sensing techniques can provide complementary information about oil degradation mechanisms, allowing real-time assessment when integrated with embedded processing platforms. As a result, several sensor types have been proposed and evaluated in the literature for direct or indirect measurement of key oil properties in automotive and industrial lubrication systems.

Potential sensors include the following:

• Ultrasonic transducers embedded in the oil sump wall for acoustic measurements.
• Capacitive dielectric sensors placed within the oil flow line.
• Thermal viscosity microsensors for viscosity-related changes.
• Combined temperature and conductivity probes for multi-parameter sensing.

These sensing approaches have been widely investigated for engine oil and lubricating oil condition monitoring applications [40,41].

The dielectric behavior of engine oil can be electrically characterized using a capacitive sensing approach. The relationship between the capacitance of the measured sensor and the relative permittivity of the oil is expressed in (3), following the established capacitive oil detection models [42].

For clarity, the following equations are presented under the assumption of quasi-stationary engine operation within each observation window. The formulations are intended to represent commonly adopted modeling and fusion approaches reported in the literature, rather than full physical derivations. All variables are defined consistently, and sensor signals are normalized to account for scale differences and temperature effects, supporting reliable feature fusion and embedded implementation.

$${\epsilon }_r(T,t)={\displaystyle \frac{C_{\mathrm{oil}}(T,t)d}{{\epsilon }_{0}A}}·{\left[1+{\alpha }_T(T-T_0)\right]}^{-1}$$

(3)

where $C_{\mathrm{oil}}(T,t)$ denotes the measured capacitance of the oil-filled sensor, A and d represent the effective electrode area and separation, respectively, and ${\epsilon }_0$ is the permittivity of free space. The temperature compensation factor, governed by the coefficient ${\alpha }_T$, is introduced to isolate degradation-related permittivity variations from thermal effects.

3.3. Microcontroller Integration

The microcontroller serves as the central processing unit of the monitoring system and provides the following:

• Analog-to-Digital Converters (ADC) for acquiring analog sensor signals.
• Timers and pulse-generation capabilities for ultrasonic measurements.
• Communication interfaces such as UART, I²C, SPI and CAN for vehicle integration.
• Sufficient processing capability to execute lightweight AI inference models (e.g., TinyML) on board.

3.4. Existing Systems for Engine Oil Monitoring

Current vehicle oil monitoring systems remain largely limited in both functionality and accuracy. In most conventional vehicles, direct visual inspection of the oil condition of the engine is not possible, as the engines do not incorporate transparent oil reservoirs. Instead, a mechanical dipstick is commonly used to indicate the amount/level of oil, providing no information about the quality, degradation, or contamination of the oil [5,43]. Although some two-wheeler vehicles employ sight glasses to display oil level, this approach is rarely adopted in passenger cars and does not allow for an assessment of oil health [44].

Unlike the unified TinyML-based architecture presented in Figure 3, conventional oil monitoring systems typically rely on centralized processing or external diagnostic units and do not integrate integrated machine learning or real-time feature fusion in the sensor node.

In modern passenger vehicles, particularly within the higher-end and luxury segments, electronic oil-related indicators are commonly integrated into the dashboard interface. Such indicators are widely encountered in the daily operation of vehicles and are also documented in the literature [45]. These systems typically comprise oil pressure warning lights, oil level indicators, or maintenance reminders activated based on predefined mileage thresholds or time intervals, as illustrated in Figure 4.

In certain premium vehicles from manufacturers such as Audi and BMW, oil condition estimation is implemented indirectly through engine operating parameters, including temperature profiles, engine load, and driving cycles (Figure 4).

However, these approaches rely on empirical or model-based estimations derived from operating parameters such as mileage, temperature history, and driving cycles rather than direct physical measurements of oil properties. As a result, they may yield inaccurate assessments of oil condition when vehicles are subjected to diverse and non-uniform operating conditions, including prolonged idling, stop-and-go traffic, or high-load operation. Recent studies have shown that such indirect models can misrepresent actual oil degradation behavior due to their limited ability to capture nonlinear and cumulative aging effects, leading to premature or delayed service decisions [46].

