Development of a Hardware-in-the-Loop Framework for Marine Engine Diagnostics: Performance Evaluation of a Standalone Hybrid ANN Application using PLC-Based Signal Emulation

1.1 Background and motivation

Marine diesel engines, such as the Daihatsu 6PSHTb-26D, serve as the prime movers for a vast majority of merchant vessels. Ensuring their reliability is paramount for maritime safety and economic efficiency. Conventionally, maintenance strategies have relied on scheduled preventive maintenance. However, with the advent of Industry 4.0, there is a paradigm shift towards Condition-Based Maintenance (CBM). Recent systematic reviews indicate that data-driven approaches, particularly Deep Learning (DL), are becoming the dominant trend in maritime predictive maintenance [1]. Algorithms such as Convolutional Neural Networks (CNNs) and Temporal Convolutional Networks (TCNs) have demonstrated superior performance in feature extraction compared to traditional methods [2].

1.2 The challenge: Data scarcity and validation paradox

Despite the theoretical potential of AI, its deployment in the actual maritime sector faces a critical bottleneck: the lack of a verifiable testing environment. Advanced models require not only massive amounts of data for training but also ground-truth scenarios for validation. Currently, three specific limitations hinder the transition of research algorithms to shipboard application:

• Data Scarcity and Class Imbalance: Marine engines typically operate in healthy conditions for the majority of their lifecycle. Collecting run-to-failure data is rare and expensive.
• The Threshold Validation Paradox: Validating the accuracy of diagnostic thresholds presents a dilemma: one cannot scientifically confirm an alarm threshold is correct without triggering an actual fault, yet inducing critical failures on a sailing vessel is strictly prohibited due to safety and economic risks.
• Integration Complexity: Many existing studies focus solely on algorithmic accuracy using offline datasets, neglecting the complexity of integrating these algorithms with real-time industrial communication protocols (e.g., OPC UA, Modbus, SQL) found in modern ship automation systems.

1.3 Related works and research gap

To address data limitations, researchers have explored several approaches. Transfer Learning has been utilized to adapt knowledge from related domains [3, 4], but this still relies on the existence of a high-quality source dataset. Others employ thermodynamic software (e.g., AVL Boost, GT-Power) or Digital Twins to generate synthetic data [5, 6]. While effective for generating data, these purely software-based simulations often lack the Hardware-in-the-Loop (HIL) connectivity required to validate the end-to-end data pipeline. There is a notable gap in the literature regarding low-cost, physical testbeds that allow researchers to validate not just the algorithm, but the entire monitoring architecture—from sensor signal generation to database logging and SCADA visualization—before deploying on a real ship.

While commercial HIL platforms such as dSPACE or OPAL-RT have been successfully employed in marine research for high-speed combustion control and fuel injection analysis [7], they are often prohibitively expensive and rely on proprietary hardware that differs significantly from actual shipboard automation. These systems operate at microsecond-level cycle times, which is essential for crank-angle resolved simulations but excessive for general condition monitoring. In contrast, this study proposes a cost-effective framework using the industrial Siemens S7-1200 PLC. Although its millisecond-level scan time is slower, it is sufficient for emulating thermodynamic process parameters (e.g., temperatures, pressures) characterized by high thermal inertia. The primary distinction is our focus on the "Operational Technology and Information Technology (OT-IT) validation" of the diagnostic pipeline rather than combustion physics. A detailed quantitative comparison between the proposed PLC-based approach and commercial HIL solutions is presented in Table 1.

Table 1. Quantitative comparison between commercial HIL and proposed PLC-based emulation

1.4 Contribution of this paper

Addressing these gaps, this study proposes a HIL Framework using a PLC-based Signal Emulation Strategy. Unlike static datasets or purely virtual models, our approach provides a physical interface to validate the real-time diagnostic pipeline. Specifically, this study offers three key contributions:

