1   Introduction

Accurate and robust detection of wax appearance is crucial for efficient hydrocarbon transportation in the petroleum industry. Wax is a medium- to high-molecular-weight component of crude oil (Bian et al., 2019). The wax appearance temperature (WAT), also known as the cloud point, is the lowest temperature at which the smallest amount of wax forms (Coutinho and Daridon, 2005). When crude oil (i.e., waxy crude) cools to its WAT, wax crystals precipitate out of solution, forming deposits on the pipeline as the solubility limit is reached. This process poses a notable challenge in hydrocarbon transportation, as wax deposits can negatively affect fuel flow (Ahmed, 2007). The WAT is a critical quality assurance parameter, particularly when for modern fuel oil blends. These oils are blended to comply with strict International Maritime Organization (Li et al., 2020) regulations, and such blending operation usually produces fuel oil with high paraffinic content and increased risk of wax deposition. When the WAT values are known, vessel operators can safely store fuel oils above this temperature (VPO, 2019). Wax deposition is also a major cause of stuck pigs and increased pressure drops in pipelines and hydrocarbon processing (Mansourpoor et al., 2019). Re-establishing flow in pipelines where crude oil or fuel has solidified below the WAT is expensive because it presents a flow assurance challenge (de Oliveira et al., 2012). Therefore, the efficient and safe operation of crude oil and hydrocarbon pipelines relies on the accurate determination of the WAT (Zhao et al., 2022).

Measuring and detecting WAT is a challenging task influenced by procedural sensitivities and uncertainties (Kok et al., 1996). Factors like cooling rates affect results, with faster cooling rates often leading to underestimated WAT values (Japper-Jaafar et al., 2016). Many known procedures detect wax in crude oil by measuring the size or the number of wax crystals. Some studies have also proposed laser-induced voltage technology for WAT detection (Zhang et al., 2022). The detection of light scattered by wax crystals has been used to measure WAT values in very low sulfur fuel oils, as specified by ASTM D5773 (Thomas, 2019).

Some methods have determined WAT by microscopically observing the onset of wax crystal formation. However, this approach is questionable, as it relies on a visual assessment of the onset of wax buildup, which might be misleading—it requires a substantial amount of wax buildup to be detectable (Kok et al., 1996). Other techniques have focused on detecting wax at an earlier stage of formation, but they failed to detect the true WAT, particularly under thermal equilibrium conditions. All other known techniques, including the “Cold finger method” and the use of a viscometer to detect the shift from the linear viscosity-temperature relationship of a Newtonian to a non-Newtonian fluid behavior, have proven ineffective for WAT detection (Speight, 2016). Laboratory-based techniques such as differential scanning calorimetry (Kök et al., 2018) and cross-polar microscopy (Taheri-Shakib et al., 2020) are highly time-consuming, causing delays of weeks or months between sample collection and analysis date. These delays may introduce sample disparities (Al Shakhs et al., 2023). Other methods, such as Karl Fischer titrations and near-infrared (NIR) spectroscopy, are also used for WAT measurement (Zhang et al., 2022). The majority of the methods used for WAT detection produce inconsistent results (Japper-Jaafar et al., 2016) and are time-intensive, as many require cooling fuel oil or crude oil at different rates (Uba et al., 2004). In addition, the ASTM D3117 method is limited to specific crude or fuel oil types. This method struggles with darker or opaque oils, as it relies on visual confirmation of wax formation (Al Shakhs et al., 2023).

Although machine learning algorithms have also been successfully applied to predict the WAT of crude oil based on parameters such as density, freezing point, and wax content (Benamara et al., 2020), this study presents the use of the one-class support vector machine (OCSVM) algorithm to predict marine fuel samples with abnormal WAT values. This method will reduce the dependence on time-consuming and uncertain laboratory analytical methods. Notably, this is the first attempt by the data supplier Intertek PLC—a total quality assurance, testing, and certification company—to utilize a machine learning algorithm to predict marine fuel oil samples (rather than their exact WAT values) with unusual WATs. This method provides an opportunity for rapid identification and segregation of high-WAT samples, even with scarce data. This process will be carried out by correlating spectral data with the assay data from laboratory-measured WAT values.

Although ISO 8217 specifies a maximum allowable WAT of −16 ℃ for distillate fuel oils to be used in marine vessels (ExxonMobil, 2022), this study focuses on a threshold of 35 ℃ based on the operational requirements of Intertek PLC and the vessel operators involved. Different crude sources show varying WAT values based on their origin (Kok et al., 1996).

