Detection of Pipe Ruptures in Shipboard Firefighting Systems Using Machine Learning and Deep Learning Techniques

Abstract

In this paper, the application of machine learning and deep learning algorithms for fault and failure detection in maritime systems is examined, specifically focusing on the detection of pipe ruptures in a ship’s saltwater firefighting (FiFi) system using pressure sensor data. Neural network models were developed to distinguish between normal operational states and anomalies, as well as to accurately locate pipe faults within the ship. Data were collected using real-world tests with FiFi system sensors, capturing both normal operations and simulated pipe ruptures, and were meticulously labeled to facilitate neural network training. Two neural network models were introduced: one for classifying system states (normal or anomalous) based on time-series pressure data, and another for identifying the location of anomalies related to pipe ruptures. Experimental results demonstrated the efficacy of these models in detecting and localizing pipe faults, with performance evaluated using mean squared error (MSE) across different network configurations. The successful implementation of these machine learning and deep learning algorithms highlights their potential for enhancing maritime safety and operational efficiency.

1. Introduction

Artificial intelligence (AI) serves as a facilitating technology extensively applied in numerous industrial sectors; however, in naval and maritime industries, its development remains nascent.

One of the potential domains where AI can be harnessed aboard ships pertains to fault detection. The introduction of machine- and deep learning-based systems for managing system faults on vessels enables the automatic (or nearly automatic) detection of system faults in real-time through the analysis of data gleaned from sensors embedded in the ship’s systems. These AI systems offer a multitude of advantages, including the following:

• Fault Detection: The identification of anomalies or data readings that fall outside the normal range, as captured by sensors distributed across various vessel systems, such as pressure sensors, flow meters, vibration sensors, and temperature sensors, among others.
• False Alarm Identification: Timely decision making is achieved to differentiate genuine system faults from events such as sensor malfunctions or signal loss.
• Precise Fault Localization: By leveraging sensor data comparisons, accurate determination of the fault’s location within the system can be achieved.
• Prescription of Action Steps: After the identification and localization of a detected fault, the system can assist in decision making by providing actionable recommendations or even autonomously executing corrective measures. For instance, in the event of a ruptured section of piping, the system may advise the individual in charge of system control to close specific valves.

In this work, the utilization of neural networks for the detection and localization of pipe ruptures within the ship’s firefighting system is proposed. This contribution centers on addressing a literature gap by applying AI technologies to the specific domain of fault detection within the maritime industry. Unlike other industrial applications that employ dense sensor networks with broad coverage, this approach utilizes the existing sensor network available, which, in the case of a medium size ship, could be limited to 15–25 pressure transmitters. This limitation presents a significant challenge for fault detection and localization, as the reduced number of sensors could affect the ability to monitor and accurately locate ruptures. However, the proposed methodology is tailored to this limitation using neural networks to process the sensor data, allowing for effective fault detection, including the localization of pipe ruptures.

To provide a comprehensive overview, this paper is organized as follows: The Section 1.1 reviews related work in the field of fault detection in maritime systems. The Section 2 describes the methodology used for data collection and neural network model development. The Section 3 presents the experimental setup and results, including performance evaluation metrics. The Section 4 discusses the implications of the findings and potential applications in maritime safety. Finally, the Section 5 concludes the paper and suggests directions for future research.

1.1. Related Work

1.1.1. Fault Detection in Pipelines

Detecting faults in pipeline systems is a challenge that has been addressed across multiple engineering disciplines for decades. The need for swift identification and localization of ruptures or blockages within pipes has prompted significant advancements in the field.

Techniques for identifying leaks in pipelines range from basic manual inspections conducted by trained technicians to the utilization of advanced satellite imagery. Conventional approaches for detecting issues in pipelines, particularly in water distribution and hydrocarbon systems, are typically grouped into three categories: biological methods, hardware-based methods, and software-based methods [1]. Biological methods focus on visual inspections, while software-based methods incorporate various computational strategies. Hardware-based methods, on the other hand, are classified based on the types of sensors and devices used for fault detection. In their work, Datta and Sarkar [2] offer a comprehensive review of the key methodologies and technologies for detecting pipeline blockages and leaks, as summarized in

Table 1. Main technologies for pipe blockage/leakage detection.