• Development of a High-Fidelity Low-Cost HIL Testbed: We constructed a signal emulation system using a Siemens S7-1200 PLC to mimic the thermodynamic behaviors of a Daihatsu 6PSHTb-26D engine. Crucially, the emulation fidelity was rigorously validated against physical engine measurements, achieving a marginal RMSE of 3.24%, proving its reliability as a cost-effective alternative to commercial HIL platforms.
• Adaptive Thresholding Mechanism using Receiver Operating Characteristic (ROC) Analysis: Beyond simple fault injection, the framework introduces a statistical approach to overcome the "Threshold Validation Paradox". By utilizing ROC analysis on the generated synthetic fault data, we established dynamic alarm thresholds that adapt to sensor noise and operational uncertainties, minimizing false alarms.
• Real-time Edge Deployment and Comparative Evaluation: We demonstrate the successful deployment of a Standalone Hybrid ANN on an industrial Edge PC. Comparative benchmarking confirms that the proposed model significantly outperforms traditional algorithms (SVM, Random Forest) with an RMSE of 0.0737℃, while maintaining an ultra-low inference time (0.0410 ms) suitable for real-time PLC synchronization.

1.5 Organization of the paper

The remainder of this paper is organized as follows: Section 2 presents the overall system architecture, detailing the closed-loop data flow between the physical emulation layer and the intelligent processing core via SQL Server integration. Section 3 elaborates on the methodology, including the hardware setup of the simplified simulation kit, the mathematical modeling for signal generation, and the implementation of the hybrid diagnostic algorithm. Section 4 provides a comprehensive analysis of the experimental results, covering the HIL framework fidelity validation, the determination of adaptive thresholds using ROC analysis, and the real-time performance evaluation on the Edge device. Finally, Section 5 concludes the paper and outlines future research directions for commercial SCADA deployment.

3.1 Development of the Simplified Physical Simulation Kit

To mitigate operational safety risks, a Simplified Physical Simulation Kit was developed as a HIL testbed (Figure 2(a)). Centered around a Siemens S7-1200 CPU 1215C PLC and a 10-inch Delta HMI, the system employs a Signal Emulation Strategy rather than static physical sensors. The PLC executes thermodynamic algorithms to generate real-time virtual signals (e.g., 4-20 mA, 0-10 V) for visualization. Crucially, the HMI serves as a Synthetic Fault Injection interface, enabling the generation of labeled data by modifying internal coefficients such as friction $\left(K_{ {fric }}\right)$ or cooling efficiency $\left(\eta_{ {cool }}\right)$.

2a.jpg

(a)

2b.jpg

(b)

Figure 2. The experimental setup: (a) The fabricated Simplified Physical Simulation Kit powered by Siemens S7-1200 PLC and Delta HMI; (b) The actual Daihatsu 6PSHTb-26D marine diesel engine used as the reference model

3.2 Mathematical modeling and diagnostic metrics

To validate the proposed diagnostic framework within a realistic automation environment, a HIL strategy was employed. Aligning with recent SCADA architectures utilizing Siemens S7-1200 controllers [8], physics-based mathematical models were embedded directly into the PLC memory.

While vibration and acoustic analysis offer high sensitivity for fault detection, retrofitting such sensors on existing fleets faces significant cost and bandwidth barriers. To ensure cost-effectiveness and alignment with the standard instrumentation of the reference Daihatsu 6PSHTb-26D engine (Figure 2(b)), this study focuses exclusively on thermodynamic and process parameters. Consequently, the emulation logic addresses five representative Diagnostic Indicators (DIs), derived from fundamental thermodynamics [9-12] and fluid mechanics [13].

3.2.1 Exhaust Gas Temperature (EGT) modeling (DI1)

The EGT dynamics are simulated in three stages: steady-state estimation, fault injection, and thermal inertia. First, the base thermodynamic temperature $\left(T_{{Base }}\right)$ is computed.

While basic control models often simplify the EGT-Load relationship as linear, established literature on control-oriented engine modeling [12] indicates that the thermal response is inherently non-linear due to the variable efficiency of the turbocharger and changing air-fuel ratios. To capture this curvature within the PLC's real-time constraints (without using computationally heavy Wiebe functions), we apply a second-order polynomial correction:

$T_{{Base }}=T_{{Coolant }}+\left(K_{R P M} \cdot n\right)+\left(K_{ {Load } 1} \cdot L \cdot n\right)+\left(K_{{Load } 2} \cdot L^2 \cdot n\right)$                (1)

where, $T_{{Coolant }}$ is the ambient coolant temperature (℃), n is the engine speed (rpm), and L is the dimensionless load factor $(0 \div 1), K_{R P M}, K_{ {Load } 1}$ and $K_{ {Load } 2}$ are empirical coefficients representing speed-dependent friction, linear combustion heat release, and non-linear efficiency corrections, respectively.