The real-world industrial NIR spectral data provided by Intertek PLC are highly imbalanced in favor of samples exhibiting higher WAT values (>35 ℃) and a lack of data for samples with lower WAT values (<35 ℃). This imbalance makes standard binary classification algorithms ineffective, necessitating the use of an anomaly detection algorithm.

This study aims to showcase the OCSVM algorithm and its ability to classify and predict marine fuel samples based on spectral data from an imbalanced dataset (Cohen et al., 2004).

The remainder of the paper will focus on describing the OCSVM algorithm and the methodology used in this study.

2   One-class support vector machine (OCSVM)

The OCSVM algorithm was developed by Schiilkop et al. (1995), and it is an extension of the SVM algorithm. OCSVM is based on the study that showcased the possibility of building an optimal hyperplane in high-dimensional feature space by maximizing the margin between the origin and the hyperplane (Schölkopf et al., 2001) (Figure 1). Where support vectors refer to the data points close to the separating hyperplane. training samples are samples used to train the model, K (x, x_i) is the kernel function, the outliers are samples that are dissimilar to the trained class samples. f < 0 is the negative hyperplane. f > 0 shows the positive hyperplane, while f = 0 is the optimal hyperplane. $ \frac{\boldsymbol{p}}{\|\boldsymbol{w}\|} $ represents the distance from the origin to the optimal/decision hyperplane (Fletcher, 2009).

Figure  1  Mapping data onto a higher dimensional space to enable a linear separation in that feature  

Download: Full-Size Img

OCSVM is designed to detect anomalous behaviors in processes where data availability for such abnormalities is complicated. This explains why OCSVM is used in novelty and fault detection since many processes store data generated with optimal conditions, resulting in data scarcity for suboptimal situations (Xiao et al., 2015).

As a semi supervised technique, there is no requirement for all the data to be labelled, which mitigates the problems experienced with supervised algorithms, making OCSVM attractive in many industries, including in clinical applications (Lang et al., 2020). To detect nosocomial infections in healthcare environments, a true positive rate (TPR) of 92.6% was achieved with OCSVM, in contrast to a TPR of 50.6% achieved with a regular SVM for binary classification. This was a remarkable result because the available data were greatly imbalanced, providing only 11% of abnormal (infected) cases and 89% of normal cases (Cohen et al., 2004). In the processing industry, OCSVM was useful in the detection of security breaches in a water treatment plant, using data collected during normal operations from subprocesses as single-class training data (Yau et al., 2020). In the oil and gas exploration and production industry, OCSVM was applied in combination with a segmentation algorithm to detect abnormal sensor information in turbomachines in offshore oil installations, enabling effective decision-making by stakeholders. This was seen as a resounding success because of the absence of labelled data generated from the process, which made it difficult to employ a standard supervised learning algorithm (Martí et al., 2015). OCSVM builds and trains the algorithm with the available data from a single class and proceeds to detect samples (or outliers) that are dissimilar to the trained class samples with the aid of a decision boundary. The decision boundary (or hyperplane) separates the outliers from these training samples (Seo, 2007). The hyperplane maximizes the distance from the origin to the samples, with the samples lying far away on the opposite side of the origin.

The margin maximization is achieved by solving the primary optimization (or quadratic programming) problem in the following equation (Seo, 2007), where the origin is positive (Xiao et al., 2015) and a negative label represents the second class as follows:

where parameter v is defined by the user. It explains the upper bound of the training errors or outliers and the lower bound of the fraction of support vectors. This means that it controls the trade-off between maximizing the distance between the hyperplane and the origin and incorporating most of the data points inside the hyperplane region. The offset is defined by ρ, and w is the weight of the vectors. The transformed image of x_i in the Euclidean space is depicted by $ \boldsymbol{\phi}\left(\boldsymbol{x}_i\right) $, where i ∈[l] and the non-zero slack variable is ξ (Long and Buyukozturk, 2014).

As established earlier, OCSVM aims at classifying data objects based on their similarity to the trained single class. Objects are rejected and classified as outliers if they are dissimilar to the trained class.

For linearly separable data, the hyperplane’s distance is $ \left(\boldsymbol{w} \cdot \boldsymbol{\phi}\left(\boldsymbol{x}_i\right)\right) \geqslant \boldsymbol{\rho} $ and ρ > 0.