Blockage Detection Techniques	Leakage Detection Techniques
Vibration analysis: This technique involves monitoring the vibrations in pipes to identify potential blockages. Changes in vibrations can indicate flow issues.	Negative pressure wave-based leak detection system: This technique is based on the detection of negative pressure waves generated by leaks in pipes.
Pulse echo methodology leak detection: This technique uses ultrasonic pulses to detect leaks in pipes. The reflection of pulses can reveal the presence of leaks.	Fiber sensor based: It uses fiber optic sensors to detect changes in tension or deformation caused by leaks in pipes.
Acoustic reflectometry: It relies on the emission and detection of acoustic waves to locate blockages in pipes. The reflection of waves can indicate obstructions.	Support vector machine-based pipeline leakage detection: This technique employs machine learning algorithms like support vector machines to identify patterns related to leaks.
Transient wave blockage interaction and blockage detection: This technique utilizes transient waves to detect blockages in pipes. Changes in wave propagation can signal blockages.	Piezoelectric acoustic emission sensor: This sensor detects acoustic emissions caused by leaks in pipes.
Stochastic successive linear estimator: It is based on mathematical and statistical models to estimate the presence of blockages in pipes.	Filter diagonalization method: This method utilizes signal processing techniques to identify noise patterns related to leaks.
	Harmonic wavelet analysis: This method focuses on decomposing acoustic signals into harmonic components to detect leaks.
	Genetic algorithm in combination with the inverse transient method: This method combines genetic algorithms with modeling techniques to detect and locate leaks.
	Computational fluid dynamic simulation: This method uses numerical simulations to predict and detect leaks based on fluid flow behavior in pipes.

.

These technologies are diverse and are used for detecting both blockages and leaks in piping systems, each with specific approaches and methods to address these issues. The choice of the appropriate technology will depend on the specific needs and conditions of the pipeline network being monitored.

Similar classifications can be found in different studies related, above all, to the oil and gas sector. For example, Lu et al. [3] classifies detection techniques into hardware-based and software-based methods (

Table 2. Software and hardware detection methods.

Hardware-Based Method	Software-Based Method
Optical method: This method relies on optical technologies, such as lidar (light detection and ranging), laser diodes, thermal imaging, and spectral imaging. These techniques use light or electromagnetic radiation to detect and analyze objects or phenomena.	Signal processing-based method: Signal processing techniques are employed to analyze data collected from sensors placed along the pipeline. Signal processing algorithms, such as filtering, noise reduction, and feature extraction, are used to identify anomalies in the collected data. Deviations from normal operating conditions, such as sudden pressure drops or irregular flow patterns, can trigger alerts for potential pipeline faults.
Acoustic method: Hardware-based acoustic methods involve the use of sound waves, acoustic emissions, ultrasonic waves, and sonar for detection and analysis. They are often used in applications like underwater sonar systems and structural health monitoring.	Real-time model-based method: Real-time models of pipeline systems are developed using mathematical and computational techniques. These models simulate the expected behavior of the pipeline under various operating conditions. During pipeline operation, the software continuously compares real-time sensor data with the predictions from the model. Any significant deviation between observed data and model predictions can indicate a fault in the pipeline. For example, a sudden drop in pressure that is not predicted by the model may suggest a leak or blockage.
Distributed optical fiber sensor method: Distributed optical fiber sensors use optical fibers as sensing elements. They can measure various physical parameters like temperature, strain, and vibration along the length of the fiber, making them suitable for structural health monitoring and environmental sensing.	Neural network method: Artificial neural networks are trained using historical data from pipeline operations. These networks learn to recognize patterns and correlations within the data. Once trained, neural networks can analyze real-time sensor data and identify abnormal patterns that may indicate pipeline faults. For instance, a neural network can detect changes in flow rates, temperature, or pressure that deviate from expected behavior, suggesting the presence of a fault.
Dynamic pressure transmitter method: This method uses dynamic pressure transmitters to measure pressure changes in real-time. It is commonly employed in industries like oil and gas for monitoring pipelines and wellbore pressures.	Piezoelectric acoustic emission sensor: The piezoelectric acoustic emission sensor is a device used for detecting and analyzing acoustic emissions in pipelines. These sensors are sensitive to acoustic signals generated by pipeline defects such as cracks or leaks. When a fault occurs, the sensor converts acoustic waves into electrical signals. These signals can be analyzed to pinpoint the location and severity of the fault. This method is valuable for early fault detection and structural integrity monitoring in pipelines.
Tracer method: Tracer methods involve introducing a substance or marker into a system and tracking its movement to gather information. This is used in various applications, including hydrology, environmental monitoring, and industrial processes.	Statistical method: Statistical methods involve the use of statistical analysis to detect anomalies in pipeline data. By establishing statistical norms for various pipeline parameters, such as pressure or flow rates, any data that fall outside these norms can be flagged as a potential fault. Statistical process control charts and outlier detection techniques are often used to identify such anomalies. Unusual data points or trends in the statistics can indicate a pipeline fault.
GPR: GPR uses radar pulses to image the subsurface of the ground. It is commonly used in geophysical and engineering applications for locating objects or anomalies underground.	Harmonic wavelet analysis: Harmonic wavelet analysis is employed in pipeline fault detection to extract and analyze harmonic components from sensor data. By identifying specific harmonic patterns in the data, it can reveal irregularities or resonances in the pipeline, often linked to defects. This analysis is particularly useful when investigating faults that result in characteristic harmonic responses.
SmartBall method: The SmartBall method typically involves a free-swimming tool equipped with sensors that can be used for inspecting pipelines. It is used in the oil and gas industry to detect leaks and anomalies.	Harmonic analysis: Harmonic analysis is applied in pipeline fault detection to decompose sensor data into harmonic components. It helps identify the fundamental frequencies and harmonics associated with pipeline conditions. When faults cause changes in harmonic patterns, this method can signal abnormalities in pressure, flow, or other parameters. It aids in the recognition of deviations from normal pipeline operation.

