Davari et al. [1] proposed an anomaly detection framework in air production units of railway industries using deep learning. The work proposes a sparse autoencoder network that, by aggregating data from various sensors, detects failures in advance to increase the efficiency of maintenance and reduce downtime in the operation of trains at Metro do Porto. Wang et al. [2] developed an enhanced LSTM model for time-series anomaly detection tailored to rail transit systems. It overcomes the challenges of diversified data distributions and the lack of labeled anomaly data facing rail transit operations through the employment of IoT and cloud technologies. The proposed model, in addition, defines thresholds on a dynamic basis and includes an error-pruning algorithm for the least number of false positives. This approach has been tested in a real metro environment, where it gave an even better model of performance than existing ones in a proper identification of device-related anomaly detection for different rail transits.
Davari et al. [3] explore the integration of ML and DL technologies into predictive maintenance in the railway sector. This review identifies ML and DL techniques as pivotal in advancing the analysis of industrial equipment for predictive maintenance purposes. Despite significant strides, the field faces challenges in data handling, method selection, and application performance. The authors highlight the growing research attention towards data-driven PdM in railways, emphasizing its critical role in enhancing system efficiency, safety, and cost-effectiveness. Veloso et al. [4] introduced the MetroPT dataset, a comprehensive collection of data from Porto’s metro public transportation service, designed to support the analysis of AI methods for failure prediction. This dataset encompasses a set of analog sensor readings (e.g., temperature, pressure), digital readings (control and discrete signals), and location data, offering a rich resource for predictive maintenance (PdM) research. Aimed at reducing operational disruptions and transitioning from reactive to predictive maintenance approaches, the MetroPT dataset presents real-world anomalies confirmed by maintenance reports, making it an ideal benchmark for PdM models.
Zhao et al. [5] suggested a hybrid model that integrates time series decomposition with ”deep learning for short-term passenger flow prediction in urban railway systems”. By STL categorizing passenger flow into seasonal, trend, and residual components, the study enhances prediction accuracy by addressing varied characteristics of passenger flows. The methodology combines the Holt-Winters (HW) method for predictable components and LSTM networks for residuals, demonstrating superior prediction performance over traditional models. Li et al. [6] explored deep learning applications in failure detection, categorizing anomalies into abnormal time points, intervals, and series. They highlight Long Short-Term Memory (LSTM) and autoencoders as prevalent techniques for identifying abnormal points and intervals, with dynamic graphs employed to reveal relational features among time series. Despite advancements, the field grapples with explainability limitations, noting a gap in understanding the underlying causes of anomalies. Addressing this gap, enhancing anomaly explainability emerges as a critical future research direction.
Berroukham et al. [7] talk about ”deep learning-based methods for anomaly detection in video surveillance”. It highlights the improvement from selected features to deep learning models, emphasizing their efficacy in learning complex representations of normal and abnormal events. The paper surveys various datasets, evaluates methods through standard metrics like AUC and EER, and discusses challenges like explainability and real-time processing. It suggests potential research directions, including the application of vision transformers for improving anomaly detection accuracy and computational efficiency. Alamr et al. [8] introduced a transformer-based architecture for failure detection in ECG signals, showcasing its effectiveness in identifying irregular patterns within electrocardiogram data. The model employs a transformer encoder to analyze time series ECG data from two prominent datasets, ”ECG5000 and MIT-BIH Arrhythmia”, achieving notable results. We substantiate that the proposed approach exhibits improved performance in considered performance metrics over several state-of-the-art deep learning models and generalizes anomaly detection across healthcare data, holding the potential to be a strong tool for the early detection of cardiac irregularities.
Luo et al. [9] investigate data-supported decision help in the control of rail traffic and find the place of the predictive model in improving the reliability of the transportation system. Zeng et al. confirm only how burdensome it is to deal with rail traffic and what potential the predictive analysis has in its turn to polish the operational strategies. It proposes a new hybrid prediction model that uses optimized synthetic minority oversampling techniques with a deep forest ensemble learning model. It was tested against actual high-speed railway data and found to improve the decision-making capability of rail traffic controllers. It offers a vast improvement in controlling the dynamic rail traffic conditions. Girish et al. [10] proposed an AI-based model using a Stjson file that achieves over 93 percent accuracy in the OpenStack cloud environment for it to be capable of detecting anomalies. In this approach, they significantly improved upon the conventional manual way of using advanced machine-learning techniques for the preemptive detection of anomalies.
Among the big contributors to the security of the IoT, Abusitta et al. [11] came up with a new deep learning model for the detection of anomalous behavior which performs well in light of the diversity of data and disturbances to the system. They used a denoising autoencoder so that only clean, non-malicious data populates the IoT systems. The remarkable superior performance of our model in real-life IoT datasets testifies very truly that our model has robust feature extraction capability, setting new standards in this domain for IoT-based anomaly detection. Cerd`a-Alabern et al. [12] focus on machine learning in fault detection for wireless community networks. Its actual area of research is not that vast, since there are no shared datasets available. Their study shows such potential of ML by implementing four unsupervised machine learning techniques on real network data. This result, therefore, recommends feature selection and model tuning to show clear performance differences between the ML approaches used.
Pan et al. [13] address anomaly detection in SHM for civil infrastructures through transfer learning-based methods. Their approach works by exploiting pattern similarity that occurs across several different bridges, dealing with sensor faults and environmental noise. Their approach outperforms in classification capability, proven to outperform in the assigned task of new bridges having the smallest target data by building up a multivariate database and using a CNN for the classification of an anomaly. The results of two long-span bridges show that transfer learning is effective in improving detection performance with very few labeled data, many better than classical methods. In one of the first works on the topic, Sousa Tom´e et al. [14] developed an online, data-driven predictive maintenance framework for railway switches, an essential element within the railway infrastructure. Their approach uses real-time data logging from the railway’s interlocking system to predict maintenance needs, aiming to improve maintenance efficiency and railway switch reliability. The methodology, which includes anomaly detection and remaining useful life prediction phases, was tested using seven months of data from the Metro do Porto. This study represents a significant stride towards optimizing railway switch operations through advanced data-driven methods.