The result of the solution of min $ \min \limits_{\boldsymbol{w} \in \boldsymbol{F}} \frac{1}{2}\|\boldsymbol{w}\| \text { with }\left(\boldsymbol{w} \cdot \boldsymbol{\phi}\left(\boldsymbol{x}_i\right)\right) \geqslant \boldsymbol{\rho} $ for every i is the distinct hyperplane closer to the origin than all other data points with a maximum distance from the origin to any other hyperplane.

In reality, not all data can be easily divided linearly into positive and negative cases; therefore, a certain error in classification is allowed (Seo, 2007). To solve the linearly inseparable problem, as earlier shown, a kernel is used to map the data onto a higher dimensional space to enable a linear separation in that feature space. Different types of kernels are applicable in OCSVM; however, the Gaussian radial basis function (RBF) is the most commonly used kernel (Cohen et al., 2004), with an ability to handle nonlinear problems, leading to a classifier with a universal optimal predictor with lower estimation errors (Martí et al., 2015). This paper uses the Gaussian RBF kernel.

A kernel function k (x, y) for a feature vector in a low dimension x and y is expressed as follows (Long and Buyukozturk, 2014):

The RBF kernel is expressed as follows (Long and Buyukozturk, 2014):

In Equation (3), σ is the optimizable Gaussian distribution hyperparameter. The model’s performance hugely depends on the accurate selection of the kernel parameter (Xiao et al., 2015). A larger value of the Gaussian parameter translates to a narrow distribution with a more concentrated sample distribution. Conversely, a smaller value of the parameter depicts a broader Gaussian distribution with a denser sample distribution (Nie et al., 2022), thereby overfitting the model and significantly reducing the generalization ability of the OCSVM model, whereas a larger value of σ will under-fit the model, which makes it unable to find even common patterns within the data. The other optimizable parameter v, as earlier stated, determines the number of training samples that can be classified as outliers. A smaller value means a lower number of false positives (i.e., low WAT samples predicted as high WAT samples), which is a preferred option. Both parameters can be selected or tuned with the aid of a cross-validation technique if samples of the minority class exist (Long and Buyukozturk, 2014).

To classify a new marine fuel sample data point x, Lagrange multipliers α_i and the kernel trick are employed to solve the dual problem as follows (Long and Buyukozturk, 2014):

Using the Gaussian equation for mapping onto a feature space the following can be obtained (Long and Buyukozturk, 2014):

where ρ can be retrieved as follows (Long and Buyukozturk, 2014):

The support vectors (i.e., all nonzero Lagrangian multipliers α_i) are utilized in solving the decision function for the new sample or data point x among samples N.

3   Support vector machines

SVM, as proposed by Cortes and Vapnik (1995), was originally intended to be used as a two-class linear classification tool, but it has now been expanded to be applied in nonlinearly separable classification with the use of the kernel trick. The kernel trick is used to map nonlinearly separable data to a higher dimension to enable a class separation (Zhang, 2001; Li et al., 2023), as is the case in this paper. SVMs classify two classes in a dataset with the use of a decision hyperplane. Data objects are plotted as points on a feature space, and each of the features in the data is denoted by a coordinate. To classify the data, the plotted coordinates are divided with a hyperplane to separate the different groups. The coordinates closest to the hyperplane are referred to as the support vectors (Thijssen and Hadjiloucas, 2020). A new and unseen data point is classified and predicted based on the group to which it is closer on either side of the hyperplane (Bangert, 2021).

For linear classification, mathematically, the hyperplane is described as wx + b = 0, where w is normal to the hyper plane and $ \frac{\boldsymbol{b}}{\|\boldsymbol{w}\|} $ describes the distance, which is perpendicu lar to the hyperplane from the origin (Fletcher, 2009). In Figure 2, if the positive hyperplane is expressed as wx + b = 1 and the negative hyperplane is expressed as wx + b = − 1, the optimal hyperplane is mathematically shown to maximize the margin, which is the perpendicular distance between the hyperplanes closest to the optimal hyperplane and the optimal hyperplane on both sides, assuming that sample data points do not fall in between the positive and negative hyperplanes. A detailed mathematical illustration of the maximization of the margin is shown in the work by Liu (2020a).