).

The following considerations can be made regarding the analysis performed using the different methods of detecting faults in pipelines:

• Most hardware-based methods focus on the use of sensors and systems, which in some cases involve an additional cost of installation. In addition, thinking about the possible application to the ship’s firefighting system, it can be observed that many of these methods are not recommended for use in marine environments due to the problems that can be generated in systems that work with seawater.
• The use of software-based methods has the important advantage that it will require minimal or no investment in additional equipment (including additional sensors or other equipment). However, the use of software-based methods also has some disadvantages, almost always related to the time needed for their development and commissioning, including time required for the development of simulation models that can replicate the system with sufficient reliability and time for data collection in the case of models that are based on real operating data (data-driven models). The following references have been identified in which the use of simulation models for pipe fault detection is analyzed [4,5,6,7,8].
• It has been proven that many of the developments analyzed use pressure measurements for fault detection [9,10,11,12].

1.1.2. AI-Based Fault Detection in Pipelines

There are some examples of using AI in pipe fault detection. Laurentys presents the development of an expert system for fault detection [13]. However, everything seems to indicate that the tendency is to take advantage of the benefits that neural networks (NNs) can present for the resolution of problems of this type (analyzing data series, pattern recognition). As examples of what is claimed, both Sukarno and Zhao employ neural networks to detect problems in subsea pipeline systems [14,15].

Table 3. Recent developments in the use of neural networks for detection of pipe faults. Adapted from [16].

Reference	Relevant Information
Leu and Bui [17]	Leu and Bui’s study focuses on using Bayesian learning within a neural network to analyze pressure data in a water distribution system. The goal is to detect anomalies or faults in the distribution system.
Zadkarami et al. [18]	This research employs a multi-layer perceptron neural network to analyze pressure and flow data in a simulated crude oil distribution system. The neural network is trained to identify and predict potential issues within the system.
Zadkarami et al. [19]	Zadkarami uses a Dempster–Shafer multi-layer perceptron classifier to analyze pressure and flow data in a simulated crude oil distribution system. This approach aids in the classification and detection of faults.
Gómez-Camperos et al. [20]	In this study, Gómez-Camperos employs a multi-layer perceptron neural network to analyze experimental flow data in a water distribution system. The focus is on the detection of abnormalities within the distribution network.
Jia et al. [21]	Jia et al. use a neural network to analyze pressure data in an experimental water distribution system. While the specific type of neural network is not mentioned, it is used to detect potential issues within the system.
Pulido et al. [22]	Pulido and the team utilize the backpropagation algorithm within a neural network to analyze pressure data in an experimental steam plant system. This aids in identifying and addressing potential faults in the system.
Kang et al. [23]	Kang and colleagues employ a neural network in the space of states to analyze pressure data in an experimental water distribution system. This approach enhances fault detection capabilities within the network.
Zhao et al. [15]	Zhao et al. apply a convolutional network to analyze pressure data in a simulated water distribution system. The neural network is designed to enhance the identification of faults within the simulated environment.
Javadiha et al. [24]	Javadiha employs both convolutional networks and Bayesian reasoning within a neural network framework to analyze pressure data in an experimental water distribution system. This multifaceted approach aids in the detection and classification of faults within the system.

summarizes the information presented by Pérez-Pérez, compiling recent studies related to this field [16].

These studies collectively showcase the versatility and effectiveness of neural networks in fault detection within pipe systems, whether in water distribution, crude oil transportation, steam plant operations, or simulated environments. Neural networks prove valuable in enhancing the reliability and safety of these critical infrastructure systems.

Moreover, as can be seen, while there is a significant number of developments employing neural network algorithms to detect problems in pipe systems, no references have been found in which NNs are applied to ship systems. Although there are AI-based developments, particularly those involving agents, these are focused on reconfiguration after fault in piping systems on ships, rather than on detection [25].

1.1.3. Adaptability of Neural Networks for Fault Detection in Shipboard Systems

The application of neural networks (NNs) for fault detection in pipelines has been extensively studied in various domains, including water distribution, crude oil transportation, and steam plant systems [17,18,19,20,21,22,23,24]. These studies demonstrate the effectiveness of NNs in analyzing pressure and flow data, enabling accurate identification of anomalies and faults. Unlike traditional analytical methods, such as Takagi–Sugeno observers, NNs do not rely on explicit mathematical models, making them particularly suitable for complex and variable environments like shipboard firefighting systems.

