We present a smart monitoring system that combines a data management architecture with an unsupervised anomaly detection technique, targeting the automated equipment in the entertainment industry. Anomaly detection uses state-of-the-art univariate and multivariate algorithms, as well as recently proposed techniques in the field of explainable artificial intelligence, to achieve enhanced monitoring capabilities and optimize service operations. The monitoring system is here presented and tested on a real-world case study, i.e., an amusement park ride.
Antonio Zamperla S.p.A.1 designs and develops entertainment rides, continuously striving for their rides to become safer, greener and more eficient. Smart monitoring systems are expected to enrich the entertainment industry with a number of key features. For example, faults could be predicted in advance, or data analysis during the testing of the ride’s prototypes could give deeper insights onto various design choices. Furthermore, maintenance operations are nowadays performed manually, following ride’s manuals or government directives that are usually scheduled on a periodic basis, regardless of the actual conditions of the machines. This implies that, oftentimes, maintenance is performed without a real need, entailing a waste of time, human resources and material. This practice, although being cautious and robust, is quite ineficient. Smart monitoring systems would allow a change of paradigm from the current conservative approach to a greener and more eficient one. Lastly, automated supervision techniques would enforce the safeness of the rides by detecting subtle anomalies, which would be hardly identified by a human supervisor.
In the context of advanced monitoring of complex systems, two main Machine Learning
(ML)-based technologies have emerged in recent years: unsupervised anomaly detection (AD) [
In this work, we present a data management architecture that combines state-of-the-art univariate and multivariate approaches for AD to reach enhanced monitoring capabilities and optimize service operation. We use a database to store the streaming data acquired by sensors, feed them to the AD algorithms, store back the features extracted by the AD algorithms, relate them to alerts and maintenance events, and support an overall Web application.
The paper is organized as follows: Section 2 briefly summarizes related work; Section 3 presents the considered use case; Section 4 describes the Entity-Relationship schema of the datastore; Section 5 introduces the proposed approach for AD; Section 6 discusses the current AD prototype and the related Web application. Finally, concluding remarks and future works are reported in Section 7.
Unsupervised AD tools adopt (i) multivariate approaches based on tabular data and (ii)
univariate approaches working with time-series [
The case study is represented by a coaster (Zamperla’s DangleZ ) whose seats can freely move during the ride, independently of the underlying chassis, see Figure 1. In this case study, the coaster track is 133 meters long.
The data acquisition process provides, by means of an on-board PLC, a set of = 55 diferent time series acquired with irregular sampling rate. Each time series represents a signal which can be analyzed to monitor the behavior of the machine: 9 signals are generated by the equipment sensors that report the transit of the coaster over diferent locations of the rail; 15 signals describe voltages, currents, frequencies and other physical quantities denoting the consumption and the movement of the machine; 5 signals are acquired by a weather station that detects the condition of the surrounding environment; the remaining signals are needed to check the correct working conditions and the security of the machine (for instance, the state of the safety button).
Data are divided into sessions (or tests) and each session is composed of a number of cycles. Each cycle collects the data generated during one run of the coaster, from the moment it leaves the station to its return to the station. Each session contains data coming from an homogeneous set of measurements, typically measurements in the same day or under the same wheater conditions.
• analysis related data: dictionaries (the obtained prototypes) for the univariate AD and the corresponding drift value of each packet are stored (orange entities). Features for the multivariate AD and their values for each packet are stored (purple entities).