Figure  2  SVM margin and hyperplane

Download: Full-Size Img

For a sample of marine fuel data with a vector of wave numbers x′, the SVM algorithm is applied for classification and subsequent prediction, as described in the following equation (Liu, 2020b):

where $ \left\|\boldsymbol{w x}^{\prime}+\boldsymbol{b}\right\| $ represents the distance from x′ to the optimal/decision hyperplane. Here, y is equal to 1 if the distance between the sample and the hyperplane is greater than 0 and −1 if the distance is less than zero and will be assigned to the negative side of the hyperplane, as shown in Figure 2. In the case of marine fuel sample prediction, y is equal to 1 (i.e., the positive case or high WAT) if the WAT of the fuel is 35 ℃ or higher and −1 (i.e., the negative case or low WAT) when it is less than 35 ℃ by weight. A large distance between the data point and the decision boundary translates to higher confidence in the prediction. This is interpreted as a higher prediction certainty because the sample is further away from the decision boundary. Parameter C (Fletcher, 2009) is the key to a decision of how much misclassification is acceptable for a given problem, as follows:

It influences the compromise that can be made between the margin size and the penalty variable ξ. An effective SVM classification requires the optimization of parameter C. This involves the careful selection of parameter C to control how the misclassifications should be penalized (Fletcher, 2009).

4   Quality of the model

There is a possibility of errors occurring in the class assignment, which can lead to the misclassification of some of the samples. It is, therefore, important to assess the performance and reliability of a model. A classification error indicates the possibility of a flawed classification when the model is applied to unknown samples in the future (Kuligowski et al., 2016).

Although accuracy has been the common performance metric adopted by many researchers, it does not present a true picture of the classifier’s accuracy when dealing with unbalanced data. As an illustration, a high classification or predictive accuracy of 95% for unbalanced data provides no information about the misclassified minority class, but it has only provided a piece of information to show that almost all the samples have been assigned to the majority class with 95% accuracy (Ding, 2011).

Most classification algorithms assume that all datasets are evenly distributed with an almost equal number of samples for every group in the data, but unfortunately, this is far from being true. Most real-life datasets are unbalanced. An unbalanced dataset has a skewed distribution, with one class surpassing the other. The classification of such data has been a major hindrance to many researchers because the majority of classifiers favor the majority class, which is sometimes referred to as the negative class, to the detriment of the minority or positive class (Ding, 2011).

It is common knowledge that the validity of classification models is vital for the subsequent prediction of unknown samples when introduced into supervised algorithms, including SVM. The samples can be assigned to the appropriate groups or classes, but there is a possibility of misclassification. The performance of the model is measured in terms of TPR or Recall (Westerhuis et al., 2008). The OCSVM algorithm also outputs anomaly scores, which are indicators of samples seen as abnormal to the training samples based on the threshold, in this case, zero. These are seen as anomalies (MathWorks, 2023). In addition to the recall or TPR (Sensitivity), Accuracy and Precision will be used to assess the performance of OCSVM for WAT. It is common to assess OCSVM models based on the percentage of the true anomalous samples detected when compared to all anomalous samples in the dataset (Precision). As shown in the following equation (Nguyen et al., 2023), the model’s accuracy compares the predicted and original classes (Ding, 2011):

In this study, the focus is on Sensitivity, which is the correct prediction of positives (i. e., samples with WAT values above 35 ℃). The TPR represents the probability that samples are correctly identified as part of the high WAT samples (or positive class) used to train the model (Broadhurst and Kell, 2006) and labeled correctly (Bekkar et al., 2013).

The following equation (Sun, 2009) describes the performance metrics as they relate to the sensitivity of the model:

5   Methodology

This section focuses on the details of the data and the analysis carried out in this study. The experimental methodology used in this paper is depicted by the flowchart (as shown in Figure 3), describing the data cleaning, preprocessing, and modeling steps used with OCSVM and SVM algorithms.

Figure  3  Marine fuel OCSVM and SVM flowchart

Download: Full-Size Img

5.1   Data collection and analysis

The spectral data were generated by performing 32 scans on each sample of the marine fuel using the ABB MB3000 Series FTIR Spectrometer. This was chosen after several trials, and it has proven to deliver a stable result when performed five times on each sample at a resolution of 4 cm⁻¹ to record the absorbance of a typical marine fuel sample covering the wavenumber range of 4 000‒4 800 cm⁻¹. This range is associated with the fuel’s absorption in the combination region range (Blanco and Villarroya, 2002), covering the region for hydrocarbon absorbance (Lammoglia and de Souza Filho, 2011) in line with Intertek PLC’s specific operational requirements.