Ship firefighting systems operate under dynamic and challenging conditions that include the use of seawater, which can corrode or degrade certain hardware-based sensors, which can hinder effective data acquisition and nonlinear interactions among components such as pumps, valves, and piping networks, which are difficult to represent using analytical models.

Methods based on mathematical observers, such as the Takagi–Sugeno descriptor systems, require precise modeling of system dynamics [26,27]. While effective in well-defined systems, these methods face challenges when dealing with unmeasurable premise variables or highly variable operational conditions, such as those encountered in maritime applications.

NNs provide a viable alternative by leveraging historical operational data to identify patterns and correlations. This data-driven approach enables NNs to detect faults robustly, even in the presence of noise or missing data. Additionally, NNs can be adapted to different types of ships with varying system configurations by retraining the models with data specific to each application, thereby enabling both flexibility and scalability.

Although analytical methods and NNs represent different paradigms, they can be integrated to complement each other. Combining the analytical strengths of observer-based methods with the adaptive capabilities of NNs may facilitate the development of hybrid models that enhance fault detection and localization in shipboard systems, offering greater reliability across diverse maritime platforms.

2. Materials and Methods

The development presented is focused on the implementation of an AI-based algorithm that allows, from the readings and analysis of the data obtained from the pressure sensors installed in the ship’s FiFi system, the automatic detection of faults, allowing “real” time to achieve the following: (1) Detection of the fault: Identification is achieved by anomalous data readings in pressure sensors, and (2) Identification of the location of the detected fault: By comparing the data from the different sensors for the same event, the location of the fault within the system is identified.

In the development of this work, three differentiated stages can be identified:

• Analysis of the Real System: In the initial phase, an in-depth analysis of the actual system was performed, and a profound understanding of its components, characteristics, and operation was gained.
• Data Acquisition from the Real System: Following this, the acquisition of data from the actual system was carried out, ensuring that enough representative data were obtained.
• Development of the Data-Driven Fault Detection and Localization System: Leveraging the collected and labeled data, the development of a data-driven system was initiated. This system is grounded in neural networks for the detection of anomalous pressure patterns and the pinpointing of fault locations.

When selecting the most suitable technique for fault detection, the no free lunch theorem has been considered [28]. This theorem underscores that there is no universally superior algorithm applicable to all problems. Hence, the decision to employ neural networks for identifying pressure drops in a pressurized water system must be based on the method’s appropriateness for the specific problem characteristics.

After the extensive review of related studies, it was decided to use a system based on machine and deep learning compared to other methods, such as the use of a simulation model based on physics, as proposed in [29,30].

Although physics-based models have important advantages [31,32], there are limitations to their application in the proposed case. One of the most important is related to the following. The ship’s firefighting system is complex, feeding salt water into various components on the vessel. Some of these components have manual openings that are not monitored by the control system, making it challenging to create a simulation model that ensures all events occurring in the system are simulated.

On the other hand, the use of data-driven models also has important advantages [33,34]. Specifically, in this case, it is intended to differentiate the pressure drop patterns in the case of openings of components that use saltwater from the FiFi system with respect to those that can occur due to rupture of a pipe in the system. The use of neural networks for pattern recognition for anomalies detection is widely [35,36,37]. In addition, it is proposed that with these same pressure measurements, it is possible to identify the point at which the fault has occurred. In this case, a classification neural network is proposed, a methodology that is also widely tested in other areas [38,39].

The main challenge in the use data-driven models lies in obtaining enough data to cover all the possible scenarios that the ship may face in real life. For this purpose, it is important to note that it seeks to validate the use of a fault detection and location system with the added value that in the future it will continue to learn once implemented on board the ship.

Finally, it should be noted that in numerous literature, it is concluded that important advantages can be achieved using models that combine physics and real data [40,41,42].

However, in the use case considered, it should be noted that it has been applied to a real ship in operation and, particularly, in a system with a high degree of monitoring of pressure signals, which implies that a significant amount of data can be relatively easily collected. Therefore, significant benefits will not be brought by the use of a combined physical and data-driven system versus a data-driven system. However, its application should be considered in possible developments focused on the use of these systems in other contexts where, for example, the degree of sensorization is not as high, or it is intended to be applied in the early stages of the ship’s life cycle.

In conclusion, the choice of machine/deep learning methods, and specifically, neural networks are justified both by their inherent capabilities and the specific requirements of the ship’s operational environment, underscoring the no free lunch theorem’s principles. References to related studies validate the decision-making process, emphasizing the importance of choosing an algorithm based on the intricacies of the problem at hand.

3. Results

3.1. Use Case Description

The development of the use case has been guided by the methodology described in the preceding section:

(1)

The first stage focused on the analysis of the real system.

(2)

After that, data acquisition of the real system will be addressed.

(3)

Finally, from the real labeled data, the development of the data-driven detection and localization system was carried out.

