This paper provides an overview of recent advances in the application of artificial intelligence (AI) in industrial contexts such as data center and the railway industries carried out at the University of Naples Federico II node of the CINI-AIIS Lab. We discuss some challenges and opportunities associated with the adoption of AI in these industries. In data centers, AI is being used to optimize resource utilization, reduce energy consumption, and ensure high availability of services. Despite the potential benefits of AI, there are also challenges associated with its adoption, including the need for high-quality data, reliability and interpretability of AI-based systems, and ethical and legal concerns related to privacy, security, and bias. From the other hand, in the railway industry, AI is being used to optimize train schedules, reduce delays, support predictive maintenance operations and improve passenger safety by predicting and preventing accidents. In this paper, we focus on Hard Disk Drive health status assessment task and on an overview of AI techniques for railway domain in the context H2020 Shift2Rail project.
the application of Artificial Intelligence (AI) in
industrial contexts such as manufacturing, transportation, and
data centers [
Since hard drives often deteriorate gradually rather than abruptly, we argue that temporal analysis methods should be employed to model the sequential nature of the dependencies within SMART attributes over time. Thus, an approach to estimate the Remain Useful Life (RUL) of a HDD has been proposed, by automatically identifying specific health conditions on the basis of SMART attributes values. This methodology is grounded in three main steps: • Hard drive health degree definition : in which status (or health level) is defined for each hard drive according to its time to failure; • Sequences extraction: in which sequences in a specific time window are extracted for each hard drive; • Health Status assessment through LSTM: in which a health level is associated to each temporal sequence.
temporally dependent. Since the selected Regression Tree (RT) presents splits only on the feature Time-to-failure, the latter is used to distinguish hard drive health levels according to the time before failure.
A diferent level or status should be assigned to samples belonging to hard drives that will not fail since they have been excluded in this step. More specifically, the samples belonging to the hard drives that will not fail are labeled as Good.
The degradation of hard drives in real-world data centers 2.2. Sequence extraction
is a gradual process that occurs over time. To account
for this gradual decay, a system has been developed to To explore the temporal dependencies within the SMART
measure the health status or level of a hard drive based features periodically collected for each hard drive,
feaon how much time is left before it fails. In contrast to ture sequences in specific time windows (TW) have been
the approach proposed in [
More specifically, in this step has been considered only posed model aims to predict hard drive health status at the hard drives that are going to fail, introducing for each time + 1 ( ( + 1) ) considering the sequence ( −+1 ..., of them an additional feature representing the time be- −1 , ). For each , the health status () is defined fore failure. The data-set reports, for each hard disk, the as the Regression Tree built , and the feature sequence temporally sorted sequence of SMART attributes with a for each hard drive at time is extracted considering the specific sampling period. Denoting with be the num- − 1 previous samples (cf. Figure 2). Each sequence ber of samples for the hard disk , it is possible to associate results in a bidimensional array of size × , where each sample with an index from 0 to − 1, represent- is the number of SMART features considered. For each ing the number of samples that follow it in the sequence hard drive, sequences are extracted with a stride of one. describing hard disk failure. As a consequence, the sam- It follows that − + 1 sequences are extracted for each ple with index = 0 is the last sample before failure. In hard drive, where is the number of samples for the Figure 1, Time-to-failure is the feature representing the disk j. time before failure for each hard drive whose meaning For each sequence ( −+1 ..., −1 , ), the hard drive’s depends on the sampling period while 1, 2, ..., are health level is defined by the health level of the set of the SMART attributes. features +1 . The result of this step is a sequence-based
The idea is to build a Regression Tree (RT) for each data-set — a set of bidimensional arrays, each associated SMART attribute with = 1, 2... , having the feature to a health level representing the hard drive’s health representing the time before failure as a predictor and condition between two consecutive samples (i.e., and as the numeric target value. Among all the resulting trees +1 ). (one for each SMART attribute ), the one with the highest performance is selected, showing the attribute most
the SMART features, it is important that designed model is able to capture dependencies across features over time. the accuracy of classifying good and failed sequences Long Short Term Memory networks (LSTMs) are exten- for a tolerance of misclassification up to one health level on Multiclass Neural Networks. 2.4.1. Performance Evaluation
presented in Section 2.1, the HDD’s health level assessment as defined in this approach is a multiclass classification problem with multivariate input variables.
The performance of proposed approach is first
evaluated in terms of accuracy, precision, and recall. Since
the distinction between good and failed hard drives is
preserved in the labelling of the data-set, we express the
results in term of accuracy on good sequences (
)
and accuracy on failed sequences (
) — respectively,
the fraction of sequences correctly classified as
Good,
and the fraction of sequence classified as the health
levels suggested by the regression trees. We also consider
the evaluation criteria introduced in [
sion to recurrent neural networks, explicitly designed
with the purpose of learning long-term dependencies [
layer is a three-dimensional data structure of size × × where: is the total number of sequences (or the batch size at each iteration); is the size of each sequence — that size of a time window, in terms of time steps; is the total number of features describing each time step.
compared to the percentage of good hard drives, the sequence extraction step may result in an unbalanced data-set with the majority of sequences belonging to the
introduced, so that the input to the network is a set of balanced data.