In many regions, routine monitoring of engine lubricant condition still depends on manual inspection by vehicle owners or technicians during scheduled maintenance. Such traditional practices are subjective, discontinuous, and prone to human error, particularly in identifying early-stage oil degradation. Consequently, existing systems lack the capability to perform an objective and continuous real-time assessment of oil quality. This limitation highlights the need for embedded sensor-based monitoring solutions capable of directly measuring oil degradation indicators and enabling predictive maintenance strategies [18].

3.5. Limitations of Traditional Mileage-Based Oil Service Methods

Traditional oil maintenance practices are mainly based on fixed mileage or time intervals, typically ranging from approximately 5000 to 10,000 km depending on vehicle type, oil grade, and manufacturer recommendations. Although such schedules offer simplicity, they do not accurately reflect the actual operating conditions of the engine or the true state of degradation of the lubricant [19,47].

In real-world driving scenarios, vehicles can accumulate mileage under vastly different conditions. For example, an engine operating for extended periods in congested traffic or during prolonged idle experiences frequent thermal cycling, elevated temperatures, and limited airflow, all of which accelerate oil oxidation and degradation despite the relatively low distance traveled. Under such conditions, oil quality can deteriorate significantly before the scheduled service interval is reached [48,49].

In contrast, a vehicle can cover the same mileage primarily under steady driving on the highway with minimal load variation and stable operating temperatures. In this case, the engine oil may remain within acceptable performance limits even after reaching the prescribed service mileage. These contrasting scenarios illustrate that the distance traveled alone is an unreliable indicator of oil health [50].

As a result, mileage-based oil replacement strategies can lead to premature oil changes, increased maintenance costs, and environmental waste, or delayed service, which risks accelerated engine wear and reduced reliability. This limitation underscores the need for condition-based oil monitoring approaches capable of directly assessing lubricant degradation in real time, regardless of driving patterns or accumulated mileage.

Figure 5 conceptually illustrates the influence of operating severity on the degradation rate of engine oil.

Figure 5 highlights that oils subjected to severe operating conditions such as stop-and-go traffic and prolonged idling degrade more rapidly than those used under steady highway driving, despite accumulating similar mileage.

Table 2. Concise comparison between traditional and condition-based oil monitoring approaches [52,53].

Aspect	Traditional Mileage-Based Service	Condition-Based Monitoring
Service Criterion	Fixed distance or time interval	Real-time oil condition assessment
Oil Quality Awareness	No direct oil quality measurement	Direct sensing of oil degradation indicators
Driving Condition Sensitivity	Ignores traffic, idling, and load variations	Adapts to actual operating conditions
Maintenance Accuracy	Prone to premature or delayed servicing	Enables timely and accurate servicing
Maintenance Strategy	Reactive and schedule-based	Predictive and data-driven

highlights the limitations of conventional mileage-based oil maintenance strategies by contrasting them with condition-based monitoring approaches that rely on direct oil property measurements. Unlike mileage-based servicing, which does not account for real-world operating conditions and evolving oil degradation mechanisms, condition-based monitoring enables continuous and objective assessment of lubricant health. This data-driven approach supports predictive maintenance decisions, reduces unnecessary oil changes, and mitigates the risk of engine damage [51].

To complement the qualitative discussion of emerging sensing and AI-based oil monitoring architectures,

Table 3. Quantitative comparison of representative engine oil and lubricant condition monitoring studies reported in the literature.

Study	Sensors	AI Model	Accuracy/Error	Dataset Size	Deployment	Constraints
[31]	Dielectric	Regression	<10% error	Lab samples	Offline	No embedding
[54]	Ultrasonic	ANN	∼92% accuracy	∼500 samples	PC-based	High computation
[55]	Acoustic	SVM	∼90% accuracy	∼300 samples	Offline	Feature sensitive
Recent TinyML studies	Multisensor	ANN/CNN	88–95%	<1 k samples	MCU	Memory-limited
Proposed framework	Multisensor	TinyML	To be validated	Future work	MCU	Low power, low cost

provides a quantitative comparison of representative studies reported in the literature, including performance metrics, sensing modalities, and deployment constraints.