3.2. Real System Description

The saltwater firefighting system delivers pressurized saltwater to the ship’s fire nozzles. It also provides pressurized saltwater to other specialized firefighting services, such as sprinklers and foam stations, along with other shipboard systems and consumers, including the wastewater treatment plant, garbage cooling system, stabilizing fins, propellers, and transversal thrusters, among others.

This firefighting system will be pressurized and supplied with saltwater by four centrifugal pumps designed for continuous operation. Each pump is equipped with a speed regulator and draws water directly from the respective saltwater collectors located within the same chamber. The ship is divided into two main fire zones, with two pumps assigned to each zone.

Furthermore, the FiFi pumps must have adequate capacity to provide the maximum flow required by the firefighting system, even with one pump out of service. The system will be arranged in a vertical ring configuration, subdivided into two main horizontal collectors. The upper collector will be positioned on the main deck, while the lower collector will be located beneath the second deck. Each of the two longitudinal horizontal collectors that form the system can supply the total required flow to the FiFi system under emergency conditions.

3.3. Data Acquisition

Data acquisition tests were conducted on a real ship. For this purpose, a custom datalogger was developed and integrated into the ship’s control system network. The historical data gathered under controlled conditions were subsequently utilized for training the neural networks.

It is important to note that the sensor information data used in this study is confidential. The data were collected under strict confidentiality agreements to protect the proprietary information of the ship’s systems and ensure the privacy of the operational details. As such, specific details regarding the sensor data cannot be disclosed.

Several tests were performed with the objective of collecting a substantial amount of data, categorized into two types: normal operation and fault behavior.

First, data regarding the normal operation of the system (the FiFi system consuming water) were collected. To collect data corresponding to the normal operation of the system, all relevant consumers were opened, including firefighting nozzles with different flow rates, space sprinklers, and foam stations. Additionally, other shipboard systems were activated, such as bilge ejectors, the wastewater treatment plant, the waste cooling system, stabilizing fins, propellers, and transversal thrusters. These actions allowed for the collection of diverse data on the system’s behavior under typical operational conditions. Figure 1 presents two examples of the pressure drop data series collected during the system’s normal operation.

On the other hand, efforts were made to replicate pipeline faults, such as ruptures. For this purpose, quick-opening valves were installed at various points in the system, under the assumption that their behavior would mimic the effects of a pipe section breaking. These quick-opening valves were strategically placed in different locations within the system branches, with varying pipe diameters, to simulate ruptures with different flow rates. Figure 2 illustrates the pressure drops recorded during two pipe break events.

Although this approach aimed to cover a range of rupture scenarios, it is acknowledged that the data collected from these simulated breaks may be insufficient to represent all possible pipe rupture scenarios comprehensively. Nonetheless, the primary goal at this stage of development was to validate the proposed methodology. Future phases will involve extensive neural network training incorporating various fault scenarios, combining real-world data with simulated data. This approach will help address a broader spectrum of pressure drop profiles that may occur under different fault conditions.

It is important to highlight that adjustments to the signal sampling frequency and the precision of pressure measurements (significant digits) were necessary to effectively characterize the pressure drops. These considerations, which may initially appear minor or insignificant, must be carefully addressed when planning data acquisition for digital twin developments. This is because certain functionalities, particularly those leveraging AI or advanced data analysis, may require high-frequency data to detect transient events, whereas routine ship operations do not typically demand high-frequency or high-precision data collection.

Based on the tests conducted, the following key recommendations and limitations should be considered for the training of neural networks aimed at detecting and localizing faults using real data collected from the ship:

• Data acquisition should be conducted at a frequency of no less than 10 Hz, with a recommended frequency of 20 Hz or higher to achieve optimal system performance. This conclusion is based on the observation that, in cases where data were recorded at lower frequencies, the collected information was insufficient to capture critical events and, therefore, could not be effectively utilized for analysis. Higher sampling frequencies are particularly important for detecting rapid changes in pressure or flow, which are often indicative of faults or anomalies. Insufficient sampling rates may result in data that fail to accurately represent the dynamics of the system, potentially leading to undetected issues or delayed responses.
• For pressurized water systems operating within a range of 8 to 12 bar, it is essential that the pressure signal be recorded with an accuracy of at least one-hundredth of a bar. This conclusion stems from tests conducted with data recorded at a lower precision, specifically with a resolution of one-tenth of a bar, which proved unsuitable for detecting subtle pressure variations.
• The duration of data samples is another critical factor in the detection and diagnosis of faults. Time series shorter than 2 s have been shown to be inadequate for accurately identifying anomalies in system operation, as they do not provide enough information to characterize the behavior or trends in the data. Samples of at least 3 s yield significantly better results, enabling more reliable fault detection. Furthermore, extending the sample duration beyond 3 s enhances the accuracy and robustness of anomaly identification, as larger datasets allow for a more comprehensive analysis of system behavior. However, it is also important to balance sample size with real-time processing capabilities, as excessively long durations may increase computational demands and affect the responsiveness of the fault detection system.