In particular, the sequence-based data-set is balanced by replicating the sequences belonging to the minority classes. Sequences replication is an eficient balancing strategy that avoids the polarization of the classification model on a single class without creating synthetic data or reducing the data-set size by sampling the instances beloging to the majority class. The implemented classiifcation network has two stacked LSTM layers with 128 units, followed by a single dense layer.
tested on the two diferent SMART data-set (Baidu and Backblaze)12, and then compared with those of three other methods explored in the literature: a Classification 1http://pan.baidu.com/share/link?shareid=189977&uk=4278294944 2https://www.backblaze.com/b2/hard-drive-test-data.html ( and
).
prediction, by assessing failure detection rate (FDR) and , false alarm rate (FAR) for each model. This is done by considering the levels Good, Very Fair, and Fair as Hard drive good statuses; and the levels Soft Warning , Warning, Alert, and Red Alert as Hard drive failed statuses.
status assessment exploiting the temporal dependencies of SMART attributes. In order to asses the efectiveness of the proposal, this section reports the performance of proposed methodology, and a comparison with several state-of-art approaches.
window (TW) used in the sequence extraction step. We explored time window sizes from 4 to 48 hours for the Baidu data-set, and from 5 to 15 days for the Backblaze data-set. For the latter, we considered a prediction window (q) varying from 15 to 45 days. As expected given the ability of LSTMs to learn long-distance dependencies, the best results are obtained with time windows of 48 hours and 15 days for the Baidu and Backblaze datasets, respectively.
lect data for algorithms’ training and/or validate their efectiveness. Data Augmentation and Transfer Learning techniques were then (especially for the first PoC) used The main objective of the H2020 Shift2Rail (S2R) project to increase the dimension of the dataset and improve AI RAILS (Roadmaps for AI Integration in the raiL Sector) algorithms performances respectively. is to investigate the potential of Artificial Intelligence Concerning Predictive Maintenance and Defect De(AI) techniques in the rail sector to contribute to the tection, attention is being posed to two of the most definition of recommendations and roadmaps for their safety-critical railway assets. Investigations are being implementation in various railway applications with a performed towards non-intrusive, cost-efective, and conspecific emphasis on three main pillars: Safety and Au- tinuous monitoring of Level Crossings (LCs) by exploittomation, Predictive Maintenance and Defect Detection, ing cameras and microphones in combination with Deep and Trafic Planning and Management. To be specific, Learning approaches such as Convolutional Neural Netthe RAILS project has been focusing on six Proofs-of- works and Object Detectors. The scope of this PoC enConcept (PoCs), two for each pillar, intending to explore compasses two main directions: the first is to define a AI techniques to eventually identify their limitations, multi-modular approach that could support the identipossible opportunities they can introduce, and directions ifcation of the health status and the estimation of the that should be taken for the fast take-up of AI in railways. Remaining Useful Life of LC components; the second, on
In the last years, there has been an increasing interest the other hand, is to explore techniques that can help in automatic train driving which led to the delineation to cope with the issue of the lack of available data, inof some promising approaches that, whether based on deed, also in this case, one of the main challenges was to AI or not, could help to migrate towards a higher Grade build suitable datasets for the efective training of AI alof Automation, up to GoA 3/4, of railway lines. One of gorithms and online repositories, real-life simulators, and the main issues that are challenging this migration is the Data Augmentation and Transfer Learning techniques fact that intrusions or obstacles could afect the safety of were leveraged for this purpose. In addition to that, analysuch autonomous systems. To mitigate this factor, also ses are being carried out to define the contribution AI can under the S2R programme, investigations have already bring to continuous monitoring and predictive maintebeen performed towards the realisation of multi-source, nance of rolling stock components. Lastly, in the context multi-technology, and multi-level systems (exploiting, of Trafic Planning and Management, the project is fofor example, multiple sensors, drones, and components cusing on modelling and predicting train delays, a topic installed on-board the train and on the trackside) which that is of primary importance when it comes to undercould help to improve the environmental perception of standing the possible congestion of rail networks and, running trains. However, these systems are as efective potentially, taking actions to mitigate this phenomenon. as they are complex and potentially expensive. There- In this context, eforts have been posed on primary defore, within the RAILS project, eforts have been oriented lays (i.e., delays generated by accidents/intrusions on rail towards the investigation of a “light” and cost-efective tracks) prediction and estimation by means of a Graph solution that could potentially help to deal with the detec- Embedding Approach, while Graph Neural Networks tion of any kind of obstacles (both known and unknown were investigated to understand how accidents could a-priori) laying on rail tracks by exploiting of data coming trigger minor disturbances that, in turn, cause delays from a camera in combination with supervised and unsu- which propagate throughout the railway network and pervised AI-based Computer Vision models (specifically, will cause, in the end, reactionary delays (i.e., knock-on Deep Autoencoders) to segment rails and, eventually, de- or secondary delays). tect possible obstacles. In addition to that, another, quite To conclude, it is important to underline that the priinnovative, PoC is being addressed which deals with mary purpose of the aforementioned PoCs was to gather the integration of AI techniques for the introduction of knowledge and evidence about the possible integration Virtual Coupling (VC), widely examined within the auto- of AI in railways by exploring diferent techniques that motive sector, in the context of autonomous trains. The could help to overcome possible issues (e.g., the probmain idea is to investigate the possibility of allowing two lem of the lack of data). Hence, the final goal of the or multiple trains to act as a single convoy by exploiting project would be to contribute to the definition of valuTrain-to-Train communication network, to reduce the able recommendations and roadmaps, to be disclosed in headway between them, and a Reinforcement Learning the near future, that could support future research on AI approach to implement the VC control strategy. Interest- in railways. ingly, since one of the main challenges in the rail sector is the lack of suitable datasets to train AI algorithms, for these PoCs two simulators were built in order to col