MATLAB Bioinformatics Toolbox was utilized to read and clean the marine fuel spectral data (which are available in Intertek PLC’s data repository). Missing values, invalid signals, and redundant values were also removed before computing the average of the five signals to ensure all signals for each sample were represented in the model, bringing the number of valid samples to 368 with 186 wavenumbers.

Figure 4 shows the unbalanced data distribution of the laboratory-measured WAT values, which are used to determine the output variables for this model, which showcases marine fuel samples with high or low WAT values. The distribution reveals 10%–30% of the data (i.e., low WAT samples below 35 ℃) as the minority class, necessitating a one-class classification, hence the use of OCSVM for anomaly detection. The OCSVM model is designed to detect anomalous behaviors in processes where data availability for such abnormalities poses a challenge (Xiao et al., 2015), as is the case with WAT. More than 80% of the samples belong to the high WAT group based on the threshold of 35 ℃. In line with the specific operational requirements for Intertek PLC’s customer, all marine fuels are stored above 35 ℃ to prevent wax precipitation and subsequent deposition on the lines.

Figure  4  Marine fuel WAT data distribution

Download: Full-Size Img

5.2   Data preprocessing and model building

MATLAB’s Statistics and Machine Learning Toolbox was used to normalize the marine fuel data with the area under the curve set to a unit (Ciaburro and Joshi, 2019) before scaling the same by the standard deviation, resulting in a mean of zero and a standard deviation of 1 (MathWorks, 2022b).

NIR signals suffer from baseline shifts that appear on the absorption axis as a result of additive effects or offsets. These baseline shifts could be a result of several factors, including particle size, porosity, air bubbles, and general morphology attributes. They can also occur as wavelength-based light scattering effects on the sample (Huang et al., 2010), which may appear like absorption but are not related to the chemical composition of the samples and, as such, not relevant in subsequent analyses of the sample (Sandak et al., 2016). These were mitigated using baseline shift techniques.

To ensure the accuracy and robustness of the model, the marine fuel data were normalized, as stated earlier, with the area under the curve set to 1 after carrying out a baseline shift at 4 780 cm⁻¹ to reduce the influence of the signal shifts on the normalized data, restoring all the signal peaks to a common baseline (Fanali et al., 2017). This data was further standardized by subtracting the mean from each sample vector before scaling it by its standard deviation. It is common to normalize the data by setting the area under the curve to 1 in such a way that all the signals are modified and summed to 1. This will create more uniform data with a common baseline and equal contribution to the model (Ciaburro and Joshi, 2019). This practice is recommended for SVM implementation (Meyer et al., 2003) and is in line with Intertek PLC’s standard operating procedure.

The OCSVM model was built with MATLAB Statistics and Machine Learning Toolbox (MathWorks, 2022b) for OCSVM using the RBF kernel to implement the quadratic programming algorithm. Hyperparameter tuning was enabled to optimize the nu (v) parameter. This was carried out with nu (v) values ranging from 0.1 to 0.8, with 0.5 proving to be the optimum value because it is low enough to prevent overfitting the models and still capture the data complexity (Chen et al., 2001). For internal validation, a five-fold cross-validation (MathWorks, 2022a) technique was implemented in MATLAB. For external validation purposes, 90% (i.e., 274 samples) of the samples with high WAT values were used to train/adjust the model’s parameters, whereas the remaining 10% (i.e., 31 samples) was used for validation (or test data) of the model. These test data include 14 low WAT samples, which were used to “contaminate” the model to test its sensitivity. The training data labels were also predicted using the trained classifier as detailed in (MathWorks, 2024).

90% of the data were used because the OCSVM model requires more training samples to compensate for the positive sample instances that may be rejected as a result of the hard thresholding adopted by the model. Samples are either accepted or rejected as marine fuel samples with an abnormal WAT value as a result of the OCSVM hard thresholding (Guerbai et al., 2014). MATLAB’s OCSVM toolbox computes and assigns anomaly scores to each sample according to their similarities to the training samples based on a threshold (i.e., zero). The scores were computed for the training data and compared with the scores for the test data. The scores below 0 are assigned to the samples that are dissimilar to the samples used to train the model and are therefore rejected as anomalies.