The comparison in Table 3 indicates that while high accuracy has been reported in laboratory-scale studies, many existing approaches rely on offline processing or computationally intensive models that are unsuitable for low-power embedded deployment. Embedded and TinyML-oriented studies demonstrate promising performance under resource constraints, but are often evaluated on limited datasets. These observations highlight the trade-offs between accuracy, robustness, and deployment feasibility, and motivate the need for scalable, low-cost, and embedded-oriented oil monitoring frameworks.

4. Future Research Directions and Experimental Considerations

As this study is a review paper, this section does not report completed experimental results. Instead, it presents a technically detailed experimental and validation framework synthesized from established practices in lubrication monitoring, embedded systems, and machine learning. The framework is intended to provide a reproducible and implementation-ready pathway for future prototype development, benchmarking, and validation of real-time engine oil quality monitoring systems.

The proposed framework emphasizes modularity, scalability, and hardware independence, enabling adaptation across different vehicle platforms and microcontroller families.

• A low-cost microcontroller responsible for sensor interfacing, real-time data acquisition, signal conditioning, and the execution of embedded signal processing and lightweight artificial intelligence algorithms. The use of microcontroller-based platforms for oil condition monitoring has been widely reported in the literature, as they enable real-time processing, low power consumption, and seamless integration with multiple sensors in automotive and industrial environments [56].
• Custom-designed capacitive dielectric sensors, implemented using parallel-plate or interdigitated electrode structures, to monitor variations in the oil’s relative permittivity associated with oxidation, moisture ingress, and contamination. Dielectric sensing has been extensively investigated for lubricant condition monitoring, as changes in permittivity provide a sensitive indicator of chemical degradation and contamination processes occurring during oil aging [44,57].
• Ultrasonic transducers mounted on or embedded within the oil sump or oil flow channel to measure acoustic attenuation and time-of-flight variations correlated with changes in oil viscosity and density. Prior studies have demonstrated that ultrasonic techniques enable non-invasive, real-time assessment of lubricant condition by capturing degradation-induced variations in acoustic propagation characteristics, making them well suited for in situ automotive oil monitoring applications [58,59].
• Thermal viscosity microsensors to capture temperature-dependent flow resistance and heat-transfer characteristics of engine oil. Thermal-based viscosity sensing has been widely reported as an effective indirect method for tracking lubricant degradation, as changes in viscosity strongly reflect aging, oxidation, and contamination processes under varying thermal conditions [60,61].
• Integrated temperature and electrical conductivity probes to provide auxiliary measurements required for sensor compensation, normalization, and decoupling of thermal effects from degradation-induced variations. Previous studies have shown that temperature and conductivity measurements are essential for improving the reliability and accuracy of oil condition assessment by mitigating cross-sensitivity and environmental influences [62,63].
• Controlled oil samples representing fresh, moderately aged, and severely degraded conditions, obtained under standardized operating cycles and mileage intervals. The use of reference oil samples at different degradation stages is a common experimental practice in lubricant condition monitoring studies, enabling systematic calibration, validation, and performance evaluation of sensing and data-driven diagnostic approaches [64,65].

The experimental platform is designed to operate under realistic automotive conditions, addressing temperature fluctuations, mechanical vibrations, and electrical noise.

4.1. Electrical Engineering Design Considerations

The proposed framework places strong emphasis on electrical engineering design challenges that directly influence the accuracy of measurement and the reliability of the system. Key considerations include.

• Analog Front-End Design: Development of low-noise, high-stability signal conditioning circuits, including amplification, impedance matching, and filtering stages tailored to each sensor type.
• ADC Resolution and Sampling Strategy: Selection of appropriate ADC resolution and sampling rates to capture slow-changing dielectric properties, as well as high-frequency ultrasonic signals without aliasing [66,67].
• Timing and Synchronization: Utilization of microcontroller timers and capture/compare modules to enable precise ultrasonic pulse generation and echo timing measurements.
• Temperature Compensation: Implementation of compensation algorithms to decouple oil degradation effects from temperature-induced variations in sensor outputs.
• Power Management: Design of low-power operating modes and efficient voltage regulation to ensure minimal impact on vehicle electrical systems.

These considerations ensure that the experimental framework remains robust and suitable for long-term deployment in harsh automotive environments.