3.4. AI Development Description

3.4.1. Fault Detection Network

This neural network can identify abnormal operations within the system and can additionally specify the zone of the ship where the anomaly is occurring. Neural networks (NNs) are computational models designed to simulate the functioning of neurons in the human brain, with the objective of solving complex computational problems that are not easily programmable [43]. Therefore, these artificial neural networks are trained to learn from data.

A multi-layer perceptron (MLP) architecture was employed in this study. The MLP is composed of multiple layers of neurons: an input layer, one or more hidden layers, and an output layer. Each neuron in each layer is fully connected to every neuron in the next layer. This configuration allows the network to learn complex patterns and relationships within the dataset, making it suitable for the task of detecting and localizing faults in the ship’s firefighting system. The developed algorithm has a general scheme like the following (Figure 3):

The input data for the neural network consist of a series of pressure measurements from each sensor, capturing both time and pressure values. The output is a binary indicator (0/1) that signifies either a fault or normal operation of the system. In the neural network model, each node functions as a neuron, and the connections between nodes in different layers are represented by arrows. These connections involve multiplying the neuron’s output by a weight value, which can either enhance or diminish the activation level of neighboring neurons. Furthermore, at the neuron’s output, an activation function or threshold may be applied. This function adjusts the output value or imposes a limit before it propagates to the next neuron, playing a crucial role in the network’s learning process.

The architecture chosen for the fault detection was a fitting neural network. In this context, the network is trained on a set of inputs to produce corresponding target outputs. Various configurations of the network, including different numbers of layers and neurons, as well as different sizes of input data samples, were explored to identify the optimal configuration.

A feed-forward, fully connected neural network architecture was selected due to its wide usage and versatility. Error correction learning, specifically a back propagation/gradient descent algorithm, was employed to allow the network to self-adjust its parameters. Bayesian regularization was chosen as the training algorithm to avoid potential overfitting issues. This decision was based on preliminary results obtained using MATLAB’s NFTOOL, which allows for straightforward neural network training without programming. The results with Bayesian regularization significantly improved upon those obtained using the Levenberg–Marquardt algorithm.

Since there is no established rule for determining the optimal number of layers or neurons for a neural network, a sensitivity analysis was conducted to identify the optimal structure that minimizes error: a number of layers that varies from 1 to 3 and a number of neurons that varies from 5 to 30.

Manual selection of cases for training, validation, and testing was performed to ensure a correct distribution of real-world cases.

The primary evaluation metric used was the mean squared error (MSE). The network’s performance was assessed based on the training, validation, and test errors to ensure the models’ generalization capabilities. The performance results were obtained from the three best networks stored by the program for each configuration.

Figure 4 illustrates the algorithm proposed for developing the neural network.

To develop both the detection and localization algorithms, the standard model selection and error estimation procedure was employed [44]. This procedure encompasses a set of techniques utilized in machine learning and statistics for selecting models and estimating generalization error.

The objective of this procedure is to select a model that best fits the available data while avoiding overfitting and underfitting. Overfitting occurs when a model fits the training data too closely, capturing noise and losing its ability to generalize to new data. Conversely, underfitting arises when the model is too simplistic to capture the underlying relationships in the data.

The standard model selection and error estimation procedure generally follows the following steps:

• Data Splitting: The available data are divided into training and test (or validation) sets. The training set is used to fit and train different models, while the test set is used to evaluate the performance of the models and estimate generalization error.
• Model Selection: Different models with different structures or hyperparameters are chosen, which may vary in their complexity. This may involve selecting different neural network architectures, different machine learning algorithms, or different feature sets.
• Training and Cross-Validation: Models are trained on the training set and evaluated on the validation set or through cross-validation. Cross-validation involves dividing the training set into several parts and performing multiple rounds of training and evaluation on different subsets.
• Model Evaluation and Selection: Models are compared and evaluated using performance metrics such as mean squared error, accuracy, F1 score, etc. The model with the best performance on the test set or cross-validation is selected.
• Generalization Error Estimation: Once the final model is selected, the generalization error is estimated by evaluating its performance on the test set, which contains unseen data during the training and selection processes.

The model architecture is grounded in an artificial neural network consisting of one or more hidden layers (defined by ’layers’) and a variable number of neurons in each hidden layer (defined by ’neurons’). The flexibility of the network architecture allows for the inclusion of 1 to 3 hidden layers, each featuring a number of neurons ranging from 5 to 50.

The search is constrained for the number of layers within a range of a minimum of one layer to a maximum of three layers. The models considered span from a simple neural network to a deeper network with three hidden layers.

For the numbers of neurons per layer, exploration covers a range from 5 to 50 for each hidden layer. This encompasses models with varying levels of complexity and the capacity to learn sophisticated representations.