To simulate cylinder-specific faults (DI1), a deviation offset is applied to a specific cylinder i based on a fault intensity input $F_{E G T, i}(0 \div 1)$:

$T_{ {Target }, i}=T_{ {Base }} \cdot\left(1+K_{{Fault }} \cdot F_{E G T, i}\right)$               (2)

where, $K_{{Fault}}$ is the sensitivity coefficient. Finally, to mimic the response time of industrial thermocouples [14], a discrete first-order lag filter is applied:

$T_{E G T, i}(k)=T_{E G T, i}(k-1)+\left[T_{T a r g e t, i}-T_{E G T, i}(k-1)\right] \cdot \frac{\Delta t}{\tau}$                (3)

Here, $\Delta t$ is the PLC scan cycle time (s) and $\tau$ is the sensor thermal time constant (s).

3.2.2 Lube Oil Pressure and viscosity dynamics (DI2)

The lubrication system is modeled as a coupled thermal-hydraulic system consistent with marine machinery principles [15, 16]. First, similar to the cooling system, the Lube Oil Temperature ($T_{L O}$) dynamics are simulated based on the heat balance between friction generation and oil cooling efficiency. The instantaneous temperature is updated via a lag filter:

$T_{L O}(k)=T_{L O}(k-1)+\left[T_{ {Target }, L O}-T_{L O}(k-1)\right] \cdot \frac{\Delta t}{\tau_{L O}}$               (4)

where, $T_{{Target,LO }}$ is the steady-state target temperature (℃), $\Delta t$ is the scan cycle time (s), and $\tau_{L O}$ is the thermal time constant of the oil sump (s).

Crucially, the oil viscosity ($\mu$) is dynamically updated based on this real-time temperature using the Andrade Equation [17, 18]:

$\mu=A \cdot e^{\frac{B}{T_{L O}+273.15}}$                (5)

where, A and B are empirical constants specific to the SAE-40 oil grade.

The pressure generation logic incorporates both the electric priming pump and the engine-driven mechanical pump. The mechanical pressure ($P_{ {Mech }}$) is calculated as:

$P_{ {Mech }}=\left(K_{ {Pump }} \cdot n\right) \cdot\left(1+K_{{Visc }} \cdot\left(\mu-\mu_{ {Ref }}\right)\right)$                 (6)

where, n is engine speed (rpm), $K_{ {Pump }}$ is the pump efficiency factor (bar/rpm), $K_{V i s c}$ is the viscosity correction coefficient, and $\mu_{ {Ref }}$ is the reference viscosity at normal operating temperature.

To validate the diagnostic algorithms, the final source pressure is determined by a "Check Valve" logic (max) and then degraded by Filter Clogging ($F_{ {Clog }}$) and Bearing Wear ($F_{ {Wear }}$) faults:

$P_{{Source }}=\max \left(P_{ {Elec }}, P_{ {Mech }}\right)$                (7)

$P_{L O}=\min \left(P_{ {Set }}, P_{ {Source }} \cdot\left(1-K_F \cdot F_{ {Clog }}\right) \cdot\left(1-K_W \cdot F_{ {Wear }}\right)\right)$                  (8)

where, $P_{{Elec }}$ is the priming pump pressure (bar), $P_{{Set }}$ is the relief valve limit (bar), and $K_F, K_W$ are the dimensionless fault severity scaling factors.

3.2.3 Cooling system thermal equilibrium (DI3)

The Jacket Water (JW) temperature model balances heat generation ($Q_{J W}$) against cooling efficiency [10]. The heat rejection is estimated to be proportional to speed and load:

$Q_{J W}=\left(K_{J W, n} \cdot n\right)+\left(K_{J W, L} \cdot L \cdot n\right)$               (9)

where, $K_{J W, n}$ and $K_{J W, L}$ are heat transfer coefficients. The target equilibrium temperature is derived by applying a cooling efficiency factor $\eta_{ {Cooling }}$:

$T_{{Target }}=T_{ {Amb }}+\frac{Q_{J W}}{\eta_{ {Cooling }}}$                (10)

Under fault conditions (DI3 - Radiator Fouling), $\eta_{{Cooling }}$ is reduced from 1.0 via the HMI. Finally, to simulate the significant thermal inertia of the coolant volume, the real-time temperature is updated via a first-order lag function:

$T_{J W}(k)=T_{J W}(k-1)+\left[T_{ {Target }}-T_{J W}(k-1)\right] \cdot \frac{\Delta t}{\tau_{J W}}$                  (11)

where, $\tau_{J W}$ is the thermal time constant of the cooling system.