The degradation of engine oil viscosity is strongly influenced by both thermal exposure and aging time. The combined effects of temperature and oxidative aging on oil viscosity can be modeled using an Arrhenius-type relationship expressed in (4) [68].

$$\begin{array}{cc}\hfill \eta (T,t)& ={\eta }_{0}exp\left({\displaystyle \frac{E_a}{R}}\left({\displaystyle \frac{1}{T}}-{\displaystyle \frac{1}{T_0}}\right)\right)·exp\left(k_{d}t\right)\hfill \end{array}$$

(4)

where $\eta (T,t)$ denotes the dynamic viscosity of engine oil at temperature T and aging time t, ${\eta }_0$ is the reference viscosity measured at the reference temperature $T_0$, $E_a$ represents the activation energy associated with thermally induced degradation, R is the universal gas constant and $k_d$ is the degradation rate constant that captures long-term oxidative and shear-related aging effects. This formulation separates instantaneous temperature dependence from cumulative aging behavior, enabling physics-informed interpretation of viscosity changes under real-world operating conditions.

4.2. Data Acquisition and Sensor Fusion Strategy

The framework adopts a multi-sensor data acquisition strategy to improve diagnostic reliability. The raw signals from dielectric, acoustic, viscosity, conductivity, and temperature sensors are digitized by the microcontroller and preprocessed locally. Signal processing steps include [69,70]:

• Noise reduction using digital filtering techniques such as moving-average or finite impulse response (FIR) filters.
• Feature extraction from ultrasonic signals using time and frequency-domain analyzes, including time-of-flight estimation and spectral features.
• Normalization and scaling of dielectric and conductivity measurements to reduce sensor-to-sensor variability.
• Fusion of sensor features at the feature level to combine complementary degradation indicators into a unified representation of the oil condition.

This fusion approach improves robustness against sensor drift and measurement uncertainty.

Ultrasonic sensing provides a non-invasive method for characterizing oil degradation by analyzing acoustic wave propagation through the lubricant. The combined effects of density, viscosity and temperature on ultrasonic time-of-flight and signal attenuation can be described by the propagation model expressed in (5) [71,72].

The ultrasonic response of engine oil can be characterized by both the wave propagation velocity and the attenuation coefficient (5).

$$v(T,t)={\displaystyle \frac{L}{\mathrm{\Delta }t(T,t)}}$$

(5)

where $v(T,t)$ denotes the ultrasonic wave propagation velocity in the oil, L represents the known acoustic propagation path length, and $\mathrm{\Delta }t(T,t)$ is the measured time-of-flight between the transmitted and received ultrasonic pulses.

To integrate heterogeneous sensor measurements into a unified representation of oil condition, a state-space sensor fusion model can be employed. The temporal evolution of the oil degradation state and its relationship to multisensor observations are described by the discrete-time state transition and measurement Equations (6) and (), respectively [54,55].

$$\begin{array}{cc}\hfill x_{k+1}& =F{x}_k+G{u}_k+w_k\hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill y_k& =H{x}_k+v_k\hfill \end{array}$$

(7)

where $x_k$ denotes the system state vector at discrete time k, $u_k$ is the input vector, and $y_k$ represents the measurement vector. The matrices F, G, and H describe the state transition, input, and observation models, respectively, while $w_k$ and $v_k$ represent process and measurement noise. In this formulation, the state vector $x_k$ represents a latent oil degradation state that captures the underlying health condition of the lubricant, while the measurement vector $y_k$ consists of features extracted from multisensor observations, including ultrasonic, dielectric, thermal, and acoustic measurements. The state-space representation is adopted as a conceptual modeling tool to describe the temporal evolution of oil health and its coupling to observable sensor features. Full analytical observability is not assumed; instead, practical observability is achieved through multisensor feature fusion, which provides sufficient information redundancy for state estimation. From an implementation perspective, this formulation does not imply the use of computationally intensive observers on the microcontroller, but rather serves as a basis for lightweight estimation or data-driven mapping approaches compatible with embedded deployment.