Granularity pertains to the number of different values considered for each hyperparameter. In this case, the following parameters are used:

• For the number of layers (’layers’), evaluation encompasses three different values: 1, 2, and 3. This provides a reasonable granularity for exploring models of different depths.
• For the number of neurons per layer (’neurons’), exploration spans 46 different values, ranging from 5 to 50. This search entails a relatively fine granularity, facilitating consideration of a diverse array of neural architectures with differing representation capabilities.

After the training process of the neural network, the regression curves obtained were analyzed, showing the correlation between the targets (x-axis) and the outputs of the system (y-axis). Some examples of regression curves are shown in Figure 5, Figure 6 and Figure 7.

• The network training was conducted with a limited number of cases representing normal operation and fault. To achieve better model adjustment, it is necessary to train the neural network over extended periods under real operating conditions and to simulate abnormal scenarios, such as breakages or collapses, to fine-tune the network’s performance.
• The optimal sample size for achieving the best network performance was identified as 9 s.
• Networks with 5 to 50 neurons per layer were analyzed. It was observed that using more than 10 neurons per layer did not enhance the network’s performance.
• Networks with 1 to 10 layers were evaluated, and generally, better results were obtained with an intermediate number of layers. There was no conclusive evidence that using the minimum or maximum number of layers improved the network’s performance.
• For the different sample sizes, there are one or several combinations of the number of layers and neurons that minimize the training error.

3.4.2. Fault Location Network

Using real data collected from the tests, a neural network was tested to approximate the fault location. The network uses pressure measurements from the sensors to determine signal delays. Based on these pressure drop patterns, the network attempts to identify the approximate section of the system where the fault has occurred.

To achieve this, it was necessary to characterize the locations of the sensors and consumers within the system to properly label the positions of elements. This labeling is essential for training the AI-based fault location system. Due to the limitations of the tests conducted, data were collected only from consumers and faults at specific positions on the ship. The ship was divided into nine areas, with the different tests carried out being labeled accordingly (Figure 8). The division into nine zones was determined based on the ship’s structure and operational requirements. Specifically, the ship was divided into three vertical zones, corresponding to the ship’s three main fire zones. Each of these vertical zones was further subdivided into three horizontal zones. This structured division allows for precise localization of faults and ensures that the neural network can effectively identify and differentiate between different areas of the ship.

The architecture selected for the fault detection network was a classification network. Classification networks are a type of feed-forward network designed to classify inputs into target categories.

The input data for the neural network consist of a series of pressure measurements from each sensor, capturing both time and pressure values as shown in Figure 9. The output is a numerical representation of a specific zone on the ship where the anomaly is detected. The same dataset used for fault detection is utilized for the anomaly localization network, but the preprocessing differs for training purposes. In the detection network, raw pressure values from the sensors are employed directly. Conversely, for the localization network, the data are preprocessed to compute the pressure drops per second.

Different network configurations, including variations in the number of layers and neurons, as well as the sizes of input data samples, were trained to identify the optimal network performance. Additionally, various training functions were employed, such as the scaled conjugate gradient, Levenberg–Marquardt, and Bayesian regularization. In this particular study, only single-layer networks were used, with the number of neurons reaching up to 50.

Different training functions were employed, including scaled conjugate gradient, Levenberg–Marquardt algorithm, and Bayesian regularization. In this case, only single-layer networks were used, and each configuration was initialized 10 times. For each network initialization, 10 retrainings were performed, retaining the layer with the best performance (minimum MSE).

A schematic of the location algorithm proposed is shown in Figure 10.

The model architecture is based on a single-layer artificial neural network, with the number of neurons specified by the parameter ’neurons’. This design choice reflects a simpler configuration where all processing occurs within a single hidden layer.

The exploration for the number of neurons in this single layer ranges from 5 to 50, providing a spectrum of complexity for the model. The search spans 46 different values, allowing for the consideration of various neural architectures and their respective representation capabilities.

The granularity of the search is defined by the number of different values considered for the ’neurons’ parameter, covering 46 values in this case. This fine granularity enables a detailed exploration of the model’s performance across a diverse set of neural network configurations.

This approach prioritizes simplicity by employing a single-layer architecture while maintaining flexibility through the exploration of a broad range of neuron values, ensuring a thorough evaluation of the model’s capacity to capture patterns and relationships within the data.

In this case, we analyzed the receiver operating characteristic (ROC) curves to illustrate the diagnostic ability of the neural network-based fault detection system as its discrimination threshold is varied. The ROC curve is a graphical representation that plots the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings.

• True Positive Rate (TPR): TPR is also known as sensitivity, recall, or probability of detection. TPR measures the proportion of actual positives that are correctly identified by the system. It ranges from 0 to 1 and is plotted on the y-axis.
• False Positive Rate (FPR): FPR is also known as the probability of false alarm. FPR measures the proportion of actual negatives that are incorrectly identified as positives by the system. It ranges from 0 to 1 and is plotted on the x-axis.