3.2.4 Governor dynamics and stability (DI4)

The engine speed control is governed by a Finite State Machine (FSM). To simulate the mechanical moment of inertia, the speed transition follows a ramp function with a saturation limiter (sat):

$n_{{Ramp }}(k)=n(k-1)+\operatorname{sat}\left(\frac{n_{{Target }}-n(k-1)}{\Delta t}, K_{ {Ramp }}\right) \cdot \Delta t$                  (12)

where, $K_{R a m p}$ is the maximum acceleration rate (${rpm} / {s}^2$).

Furthermore, acknowledging the importance of stability analysis in control systems [19], a deterministic noise signal is injected to validate the detection algorithm. Although actual governor instability typically manifests as a non-linear limit cycle, this study deliberately employs a fixed-frequency sinusoidal waveform as a standardized Fault Injection signal. This approach isolates the frequency-based instability from stochastic mechanical friction, allowing for the precise calibration of the Neural Network’s frequency extraction sensitivity. The injected speed is defined as:

$n_{{Final }}=n_{{Ramp }}(k)+\left(K_{{Noise }} \cdot F_{ {Gov }}\right) \cdot \sin (\omega t)$                   (13)

where, $F_{G o v} F_{G o v}$ is the fault intensity, $K_{N o i s e}$ is the maximum noise amplitude and $\omega$ is the oscillation frequency $({rad} / {s})$.

3.2.5 Starting air pneumatic decay (DI5)

The starting air pressure ($P_{A i r}$) is modeled using a differential mass balance equation [13]. The net rate of pressure change combines charging ($R_{C h g}$), consumption ($R_{ {Cons }} R_{ {Cons }}$), and leakage ($R_{ {Leak }}$):

$R_{ {Leak }}=-K_{{Leak }} \cdot F_{ {AirLeak }}$                 (14)

$\frac{d P_{A i r}}{d t}=R_{C h g}+R_{C o n s}+R_{ {Leak }}$                 (15)

where, $K_{ {Leak }}$ is the leakage coefficient and $F_{ {AirLeak }}$ is the fault variable $(0 \div 1)$.

Since the PLC operates in discrete time, it explicitly integrates this rate of change to update the system pressure in real-time:

$P_{A i r}(k)=P_{A i r}(k-1)+\frac{d P_{A i r}}{d t} \cdot \Delta t$               (16)

where, $P_{ {Air }}(k)$ is the instantaneous air pressure (bar). This formulation allows the "Air Leakage" logic (DI5) to be validated against mathematically defined decay rates.

3.3 The hybrid diagnostic algorithm

To ensure the proposed solution remains cost-effective and faithful to the configuration of the reference Daihatsu 6PSHTb-26D engine, this study restricts its scope to thermodynamic and process parameters. Consequently, five representative Dis were selected. These DIs are processed by a hybrid algorithm running on MATLAB, categorized into two groups: Rule-based Statistical Logic and Multi-Layer Perceptrons (MLP)-based Normality Modeling.

3.3.1 Group 1: Rule-based statistical logic

Parameters with linear characteristics or distinct physical thresholds are monitored using statistical rules.