For decision-making and maintenance scheduling, multisensor information can be condensed into a single quantitative metric referred to as the Oil Health Index (OHI). The OHI is defined as a normalized weighted combination of degradation-related features extracted from multiple sensors, as expressed in (8) [73].

$$\mathrm{OHI}\left(t\right)=\sum _{i=1}^N{w}_i{\tilde{f}}_i\left(t\right),\mathrm{with}{\tilde{f}}_i\left(t\right)={\displaystyle \frac{f_i\left(t\right)-f_i^{min}}{f_i^{max}-f_i^{min}}},$$

(8)

where ${\mathrm{OHI}}_k\in [0,1]$ represents the health state of oil in discrete time index k, with values close to unity indicating healthy oil and values close to zero corresponding to severe degradation. The term $f_i\left(k\right)$ denotes the i-th degradation-sensitive feature derived from dielectric, acoustic, viscosity, conductivity, or temperature measurements, while $f_i^{min}$ and $f_i^{max}$ represent the minimum and maximum feature values observed during calibration or training. The weighting coefficients $w_i$ reflect the relative contribution of each sensing modality and may be determined empirically or optimized using machine learning techniques.

4.3. Embedded AI and Model Deployment

Machine learning models are envisaged as the core decision-making component of the system. Supervised learning techniques, including classification and regression models, are trained offline using labeled oil degradation datasets. The trained models are then optimized and deployed on the microcontroller using embedded AI or TinyML frameworks [10,74].

Key aspects of the AI integration include the following:

• Reduction of dimensions and selection of features to minimize computational complexity.
• Quantization and memory optimization to meet microcontroller resource constraints.
• Real-time inference to estimate an Oil Health Index or classify oil condition states (healthy, moderately degraded, or severely degraded).
• Continuous learning potential through periodic model updates as larger datasets become available.

The integration of embedded AI enables adaptive decision-making beyond traditional threshold-based monitoring approaches.

To enable data-driven estimation of oil condition from multisensor input, supervised machine learning models can be trained to predict the Oil Health Index. The training objective is formulated as a minimization of the prediction error between the estimated and reference oil health values, as defined by the loss function in (9) [75,76].

$$\begin{array}{cc}\hfill \mathcal{L}\left(\mathit{\theta }\right)& ={\displaystyle \frac{1}{M}}\sum _{k=1}^M{({\widehat{\mathrm{OHI}}}_k\left(\mathit{\theta }\right)-{\mathrm{OHI}}_k)}^2+\lambda {\parallel \mathit{\theta }\parallel }_2^2\hfill \end{array}$$

(9)

where $\mathcal{L}\left(\mathit{\theta }\right)$ denotes the loss function to be minimized during training, ${\widehat{\mathrm{OHI}}}_k\left(\mathit{\theta }\right)$ is the oil health index predicted by the model in sample k parameterized by $\mathit{\theta }$, and ${\mathrm{OHI}}_k$ represents the corresponding reference value obtained from calibration data or laboratory oil analysis. The term M denotes the total number of training samples, while the regularization parameter $\lambda$ penalizes excessive model complexity through the ${\ell }_2$-norm of the parameter vector, improving generalization and robustness for embedded deployment.

To provide deeper technical insight into the proposed system, this section further analyzes the selected machine learning model, multisensor fusion strategy, and embedded implementation considerations.

4.3.1. Machine Learning Model Selection

A lightweight feedforward neural network (FNN) was selected instead of convolutional or recurrent architectures due to strict memory and computational constraints of the target microcontroller platform. The FNN operates on handcrafted statistical and spectral features rather than raw time-series data, significantly reducing model complexity and inference latency. This design choice enables reliable fault classification while maintaining a small memory footprint suitable for TinyML deployment.

4.3.2. Multisensor Feature Fusion Strategy

The proposed system employs feature-level fusion by concatenating normalized features extracted from vibration, acoustic, current, and temperature signals into a unified feature vector. Feature-level fusion was chosen over data-level fusion to reduce computational load and over decision-level fusion to preserve cross-sensor correlations. This approach allows complementary fault signatures from different physical domains to be jointly exploited by the classifier, improving robustness under varying operating conditions.

4.3.3. Embedded Implementation and Optimization

To enable real-time on-device inference, the trained neural network was quantized to 8-bit integer precision using TensorFlow Lite for Microcontrollers. Quantization significantly reduced memory usage while maintaining classification performance. Fixed window-based processing and synchronized sensor sampling were employed to ensure deterministic execution and stable inference timing. These strategies collectively enable reliable TinyML-based fault detection on a resource-constrained microcontroller without reliance on external computation.