For comparison purposes, the same data were also normalized and trained with a standard binary SVM classification algorithm employing the same fivefold cross-validation approach. 90% of the spectral data supplied by Intertek PLC was used to train the model, whereas 10% was used for validation. Similarly, MATLAB Statistics and Machine Learning Toolbox (MathWorks, 2022b) was used to build the model. An RBF kernel scale of 0.002 and a Box Constraint of 0.318 were adopted using Bayesian optimization for hyperparameter tuning, implementing the quadratic programming algorithm for objective function minimization. The “Box Constraint” here is the terminology used in MATLAB to represent the C parameter for the maximization of the margin separation between the support vectors and the hyperplane to minimize the misclassification error of high WAT samples and low WAT samples in the marine fuel dataset (MathWorks, 2022b). A total of 149 support vectors were used to influence the separation of high and low WAT classes.

6   Results and discussion

The proposed OCSVM model was trained using marine fuel samples with WAT values above 35 ℃, which accounted for >80% of the total available data. MATLAB assigned anomaly scores to each sample in the training dataset, where positive scores depicted similarity to the training data (WAT >35 ℃), whereas negative scores identified outliers (WAT <35 ℃).

Figures 5 and 6 show the accuracy and sensitivity plotted against various nu values, respectively, which represent different outlier fractions for each nu value. The results show that accuracy and sensitivity improved as the fraction of outliers detected increased. This observation confirmed that the samples with scores below the threshold of zero were correctly identified as anomalies by the OCSVM classifier and were treated as outliers. Figures 5 and 6 show that the proposed model detected 4% outliers, corresponding to 14 dissimilar samples, though fewer than one sample per test instance prevented clear visualization.

Figure  5  Accuracy, outlier fractions, and nu values

Download: Full-Size Img

Figure  6  Sensitivity, outlier fractions, and nu values

Download: Full-Size Img

A TPR of 96 indicates that 96% of marine fuel samples with WAT >35 ℃ were accurately classified. The model achieved an accuracy of 0.94 and a precision of 0.99, enabling informed decision-making regarding fuel stability for logistics and storage. These results highlight OCSVM as a high-performing anomaly detection algorithm for marine fuel spectral data. Table 1 presents the original and predicted labels for the test (or validation) data, demonstrating that the model classified all the positive cases. No false negatives (i.e., high-WAT samples misclassified as low-WAT) were recorded.

All fuels predicted to have WAT >35 ℃ will be stored separately and will not be taken onboard vessels. The results further showcased the validity of OCSVM over standard binary SVM when classifying and predicting with very scarce data. The standard binary SVM model recorded a TPR of 85.7% for the majority class but failed to classify any samples in the minority class. The poor predictive performance of regular two-class SVM was highlighted by a low area under the receiver operating characteristic (ROC) curve (AUROC) of 0.54, as shown in Figure 7. This observation indicates random classification, with ~50% false positives and 50% true positives. This prediction will lead to extremely inefficient decision-making by vessel operators and potential reputational and maintenance costs.

Figure  7  ROC curve of binary SVM on WAT

Download: Full-Size Img

Table 1 shows that the algorithm identified all the samples (high WAT) that are similar to the samples of the class used to train the model. This means that all the positive cases (samples with high WAT values) were successfully predicted by the model. The lower WAT classes were not predicted and there were no false negatives (high WAT samples predicted as low WAT samples) recorded.

7   Conclusions

Herein, we demonstrated the effectiveness of the application of the OCSVM algorithm in predicting suboptimal marine fuel samples using spectral data and laboratory-measured WAT values provided by Intertek PLC. Compared with the standard binary classification algorithm, such as SVM, OCSVM enables the rapid detection of fuel samples with abnormal WAT values, reducing reliance on irregular laboratory test methods.

OCSVM achieved a TPR of 96 with no false negatives using a five-fold stratified cross-validation approach in MATLAB to avoid model overfitting. This capability supports fast and well-informed decision-making for logistics and storage, even when choosing fuel oils on board the ocean-going vessels, despite the absence of sufficient data on LOW-WAT samples required for standard binary classification. In contrast, the regular SVM model yielded a TPR of 85.7, which was confirmed as a chance classification based on the ROC curve. OCSVM classified 4% of samples with WAT of <35 ℃ as outliers, indicating their dissimilarity from the training data. Although the results are promising, the proposed technique can be improved to detect more LOW-WAT samples (<35 ℃). Overall, using OCSVM is recommended, particularly when dealing with data scarcity, as is the case with Intertek PLC’s dataset.