DI1 (EGT Deviation): This indicator quantifies the thermal unevenness of combustion. It is calculated as the maximum deviation of any single cylinder's temperature from the instantaneous average of all six cylinders:

$T_{a v g}=\frac{1}{6} \sum_{i=1}^6 T_{E G T, i}$                 (17)

$D I 1=\max _{i=1 \ldots 6}\left(\left|T_{E G T, i}-T_{ {avg }}\right|\right)$               (18)

DI4 (Governor Stability): This metric assesses the mechanical stability of the speed control loop. It is defined as the Oscillation Amplitude ($A_{g o v}$) of the engine speed within a sliding time window (W), active only when the engine is in the running state (SE = 2):

$D I 4=\max _{t \in W}(n)-\min _{t \in W}(n)$                 (19)

DI5 (Air Leak Rate): This indicator detects pneumatic leakage severity by monitoring the instantaneous rate of pressure change. It is formally defined as the time derivative of the air pressure, approximated in the discrete domain as:

$D I 5=\frac{d P_{A I R}}{d t} \approx \frac{P_{A I R}(t)-P_{A I R}(t-\Delta t)}{\Delta t}$                (20)

3.3.2 Group 2: Multi-Layer Perceptrons-based normality modeling

Complex non-linear parameters are modeled using MLP based on standard pattern recognition theory. Adopting a Normality Modeling approach, the networks were trained exclusively on 11,520 samples of healthy data to learn the ideal engine behavior across the full operating range.

To ensure the reproducibility of the experimental results, the specific architecture and training hyperparameters are detailed in Table 2. The model uses a deep topology optimized for non-linear regression using the Levenberg-Marquardt algorithm.

Faults are detected by analyzing the Residual (R) between the measured real-time value and the MLP prediction. The residuals for the two non-linear indicators are defined using the general functional form as follows:

DI2 (Lube Oil Residual): The network predicts the expected oil pressure based on the current engine speed, oil temperature, and load factor. The residual reflects anomalies such as pump degradation or filter clogging:

$D I 2=P_{L O}-P_{ {pred }}\left(n, T_{L O}, L\right)$            (21)

DI3 (Jacket Water Residual): Similarly, the expected cooling water temperature is predicted to detect cooling efficiency losses (e.g., radiator fouling):

$D I 3=T_{J W}-T_{{pred }}\left(n, T_{L O}, L\right)$                (22)

Table 2. Configuration of the proposed Multi-Layer Perceptrons (MLP) network

The selection of this specific topology [3-50-90-1] was based on a preliminary empirical grid search, which identified it as the optimal configuration balancing learning capability (RMSE) and computational efficiency (inference time) for the target IPC hardware.

3.3.3 Output Bit-Packing

To optimize storage and transmission bandwidth, the diagnostic status of all five DIs (classified into Normal, Warning, or Alarm levels) is encoded into a single 16-bit integer (Packed_Alarm_Integer) before being written to the SQL Server.

This study successfully developed and validated a low-cost HIL framework to bridge the data scarcity gap in marine engine diagnostics. By integrating a Siemens S7-1200 PLC with a Python-based Edge IPC, the proposed system effectively emulates the thermodynamic behaviors of a Daihatsu 6PSHTb-26D engine, facilitating the safe injection of critical faults (e.g., cooling failures, lubrication degradation) without risking physical assets.

The key contributions and findings are summarized as follows:

• HIL Fidelity: The PLC-based engine model was validated using a multi-source strategy, combining experimental measurements at idle (L = 0) and manufacturer technical specifications at full load (L = 100%). The system achieved a relative RMSE of 3.24%, confirming its reliability as a digital twin for generating high-quality training data.
• Algorithmic Superiority: A comparative assessment demonstrated that the proposed Hybrid ANN (topology [3-50-90-1]) significantly outperforms established baselines. With an RMSE of 0.0737℃ and an R² of 0.9999, the ANN exhibited superior tracking of non-linear transients compared to Linear Regression, SVR, and Random Forest (RMSE ≈ 2.8236℃).
• Real-time Feasibility: The proposed model achieved an ultra-low inference time of 0.0410 ms on the Edge IPC, ensuring seamless synchronization with the PLC’s telemetry cycle and proving its readiness for real-time onboard deployment.

Future Work: Despite these promising results, the current validation relies on laboratory and simulation data ("Source Domain"). The next phase of research will focus on bridging the Sim-to-Real gap. Specifically, we aim to collect extensive sea-trial data from operating vessels under varying load conditions. To address the inevitable distribution shift between the HIL simulation and the stochastic marine environment, Domain Adaptation techniques such as Maximum Mean Discrepancy (MMD) [23] will be applied. Furthermore, Transfer Learning strategies, including freezing the early feature-extraction layers and fine-tuning the fully connected layers [24], will be investigated to adapt the pre-trained model to specific individual engines.