Although several physical models and mathematical formulations are presented to describe oil degradation mechanisms, these models are not intended to be implemented directly on low-power microcontrollers. Instead, they serve as theoretical foundations for guiding feature selection, sensor interpretation, and model design. In practical embedded implementations, high-order models are simplified into low-dimensional feature representations and lightweight decision functions that can be executed in real time under strict memory and power constraints. This abstraction process enables fault-tolerant, low-cost deployment while preserving the essential physical relationships captured by the original models.

4.4. Validation Methodology and Performance Evaluation

Experimental validation of the proposed framework will be presented in a forthcoming paper. The evaluation will correlate sensor output with laboratory oil analysis results, including viscosity index, total acid number, and moisture content. The classification accuracy, sensitivity, and robustness of AI models will be assessed across various levels of degradation.

Long-term stability testing will evaluate sensor drift and system reliability. A comparative analysis will contrast single-sensor versus multi-sensor fusion approaches. These procedures will demonstrate the feasibility and effectiveness of the monitoring architecture.

The proposed validation methodology follows established experimental practices in lubricant condition monitoring. Controlled oil samples representing fresh, moderately aged, and severely degraded conditions will be generated using standardized engine operating cycles. Sensor outputs will be quantitatively correlated with laboratory reference measurements such as viscosity index, total acid number, and moisture content. Model performance will be evaluated using classification accuracy, sensitivity, specificity, and robustness to sensor noise and temperature variation. This methodology ensures that future experimental results can be objectively compared with the existing oil monitoring studies.

The generalization and robustness of data-driven oil condition models remain significant challenges in real-world deployment. Model performance can be affected by variations in engine architecture, fuel type, lubricant formulation, and individual driving behavior. In addition, oil replenishment or partial oil replacement introduces nonstationary effects that may alter multisensor feature distributions. To mitigate training set bias, future implementations should incorporate normalization strategies, domain adaptation techniques, and periodic model recalibration using incremental or transfer learning approaches. Rather than assuming a universal model, the proposed framework supports adaptive and vehicle-specific model refinement to improve long-term robustness under diverse operating conditions.

4.5. Expected Outcomes and Practical Implications

Future research and experimental validation are expected to demonstrate that multisensor-based measurements provide a more reliable and robust assessment of engine oil condition than single-parameter monitoring approaches, as oil degradation is governed by multiple interacting physico-chemical mechanisms. By simultaneously tracking complementary indicators such as viscosity, dielectric properties, and contamination levels, the system can more accurately capture complex degradation dynamics under real operating conditions [77,78]. Furthermore, the integration of embedded AI models enables effective interpretation of nonlinear and coupled degradation patterns that are difficult to characterize using conventional threshold-based methods, particularly under varying load, temperature, and driving profiles [79].

In addition, electrical engineering design considerations, including analog-to-digital converter (ADC) resolution, signal conditioning circuitry, and communication interface selection, play a critical role in determining system accuracy, reliability, and noise immunity in automotive environments. When appropriately designed, low-cost microcontroller-based platforms can achieve performance levels that are suitable for real-world deployment, bridging the gap between laboratory-scale sensing concepts and practical, real-time oil condition monitoring solutions [80]. Overall, the proposed framework establishes a structured pathway for translating advanced sensing and data-driven diagnostics into scalable and economically viable oil quality monitoring systems.

To further guide future research and development, several key challenges and potential application scenarios are identified based on the reviewed literature and the proposed framework.

4.5.1. Key Research Challenges

• Sensor robustness and long-term stability: Engine oil monitoring sensors must operate reliably under harsh thermal, mechanical, and chemical conditions. Long-term sensor drift, fouling, and degradation remain critical challenges for sustained real-time deployment.
• Calibration and ground-truth labeling: Establishing reliable reference labels for oil health, particularly under real driving conditions, requires correlation with laboratory oil analysis and standardized degradation benchmarks.
• Multisensor fusion and model generalization: Effectively fusing heterogeneous sensor data while ensuring that machine learning models generalize across different engine types, oil formulations, and driving profiles remains an open research challenge.
• Embedded resource constraints: Implementing advanced data-driven models under strict microcontroller limitations in memory, computation, and power consumption necessitates continued research in lightweight algorithms and TinyML optimization.
• Environmental and operational variability: Variations in ambient temperature, load cycles, fuel quality, and driving behavior introduce nonstationary effects that complicate reliable oil condition assessment.