The ROC curve provides a comprehensive evaluation of the classifier’s performance across different threshold values. A perfect classifier would have a point in the upper left corner of the ROC space (TPR = 1, FPR = 0), indicating 100% sensitivity and 0% false alarms.

Figure 11, Figure 12 and Figure 13 show the ROC curves obtained from our analysis, demonstrating the effectiveness of the neural network models in distinguishing between normal operation and fault conditions.

• Given the limited amount of available data, the ship was divided into zones, and each consumer opening/pipe break was labeled according to its location within these defined zones. To achieve greater precision in fault detection, it is necessary to acquire a substantial amount of data for training the neural network-based algorithm.
• Several different configurations for the neural network (NN) were defined to determine the best training algorithm and the most effective network structure (number of neurons). The training algorithms used included the scaled conjugate gradient, Levenberg–Marquardt, and Bayesian regularization. Networks with one hidden layer, containing 5, 10, 20, and 50 neurons, were trained.
• The best results were obtained using Bayesian regularization as the training function. The Levenberg–Marquardt algorithm also produced good results, as it did not require a high number of neurons to perform effectively. The poorest results were obtained with the scaled conjugate gradient method.
• It is important to note that the conclusions drawn from these results are based on the premise that the training was conducted with a limited number of training cases, and the identification was performed for relatively large areas of the ship.

4. Discussion

In this work, the development of a method based on the use of neural networks for the detection and location of faults in the FiFi system of a ship has been presented.

This neural network-based system uses the data collected from the sensors of the firefighting system to determine whether the series of pressure data corresponds to normal operation or to the possible rupture of a section of pipe in the system. In addition, the system makes it possible to determine the area of the vessel in which the corresponding breakage has occurred.

To do this, the system has been trained using real data from a ship’s FiFi system. Data of normal operation as well as breaks in the pipes of the system have been collected. These data have been labeled for use in training. The results obtained in the training of neural networks for detection and localization of anomalies have been satisfactory.

Data from pressure sensors in the FiFi system must be collected at a higher sampling rate than normally used for system monitoring. However, it has been demonstrated that modifying the sampling rate is feasible. In the case of use considered, with a sampling frequency above 10 Hz, the results obtained were satisfactory, but a sampling rate above 20 Hz is recommended. After analyzing different alternatives in training, it can be affirmed that it is possible to detect and locate the anomaly with a sample size of 10 s.

The data utilized in this study were processed as received from the ship’s control system, with modifications made only to the sampling frequency and significant figures. Despite potential challenges, such as noisy data, varying sampling rates, or missing values, the inherent robustness of our neural network models allowed for effective fault detection and localization. It is noteworthy that the noise present in the signals did not pose significant issues during the fault detection and localization processes. This illustrates the practicality of our approach under realistic operational conditions and emphasizes the effectiveness of the methodology even with minimal preprocessing.

Future work will focus on expanding the system by continuously collecting onboard data, enabling the system to learn from new fault scenarios and further improve its fault localization capabilities. Additionally, deploying this system on different ships may require retraining of the neural network to accommodate varying ship configurations and sensor setups, ensuring the model’s adaptability to different operational conditions. The integration of hybrid models that combine analytical methods with neural networks will also be explored to enhance fault detection and localization in shipboard systems.

The developed system will be implemented in the ship’s digital twin, providing it with intelligence and allowing autonomous detection of faults. Although this methodology has been validated, employing this development on another ship could require retraining of the neural network. This is due to differences in ship configurations and sensor setups that may necessitate the model to adapt to new conditions and ensure accurate fault detection and localization.

5. Conclusions

This work has successfully demonstrated the development of a neural network-based system for detecting and localizing faults in a ship’s FiFi system. The use of pressure sensor data collected during both normal operation and pipe break events has shown that neural networks can effectively detect anomalies and pinpoint their location within the vessel.

The results of the study highlight the importance of collecting data at higher sampling frequencies, particularly above 10 Hz, with a recommendation to use rates above 20 Hz for optimal performance. The neural network models have been demonstrated to be robust, effectively handling noisy data and varying sampling rates, and can reliably detect faults with minimal preprocessing.

The methodology presented here holds great promise for real-world applications, especially when integrated into a ship’s digital twin for autonomous fault detection. Future work will focus on expanding the system by continuously collecting onboard data, enabling the system to learn from new fault scenarios and further improve its fault localization capabilities.

Furthermore, deploying this system on different ships may require retraining the neural network to accommodate varying ship configurations and sensor setups, ensuring the model’s adaptability to different operational conditions.

In conclusion, this study demonstrates the feasibility and potential of using neural networks for fault detection and localization in ship firefighting systems, contributing to improved safety and operational efficiency in maritime operations.