The proposed approach integrates univariate and multi-variate methods, as illustrated in Figure 3: • at a given machine cycle, a set of raw signals = {1, . . . , } coming from the equipment sensors (and additional context information) are fed to a feature extraction block that computes features 1, . . . , , where it typically holds ̸= . While signals are usually time-series, features are scalars. Here, we consider the case where each feature is computed from a single raw signal and we denote the set of features computed from by ; • such features are used within a multivariate AD block having two objectives: (i) obtaining an anomaly score (· ), a quantitative indicator that summarizes the degree of “outlierness” of the machine cycle under exam; (ii) using an XAI approach to obtain a feature
FEATURE EXTRACTION
Features A Priori Knowledge
MULTIVARIATE ANOMALY DETECTION
List of signals to be monitored
univariately
Ranking of Variables AnoUmnaivlyarSiactoeres Concept Drift
Measure
importance score ( ), ∀ ∈ {1, . . . , }. ( ) is a quantitative index that summarizes the impact of feature in identifying machine cycle as anomalous. • raw signals that deserve to be monitored separately with a time-series based (univariate) approach are detected based on (i) expert knowledge and on (ii) the feature importance ( ) coming from the AD module. The rationale is that some anomaly types are better identified by analyzing a specific time-series. To be reliably detected, the respective signal should be independently assessed, thus avoiding the loss of information descending from a multi-variate feature extraction procedure. To identify such signals, all data points are considered. Raw signals to be processed by a univariate approach are identified by applying the following conditions: 1. a cycle is tagged as “anomalous” if it holds
( ) > , where is a pre-defined threshold on the anomaly score (under the assumption that “high” values of the anomaly score indicate a high degree of outlierness). For the problem at hand, we used = 0.55 w.r.t. the anomaly score generated by an Isolation Forest; 2. if, for an anomalous cycle, , we detect that the features obtained from a single time-series are more important than all the others in explaining the anomaly, then the corresponding time-series is sent to the univariate monitoring block. Formally is selected if there exists where and ( ) > ( ), , ∈ {1, . . . , }, ∈ , ∈/ , where > 1 is a pre-defined quantity (for our results, we have set = 2); 3. if, besides the two previous steps, some expert knowledge indicates that some raw signals are particularly relevant by themselves, then, they should be monitored univariately as well. (1) (2) 40 35 30 ]A25 [ A r r u20 C t u O l.c15 D 10 5 0
GWR prototypes 20 input data prototype 0 prototype 1 prototype 2 prototype 3 prototype 4 prototype 5 prototype 6 0 10 20
The user is finally provided with complementary information coming from the two module types, with anomaly scores, system monitoring indications, ranking of variables and concept drift estimations, allowing for a guided Root Cause Analysis.
We remark that features are not only related to physical signals acquired from the machine, but also to parameters that characterize its working condition and the surrounding environment, like the machine load (number of customers carried), the lubrication, the ambient temperature and the presence of rain. To exemplify, Fig. 4 shows a particular electric current consumption during a machine cycle: this signal was processed by extracting features related to the value of the first current peak, the rising time (to rise to 90% of the peak’s amplitude), the maximum current value, a standard deviation in its neighborhood, and the value of the last peak before the signal drops to zero. Once all the features were computed, a correlation analysis was performed to remove redundancy, obtaining vectors composed of = 23 independent features ready to be inputted into the Isolation Forest. For the concept drift estimation, we focused on one specific signal, namely, the current consumption of a particular engine, since it highly correlates with the lubrication that is performed periodically on the machine. We used, as a reference, measurements collected in a day when the gears were lubricated and during which the machine was tested in all its possible load configurations (i.e., recalling that, as shown in Fig. 1, four seats are available, the possible configurations correspond to 1, 2, 3 or 4 occupied seats). The left part of Fig. 4 illustrates these data. Some of these configurations generate traces that are clearly identifiable also by visual inspection, as a higher number of occupied seats implies more weight and, therefore, a higher current consumption. However, the proposed GWR approach does this automatically, extracting and then storing the typical patterns (the prototypes) into the dictionary depicted in the right part of Fig. 4.
Data acquired from the sensors are temporarily stored into the file system where there is a folder for each Ride containing metadata about the Ride and a set of csv files corresponding to each Session for data Ride. These files are parsed by using Python scripts and stored in a relational database, namely PostgreSQL2. Then, a data access layer, written in Python, provides API to query the data to both the anomaly detection algorithms and to the Web application. The server-side of the Web application is implemented by using the Django REST framework3. On the client-side, we rely on AJAX calls implemented by using jQuery4 and the PLOTLY5 libraries for delivering interactive plots.
Figure 5 shows an example screenshot of the Web application we have developed. It allows designers to explore, visualize, and analyze the acquired data, according to the approaches 2https://www.postgresql.org/ 3https://www.django-rest-framework.org/ 4https://jquery.com/ 5https://plotly.com/ described above. In particular, Figure 5 shows the side-by-side comparison of two diferent sessions of acquisitions and the related analysis. In this work, an AD system to monitor amusement rides has been presented. The system combines a database management system and a multi-faceted AD engine performing univariate and multi-variate analyses, exposed through a Web application. The approach is unsupervised, making it appealing for scenarios where tagged information is unavailable or unreliable. Also, features are ranked according to their importance in explaining an alarm, using an explainable artificial intelligence method.
Future work may concern diferent areas: improvements of the current framework by (i) using
other norms (e.g., | · | ∞ or DTW [
This work has been supported by the Regione Veneto POR FESR 2014-2020. Asse 1. Azione 1.1.4 (Bando per il sostegno a progetti sviluppati da Aggregazioni di imprese) initiative for the project “Trasformazione digitale innovativa nell’industria dell’entertainment” and by MIUR (Italian Minister for Education) under the initiative “Departments of Excellence” (Law 232/2016).