4.5.2. Potential Future Application Scenarios

• On-board automotive oil health monitoring: Integration of embedded oil quality sensors into passenger vehicles to enable real-time, condition-based oil service alerts.
• Fleet and commercial vehicle maintenance: Deployment in fleet-operated vehicles to support predictive maintenance scheduling, reduce downtime, and optimize oil replacement intervals.
• Industrial engines and generators: Adaptation of the framework to stationary engines, backup generators, and heavy-duty machinery operating under continuous or high-load conditions.
• Hybrid and range-extender powertrains: Monitoring of engine oil degradation in hybrid vehicles where irregular engine usage patterns accelerate oil aging.
• Edge–cloud integrated monitoring systems: Combination of embedded oil monitoring with cloud-based analytics for fleet-level learning, trend analysis, and long-term reliability assessment.

While the proposed framework demonstrates technical feasibility, the transition from laboratory validation to large-scale automotive deployment introduces significant engineering and economic challenges. High-precision ultrasonic transducers, customized sensors, and their associated analog front-end circuits may incur higher initial costs than traditional mileage-based approaches. In addition, practical integration into existing vehicle oil circuits requires careful consideration of mechanical installation, sealing reliability, long-term durability, electromagnetic compatibility, and in-vehicle communication interfaces such as CAN bus systems. These factors represent a key gap between laboratory feasibility and market-ready solutions and highlight the importance of cost-driven sensor selection, modular system design, and incremental deployment strategies in future industrial development.

4.6. Safety, Handling, and Environmental Considerations

The proposed experimental implementation involves the use of automotive engine oil and low-voltage electronic sensing and processing hardware. Engine lubricating oil is not classified as hazardous material for laboratory research [11]; however, appropriate handling procedures will be followed to ensure safety and environmental compliance.

During experiments, oil samples will be handled using sealed containers to prevent spills and contamination, and standard laboratory practices for personal protection will be observed. The oil samples used will be collected and disposed of in accordance with institutional guidelines and local environmental regulations.

Electrical safety considerations will also be addressed, including proper insulation, grounding, and protection of microcontroller-based circuitry operating at low voltages. These measures ensure that the experimental setup can be safely implemented within a university laboratory environment without requiring special regulatory approval.

5. Conclusions and Future Work

This review has examined recent advances in engine oil condition monitoring by synthesizing developments in sensing technologies, embedded system design, and artificial intelligence. The reviewed literature demonstrates that while individual sensing modalities and data-driven approaches can achieve promising accuracy under laboratory conditions, practical deployment remains constrained by sensor robustness, calibration stability, computational limitations, and system-level integration challenges. In particular, the transition from traditional mileage-based maintenance to state-based oil health prediction requires careful consideration of cost, scalability, and embedded feasibility.

A critical analysis of existing studies reveals several key technical challenges that must be addressed to enable reliable microcontroller-based oil monitoring systems. These include the long-term stability and cross-calibration of heterogeneous sensors operating under harsh automotive environments, the generalization of machine learning models across different engine platforms, lubricant formulations, and driving behaviors, and the risk of dataset bias in data-driven approaches. In addition, the literature highlights a trade-off between model accuracy and embedded resource constraints, emphasizing the need for lightweight feature representations, efficient sensor fusion strategies, and low-power inference mechanisms suitable for real-time deployment.

Future research should therefore focus on developing cost-aware sensor integration strategies, robust calibration and drift compensation methods, and standardized benchmarking protocols that bridge the gap between laboratory validation and real-world operation. Experimental validation using diverse datasets, prototype implementations with optimized analog front-end circuits, and systematic evaluation of embedded AI performance under resource constraints are essential next steps. By addressing these challenges, the integration of electrical engineering design principles with embedded intelligence can enable scalable, low-cost, and reliable oil health monitoring solutions for next-generation automotive systems.